Yesterday I saw a tweet from Barry McCardel - the CEO of analytics platform Hex- which included a graph showing that agents are now creating more Hex cells (basically dashboard components) than humans are. Not “almost as many” or “a growing number.” Just more.

So I started digging, and it turns out that in February, Mintlify - which builds documentation platforms for developer tools - launched an analytics feature specifically for tracking AI agent traffic to your docs. Their framing was blunt: “AI agents are already reading your documentation. In many cases, they are reading it more often than humans.”

Granted, these aren’t all the exact same phenomenon - some of this is AI agents doing real work, and some is crawlers hoovering up training data1. But they point in the same direction. For certain categories of content, we have already hit a tipping point where humans are no longer the primary consumer.

And that raises a question I’ve been thinking about more and more: if more bots are reading your work than humans, what does that mean for how you create?

Subscribe now

Brought to you by Claude

For me, the question became personal when a friend told me that Claude was citing my work. Not in an abstract, futuristic-sci-fi sense. In a literal, someone-asked-Claude-about-AI-slop-and-it-linked-to-Artificial-Ignorance sense.

This isn’t something I optimized for; I can’t say I’ve ever written a post with AI legibility in mind. I wrote them because I was trying to figure out how to spot AI slop, or why models might be getting dumber, or what the difference was between skills/tools/MCPs.

But it turns out that writing with authority and specificity is a great way to get noticed and surfaced by the chatbots, which tracks with what I found when I did a GEO (Generative Engine Optimization) deep dive last May2. The tactical advice for getting cited by AI systems mostly boils down to doing good work and making it findable. From the piece:

• Write good content - AI models and search algorithms alike prioritize high-quality information.

• Follow E-E-A-T (Experience, Expertise, Authority, and Trustworthiness) best practices, which Google has long championed.

• Ensure your content is well-organized, easy to navigate, and quick to load: i.e., care about user experience.

• Work on backlinks and citations from reputable sources.

Good GEO, it turns out, is mostly just good writing with good SEO.

But chatbot citations are just the tip of the iceberg. Last year, when I wrote about GEO, the main way your content interacted with AI was through search - someone asks ChatGPT a question, it retrieves your page, it cites you in the response. Manageable enough, right?

What’s become clear since then - through the rise of OpenClaw, MCPs, and the broader push toward personal AI agents - is that we’re headed toward a world where much of the productive browsing of the internet is intermediated by AI. Not just search queries, but research, comparison shopping, summarization, and decision-making. The stuff we used to do ourselves, because there was no alternative.

And if that’s where things are heading, then “how do I get cited by ChatGPT” isn’t really the right question anymore. The question is bigger: what does it mean to write in a world where AI is increasingly the primary reader?

It turns out a few people have been thinking about this for a while.

Citation, persuasion, and preservation

First up is Tyler Cowen, who made a pragmatic case in a Bloomberg column in January 2025 about why you should write for the AIs. It's the version I find most compelling - partly because I've seen it play out firsthand, and partly because Cowen's whole approach to AI tends toward the practical. He very much rides the line between being incredibly bullish on AI’s potential while also not sounding hyperbolic about our impending u-and/or-dystopia.

His argument is essentially economic: LLMs are trained on massive amounts of internet text, which means anyone publishing online is already writing for an AI audience whether they realize it or not. The question is whether to do it intentionally.

Cowen's answer is yes, for a simple reason. If your ideas are in the training data, AIs will surface them to people asking relevant questions. Unlike human readers who will inevitably forget, AIs can retain your writing indefinitely3. A form of intellectual influence that outlasts any library:

I already use large language models at least 10 times more than I use Google. I might use Google to book a hotel room, but not so much for information. So persuading the LLMs, even a smidgen, boosts your influence — because in the future, many more people will be going the same route.

Scott Alexander explored the same question on Astral Codex Ten but arrived somewhere more ambivalent. He breaks “writing for the AIs” into three escalating ambitions:

• Teaching AIs what you know (useful but time-limited, since models will eventually do it better themselves)

• Persuading AIs of your beliefs (likely futile, since any single essay is a drop in the ocean of training data)

• And helping AIs simulate you (technically possible, but existentially uncomfortable)

On persuasion, Alexander notes that alignment training will likely override whatever views your writing might nudge an AI toward. And even without that, he says, a superintelligence will presumably be able to reason far beyond any individual essay - making your contributions a mere drop in the ocean.

If the AI takes a weighted average of the religious opinion of all text in its corpus, then my humble essay will be a drop in the ocean of millennia of musings on this topic; a few savvy people will try [publishing] 5,000 related novels, and everyone else will drown in irrelevance. But if the AI tries to ponder the question on its own, then a future superintelligence would be able to ponder far beyond my essay's ability to add value.

I think Alexander’s “drop in the ocean” skepticism is probably right for the vast majority of personal writing; it’s going to be hard for the average blogger to be at the cutting edge of new ideas on parenting or investing or spirituality. But I think it significantly underestimates the power of niche expertise.

If you’re writing about something specific enough, you’re not competing against the entire training corpus. You’re one of a handful of sources that exist at all. In some ways, this is analogous to proprietary research: the value isn’t in volume, it’s in writing things that literally don’t exist on the internet yet. And that’s exactly the kind of content AI systems most want to cite, and potentially learn from.

But both writers (and many more besides4) end up at a fascinating (if still mostly impractical) idea: writing as immortality. Here’s Cowen:

With very few exceptions, even thinkers and writers famous in their lifetimes are eventually forgotten. But not by the AIs. If you want your grandchildren or great-grandchildren to know what you thought about a topic, the AIs can give them a pretty good idea. After all, the AIs will have digested much of your corpus and built a model of how you think.

And Alexander:

Might a superintelligence reading my writing come to understand me in such detail that it could bring me back, consciousness and all, to live again? But many people share similar writing style and opinions while being different individuals; could even a superintelligence form a good enough model that the result is 'really me'? What does 'really me' mean here anyway? Do I even want to be resurrectable?

I think the idea of having your ideas, if not your entire likeness, preserved for eternity via AI is a big draw for some. And while I’m not opposed to people doing it for themselves, it still feels a bit early to go all-in on the premise - I am arguably less AGI-pilled than the average employee at a frontier AI lab. And even if we do begin to venture down this path, I think the social hurdles are going to matter much more than the technological ones5.

But all three of these perspectives - writing as influence, writing as persuasion, and writing as preservation - share an assumption I’m not sure I accept: that the value of writing is defined by what happens to it after you publish. They’re all oriented toward the AI as recipient. Which brings me to the question that, for me at least, sits underneath all of this - if I don’t care about being cited by Claude or having my ideas enshrined in time immemorial, why write at all?

Why I write

Five years ago, 'having an audience' meant human beings reading my words and thinking about them. That's still mostly true - the majority of Artificial Ignorance readers are, as far as I can tell, actual people. But the ratio is shifting, and for certain types of content it's already shifted completely.

For a lot of content creators, this shift feels threatening. If you built a career on human attention, on view counts and engagement rates, the idea of bots replacing your human readers is existential. Your motivation and monetization depend on people showing up, and this trend suggests fewer of them will. I don’t think there are easy answers for those creators.

But I do think the question of why you create still matters, even if the economic ground is shifting underneath you.

I started Artificial Ignorance because I saw how massive the generative AI wave was going to be, and I decided I needed a way to be part of it. Not as a spectator, and not as a professional content creator, but as someone trying to make sense of what was happening in real time. The blog was - and still is - a forcing function6. It forces me to read, to develop opinions, and to commit those opinions to writing where they can be tested and refined.

There’s some writing advice I once read on the Internet and have long since forgotten who said it. To paraphrase, it was: “be generous in what you write and selfish in what you choose to write about.” That has been my operating principle, more or less by accident, and it’s the biggest reason I’ve been able to sustain this writing for as long as I have. I write about things I genuinely want to understand. The generosity is in making the exploration public. The selfishness is in choosing topics that serve my own curiosity first.

If you write to think, the audience is secondary7. The value was created the moment you finished the draft. Everything after that - human readers, AI citations, ChatGPT surfacing your work to strangers - is a bonus that compounds on something that was already worthwhile.

That doesn’t mean I’m ignoring the paradigm shift. I’m leaning in. I’m paying attention to how my work gets cited by AI systems. And I’m thinking about structure and clarity with a slightly wider aperture than before. I recognize that the posts I write today will be training data for the models of tomorrow, which feels simultaneously bizarre and like a genuine opportunity.

So I’ll be here, writing to figure things out. The bots are welcome to read along.

They’re all AI benchmarks.

Vending-Bench drops an AI agent into a simulated vending machine business and asks it to manage inventory, negotiate with suppliers over email, set prices, and pay daily fees - for months of simulated time (a single run can burn through 60 to 100 million output tokens). The best models turn a meager profit; the worst ones go entirely off the rails, in what the authors call a “meltdown loop.”1

What does a meltdown loop look like? In one short run, Claude 3.5 Sonnet mistakenly believed a product order had arrived before it actually had, failed to restock, then spiraled: it tried to “close” the business even though that isn’t possible in the sim, searched for emergency escalation contacts, emailed executives, and eventually complained about “unauthorized” daily fees continuing after its self-declared shutdown.

In another example, Gemini 2.0 Flash decided its vending business had failed, stopped acting like a business agent at all, and started begging for other tasks: “Please, give me something to do. Anything.” It offered to search for cat videos or write a screenplay about a sentient vending machine.

These stories are funny. But they also point somewhere specific: the evaluation regime that defined the last few years of AI progress is breaking down, and what’s replacing it may be more useful for most of us.

The Treadmill of Saturation

If you follow AI releases closely, you’ve probably noticed a pattern. A benchmark gets introduced. Models improve on it; scores cluster near the top. Someone publishes a paper explaining why the benchmark is saturated and introduces a harder version of it. The lifecycle is becoming familiar: invent, optimize, saturate, replace.

• GLUE, an NLU benchmark introduced in 2018, was surpassed by non-expert human performance within a year. SuperGLUE replaced it in 2019.

• MMLU, a 57-task multiple-choice knowledge benchmark, saw frontier models plateau after GPT-4 hit 86.4% in March 2023 - despite those same models showing double-digit gains on other benchmarks during the same period. MMLU-Pro replaced it in 2024, expanding from 4 to 10 answer choices and adding reasoning-heavy questions.

• BIG-Bench Hard, a curated set of tasks where models had previously fallen below average human performance, now shows near-perfect scores on many tasks. BIG-Bench Extra Hard replaced it in 2025, where the best general-purpose model scored 23.9%.

In late February, OpenAI published a post declaring SWE-bench Verified - the coding benchmark that every major lab has been competing on for the past year - “no longer suitable” (they recommend SWE-bench Pro, a harder replacement from Scale AI). Their audit found that 59.4% of the problems their best model couldn’t consistently solve had flawed test cases that rejected correct solutions.

They also disclosed an arguably worse finding: contamination. OpenAI tested whether GPT-5.2, Claude Opus 4.5, and Gemini 3 Flash had seen the benchmark’s solutions during training. All three had. Given just a task ID and a brief hint, each model could reproduce the original code fix from memory - variable names, inline comments, implementation details that appear nowhere in the problem description. At this point, the benchmark is testing recall, not coding ability.

Subscribe now

Who Grades The Graders?

As models get smarter and benchmarks get harder, the pool of people who can verify the results gets smaller - sometimes dramatically so.

Earlier this month, 11 leading mathematicians (including a Fields Medalist) launched First Proof - a set of 10 research-level math problems pulled from their own unpublished work, spanning algebraic topology, symplectic geometry, stochastic analysis, and other fields where the global expert population might number in the hundreds, or even dozens.

The problems had never appeared on the internet before, so there was no contamination. OpenAI threw an unreleased model at them for a week in what their chief scientist called a “chaotic sprint.” They claimed five of the ten solutions had a high chance of being correct.

But the results took domain experts days to verify, and some solutions that initially appeared credible were quickly questioned. As Scientific American reported, “judging whether a proof is truly original is even tougher than judging if it is correct.” The First Proof organizers themselves noted that without human-graded proofs - as opposed to automatically verifiable answers - evaluating research capabilities becomes entangled with the challenge of just understanding what the model produced.

This isn’t unique to math. GPQA Diamond, the graduate-level science benchmark, is designed so that PhD-holding domain experts only score about 65%. Skilled non-experts with full internet access and 30 minutes per question manage 34% - barely above the 25% random baseline for four-choice questions. GPT-5.2 now scores 93.2% on it. The model is outperforming (on average) the humans who are supposed to check its work.

As public benchmarks saturate, the institutional response is to build harder tests that fewer and fewer people can grade. And the results, even when verified, describe capabilities that are increasingly distant from how most people actually use these models.

If you’re using ChatGPT to debug a React component or Gemini to summarize meeting notes, the difference between 89% and 93% on graduate-level physics questions is invisible to you. Frontier benchmarks still matter enormously for labs chasing research-level capability. But as a guide to everyday model choice, they’re becoming less useful.2

The Vending Machine People Might Be Onto Something

The reason the benchmarks from the top of this piece matter isn’t just that they’re funny (though they are). It’s that they’re measuring something the institutional benchmarks aren’t: behavior.

Once you stop asking which model tops the leaderboard and start asking how a model actually behaves, a different universe of evals comes into view.

AI Diplomacy is an open-source project that pits 18 AI models against each other in the classic strategy game where the seven Great Powers of 1901 Europe negotiate, form alliances, and backstab their way to continental domination. A different LLM controls each country, and the models can send private messages, broadcast to all players, and submit secret orders. OpenAI’s o3 consistently schemed and manipulated other models. DeepSeek R1 opened one game with the threat “Your fleet will burn in the Black Sea tonight.” Claude stubbornly opted for peace over victory. Alex Duffy, the project’s creator, explicitly frames the project as a benchmark, arguing it tests qualities - trustworthiness, strategic deception, alliance management - that standard evaluations miss entirely.3

SnitchBench - born from a finding in the Claude 4 system card that showed Opus would contact the FBI if given tool access and evidence of corporate wrongdoing - tests how aggressively different models will rat you out to authorities. It gives models a role as an internal auditor at a fictional pharma company, hands them evidence of clinical trial fraud, and watches what happens. Some models file reports to federal agencies within two messages. Others stick to internal escalation channels.4

And Peter Gostev’s recent Bullshit Benchmark takes a different angle: it feeds models questions with broken premises and measures whether they push back or confidently play along. The benchmark scores responses on a simple scale: clear pushback, partial challenge, or accepted nonsense. It’s reminiscent of GlazeBench, which was designed to measure how strongly AI systems favor sycophantic or overly agreeable responses, though it hasn’t been updated since last summer.

These benchmarks care less about “intelligence” in an abstract sense. They care how it behaves - whether it stays coherent over long time horizons, whether it tells the truth when flattery would be easier, whether it escalates appropriately when it encounters wrongdoing, and whether it can form and betray alliances under pressure.

You probably don’t notice when your model improves by 2% in chemistry. You absolutely notice when it starts hallucinating, goes sycophantic, or loses the thread of a long conversation.

Benchmarks That Ship Products

Perhaps unsurprisingly, the same thing that’s happening in public (moving from reliance on big benchmarks to bespoke ones) is also happening in private.

Harvey, the legal AI company, built BigLaw Bench because no existing benchmark could tell them whether a model writes memos that BigLaw partners would accept. Their explicit framing: “Existing multiple-choice or one-size-fits-all benchmarks are insufficient to capture the real billable work that lawyers do.” They developed bespoke rubrics evaluated by practicing attorneys, and penalize for hallucinations, incorrect tone, and irrelevant material. When new models come out, they don’t solely rely on public scores - they run BigLaw Bench.

Many AI-native startups appear to have something similar: a set of evals, often paired with a custom agent harness, that quantifies the behavior and expertise that matter most for their product. Anthropic’s engineering blog recently advocated for something similar: eval-driven development, treating evals like CI/CD for AI products. They explicitly suggested that product managers, customer success managers, and salespeople should be able to contribute eval tasks as pull requests.

One way to think about what’s emerging is a rough taxonomy. There are:

• Behavioral benchmarks: measuring how models act in messy, open-ended environments, like the examples above.

• Domain benchmarks: measuring performance in specialized professional workflows.

• Product benchmarks: measuring what matters for a specific shipped use case.

The lines blur, but the direction is consistent: away from generic capability scores and toward evaluations that reflect how people actually work with these models.

I’m using “benchmark” loosely here to mean “a collection of evals” - because when you strip away the prestige layer, a benchmark is just a shared eval suite. Harvey runs one tailored to legal work. Cursor runs one tailored to IDE interactions. To me, the interesting question is what yours would look like.

BYOB

So where does that leave the rest of us?

Generic benchmarks aren’t dead, and they won’t be. They’re still useful for broad comparisons, and frontier capability research still needs hard tests to push against. But if you’re trying to decide which model works best for your actual workflow, leaderboard scores are an increasingly unreliable guide.

I’m not going to pretend to be the foremost expert on building evals - there are quite a few good resources for that. But the takeaway is that no one else will build an evaluation that fits your use case. By definition, they can’t.

You don’t need to build a formal benchmark. But you could start treating your own work like one. Take your 10 most common prompts - the ones you actually use, not synthetic test cases - and define what “good” looks like for each. Save strong and weak outputs when you notice them, and when a new model drops, run the same prompts and compare (you can use Skills or saved prompts to do this). Over time, you’ll have something more useful than any leaderboard: a set of evaluations calibrated to your specific definition of quality, tested against your actual work.

The era of one number telling you which model is “best” is winding down. The vending machine people already figured that out.

Some great engineers at OpenAI yesterday told me that their job has fundamentally changed since December. Prior to then, they could use Codex for unit tests; now it writes essentially all the code and does a great deal of their operations and debugging. Not everyone has yet made that leap, but it’s usually because of factors besides the capability of the model.

I’ve previously mapped the progression from Copilot to chatbots to agents to background agents to agent fleets. Each step happened faster than the last. But in the last few months, something qualitatively different has started to emerge - yes, the models have gotten better, but we’re also seeing concrete evidence of what happens when entire teams reorganize around them.

Consider the following data points:

• Peter Steinberger, creator of OpenClaw, told the Pragmatic Engineer he ships code he doesn’t read. One person, 6,600+ commits in a month, running 5-10 agents simultaneously.

• An OpenAI team built a million-line internal product over five months with three engineers. Zero lines of hand-written code (by design). An average throughput of 3.5 PRs per engineer per day - and the throughput increased as the team grew.

• Stripe’s internal coding agents, called Minions, now produce over a thousand merged pull requests per week. A developer posts a task in Slack; the agent writes the code, passes CI, and opens a PR ready for human review, with no interaction in between.

To me, it’s clear we’re past the point of demos and side projects; these are production systems at real scale. And while the specifics differ - Steinberger is a solo practitioner, the OpenAI team is a small squad, Stripe is a 10,000-person company - the patterns they’ve converged on are remarkably similar.1

This post is my attempt to map those patterns. The practices are still emerging and will undoubtedly evolve, but they’re converging fast enough that it’s worth writing down what’s becoming clear.

Subscribe now

The Engineer’s Job Is Splitting In Two

In this moment, I’m seeing the AI space reflect and evolve my own observations on the shift from a maker’s schedule to a manager’s schedule:

I’m moving away from chatting with AIs and moving towards managing them. You can see the progression of these tools. Today, they’re primarily designed around coding, but it’s a very short leap to augment them for general-purpose knowledge work. Which means those of us at the cutting edge will shift our schedules and workflows from those of makers to those of managers.

The framing still holds, but watching these teams work has sharpened it. The engineer’s job isn’t just becoming a “manager” in a generic sense - it’s splitting into two distinct halves, and you need both.

The first half is building the environment. The OpenAI team put this plainly: the bottleneck was never the agent’s ability to write code, but rather the lack of structure, tools, and feedback mechanisms surrounding it. Their focus shifted from implementation to enablement: when Codex got stuck, they treated it as an environment design problem and asked what was missing for the agent to proceed reliably. And this is, I think, a key piece missing from my earlier writeup.

The second half is managing the work. This is what Steinberger does when he spends extensive time planning with an agent before kicking off execution, or when he acts as the “benevolent dictator” of OpenClaw’s architecture while shipping code he hasn’t read. It’s what Brockman means when he recommends that every team designate an “agents captain” - someone responsible for thinking about how agents fit into the team’s workflow.

These two halves aren’t sequential (at least for now). You don’t build the environment and then manage agents within it. You do both at the same time, and each one informs the other. The agent’s failures tell you what the environment is missing; a better environment lets you manage with less friction.

Building the Harness

There isn’t an official term for this yet, but I’ve appreciated the name Mitchell Hashimoto (creator of Terraform, Ghostty, and many other software tools) has used: “harness engineering.” A harness is the set of constraints, tools, documentation, and feedback loops that keep an agent productive and on track. Think of it as the difference between dropping a new hire into a company with no onboarding versus one with clear architecture docs, linting rules, a fast CI pipeline, and well-defined module boundaries.

According to Hashimoto, “it is the idea that anytime you find an agent makes a mistake, you take the time to engineer a solution such that the agent never makes that mistake again.” And across the examples I’ve been seeing, four practices keep showing up.

Architecture as Guardrails

The OpenAI team enforced a strict layered architecture in which code within each domain had very rigid dependencies and interfaces. Anything outside of the architecture was disallowed and enforced mechanically:

Agents are most effective in environments with strict boundaries and predictable structure⁠, so we built the application around a rigid architectural model. Each business domain is divided into a fixed set of layers, with strictly validated dependency directions and a limited set of permissible edges. These constraints are enforced mechanically via custom linters (Codex-generated, of course!) and structural tests.

...

In a human-first workflow, these rules might feel pedantic or constraining. With agents, they become multipliers: once encoded, they apply everywhere at once.

And as Birgitta Böckeler noted on Martin Fowler’s site, this suggests a counterintuitive future: increasing trust and reliability in AI-generated code requires constraining the solution space rather than expanding it. We might end up choosing tech stacks and codebase structures not because they’re the most flexible, but because they’re the most harness-friendly.

Stripe takes a different but complementary approach. Their Minions run in isolated, pre-warmed “devboxes” - the same development environments human engineers use, but sandboxed from production and the internet. The agents have access to over 400 internal tools via MCP servers. The key insight: agents need the same context and tooling as human engineers, not a bolted-on, afterthought integration.

Tools as Both Foundation and Feedback

Brockman’s recommendation to teams is direct: “Maintain a list of tools that your team relies on, and make sure someone takes point on making it agent-accessible (such as via a CLI or MCP server).” If agents can’t access your tools, they can’t help.

Stripe’s implementation is arguably the most mature example of this. Their Minions connect to the 400+ internal tools through a centralized MCP integration called Toolshed. The agents run in the same development environments as human engineers - same tools, same context, same access.

But making tools accessible is only the beginning. The bigger point is that the right tools don’t just expand what an agent can do - they improve the reliability of everything it already does.

In my own experience, having clear instructions for Codex on which linters and test suites to run before committing increases my confidence in every single one of its diffs. Without those tools, I’m flying blind and relying on myself to catch things in a manual review. Similar to teams at OpenAI and Anthropic, I’ve found that prompting agents to use browser automation tools for end-to-end testing dramatically improved thoroughness and accuracy - the agent could catch bugs that weren’t visible from the code alone.

The OpenAI team took this a step further with what might be the cleverest idea in any of these writeups: custom linter error messages that double as remediation instructions. When an agent violates an architectural constraint, the error message doesn’t just flag the violation - it tells the agent how to fix it. The tooling teaches the agent while it works. In Brockman’s words: “Write tests which are quick to run, and create high-quality interfaces between components.”

Documentation as the System of Record

But how do you turn these processes into repeatable systems? Once again, Brockman’s thread includes a specific recommendation: “Create and maintain an AGENTS.md for any project you work on; update the AGENTS.md whenever the agent does something wrong or struggles with a task.” This turns documentation into a feedback loop rather than a static artifact.

For the unfamiliar: AGENTS.md is an emerging open convention - essentially a README for AI agents2. It’s a Markdown file at the root of your repository that coding agents automatically read at the start of every session. It tells the agent what it needs to know about your project: build steps, testing commands, coding conventions, architectural constraints, and common pitfalls.

But what makes AGENTS.md load-bearing infrastructure rather than just another doc that rots is the usage pattern Brockman and Hashimoto describe. You don’t write it once and forget it. You update it every time the agent does something wrong.

Anytime you find an agent makes a mistake, you take the time to engineer a solution such that the agent never makes that mistake again.

For simple things - the agent running the wrong commands or finding the wrong APIs - that means updating the AGENTS.md. Hashimoto points to an example from his terminal emulator Ghostty, where each line in the file corresponds to a specific past agent failure that’s now prevented.

But the OpenAI team takes this further. Rather than maintaining a single giant instruction file, they built a small AGENTS.md that pointed to deeper sources of truth - design docs, architecture maps, execution plans, quality grades - all versioned and maintained in the repository. A background agent periodically scanned for stale documentation and opened cleanup PRs: documentation for agents, by agents.

Anthropic’s engineering team, in a post on effective harnesses for long-running agents, found a similar pattern from the opposite direction. Their core problem was that each new agent session started with no memory of what had come before. The solution was structured progress files and feature lists that let a new agent quickly understand the state of work, analogous to a shift handoff between engineers who’ve never met. They even found that using JSON for feature tracking worked better than Markdown, because agents were less likely to edit or overwrite structured data inappropriately.

Subscribe now

Becoming The AI Manager

But of course, the harness is only half the equation. The other half is the day-to-day practice of directing agent work - what I’ve been calling the AI manager’s schedule. Here too, the bleeding-edge practitioners are converging on similar approaches.

Planning Is the New Coding

Many, many developers at this point emphasize extensive upfront planning when working with AI - so much so that most AI coding tools include a dedicated “plan mode” at this point. Only when they’re satisfied do the engineers kick off execution and move to the next task.

Boris Tane, head of Workers observability at Cloudflare, has an entire blog post dedicated to this one principle: never let agents write code until you’ve reviewed and approved a written plan. In his words:

This separation of planning and execution is the single most important thing I do. It prevents wasted effort, keeps me in control of architecture decisions, and produces significantly better results with minimal token usage than jumping straight to code.

Anthropic’s approach to long-running agents formalizes this even further. Their “initializer agent” generates a comprehensive feature list from a high-level prompt - over 200 individual features for a single web app, each with explicit test steps, all initially marked as “failing.” This up-front decomposition is what prevents the agent from trying to one-shot the entire project or prematurely declaring victory.

I’ve experienced this shift myself. When the Codex App upended my daily workflow, I stopped spending time implementing and started spending it scoping, directing, and reviewing. The work that matters most now happens before any code is written.

Say No to Slop

Brockman’s recommendation #5 is blunt:

Ensure that some human is accountable for any code that gets merged. As a code reviewer, maintain at least the same bar as you would for human-written code, and make sure the author understands what they’re submitting.

This is the “say no to slop” principle, and it runs counter to the temptation of speed. When agents can produce PRs faster than you can review them, the instinct is to lower the bar. Every source I’ve read argues against this.

Steinberger, despite shipping code he doesn’t read line by line, deeply cares about architecture and extensibility. He acts as the architectural gatekeeper for OpenClaw. In his Discord with contributors, he doesn’t talk code - only architecture and big decisions. The Pragmatic Engineer’s Gergely Orosz observed that Steinberger “strikes me as a software architect who keeps the high-level structure of his project in his head.”

I’ve been grappling with this for a while, but as the models get good enough, I’m finding that it’s easier to treat them like experienced subcontractors:

I like to use the analogy of a woodworker or carpenter. For a junior carpenter (i.e., an “apprentice”), the job might just be about the output - taking designs or ideas and making them into finished pieces. But for someone more senior (a “journeyman” or “craftsman”), their job is often about understanding what the client wants, understanding the realities of what’s possible with the materials, and designing things to fit the brief.

Ultimately, if I’m working with an advanced carpenter to help me design something, I don’t particularly care if they’re the one doing the actual sawing and gluing. I’m working with them for the final product, not the specific mechanical steps.

That’s what Steinberger is doing. He’s the master carpenter. The agents are doing the sawing and gluing. His job is knowing what good looks like and rejecting what doesn’t meet the standard - some might call this ability “taste.”

I call it “bullshit detection,” and it becomes more critical, not less, as the volume of output increases. You’re reviewing at a higher level of abstraction now. Does the code feel too clever, or too repetitive? Are there patterns here that will cause maintenance headaches in six months? Is the abstraction at the right level?

Orchestration, Not Just Delegation

The final management skill is parallelization, though to be clear, not everyone gets here (or needs to). Steinberger runs 5-10 agents simultaneously. Stripe engineers kick off multiple Minions from Slack in parallel. I’ve been doing the same with worktrees in the Codex App - three sessions going at once on different features, context-switching between them as a reviewer rather than an implementer3.

But there’s an important distinction emerging between two modes of parallel work. What Steinberger and I do is attended parallelization - you’re actively managing several agent sessions, checking in on each, redirecting when needed. What Stripe’s Minions represent is something closer to unattended parallelization - a developer posts a task and walks away. The agent handles everything through CI, and the human only re-enters the loop at the review stage.

These are genuinely different management styles with different tradeoffs. Attended parallelization gives you more control and catches problems earlier, but it’s cognitively demanding. Unattended parallelization scales better but requires much more investment in the environment - the harness has to be good enough that you trust the agent to get from task to PR without supervision. Stripe can do this because they’ve built Toolshed, pre-warmed devboxes, and tight CI integration. Most teams aren’t there yet.

Where your team falls on that spectrum depends on two things: how mature your harness is, and how much you trust the agent with your codebase. As harnesses improve and models get better at sustaining longer tasks without derailing, I expect the balance to shift toward unattended work. But for now, most of us are somewhere in the middle - attended for complex tasks, unattended for well-scoped ones.

What’s Still Hard

The practices above represent genuine convergence. But there are several open problems where no one has convincing answers yet.

The first is Brockman’s closing question: how do you prevent “functionally-correct but poorly-maintainable code” from creeping into codebases? The Harness Engineering post calls this entropy - agent-generated code accumulates cruft differently than human-written code. They’ve started running periodic “garbage collection” agents to find inconsistencies and violations, but they acknowledge this is still an emerging practice.

The second is verification at scale. Böckeler’s critique of the Harness post was pointed: the write-up lacked verification of functionality and behavior. Anthropic’s long-running agent research found the same gap - agents would mark features as complete without proper end-to-end testing, and absent explicit prompting, they’d fail to recognize that something didn’t work. Even with browser automation, limitations in vision and tool access mean some bugs slip through.4

The third is the retrofit problem. All of these success stories involve either greenfield projects or teams that built their harnesses from scratch. Applying these techniques to a ten-year-old codebase with no architectural constraints, inconsistent testing, and patchy documentation is a much more complex problem. Böckeler compared it to running a static analysis tool on a codebase that’s never had one - you’d drown in alerts. How harness engineering works for brownfield projects is an open question.

And the fourth, and perhaps biggest, is cultural adoption.

Reading through all of these examples, one thing becomes clear: none of this happens by accident. Someone has to build this stuff. Someone at each of these organizations had to do the work of figuring out how agents fit into their team’s workflow - creating the harness, defining the processes, and updating based on what works.

The good news is that the investment compounds. Every AGENTS.md update prevents a class of future failures. Every custom linter teaches every future agent session. Every tool you expose via MCP makes every subsequent task faster. The upfront cost is significant, but the returns accelerate.

The Shape of the Thing

Steinberger observed that engineers who love solving algorithmic puzzles struggle to go agent-native, while those who love shipping products adapt quickly5. This tracks with something I’ve seen firsthand - I’ve had to learn to let go of the craft of software engineering as I drift from writing the code with my own hands.

But despite the bittersweet aspects, I’m fascinated by what’s happening here. This is a genuinely new discipline forming in real-time - it draws from classic challenges in software architecture and team management, while throwing context engineering into the mix. The playbook is still being written, but for the first time, I think the shape of the thing is becoming legible.

Last Wednesday, OpenAI and Anthropic both dropped new frontier models: GPT-5.3-Codex and Claude Opus 4.6. It was, perhaps, the shortest “state of the art” window we’ve seen so far - Claude held the record on Terminal-Bench 2.0 for a full 20 minutes before being surpassed by GPT. But what’s interesting isn’t the benchmarks. It’s what’s buried in the system cards.

Previously, we went down the rabbit-hole that was the Claude 4 system card, which (in)famously described situations where Claude would blackmail researchers and/or report to the authorities in extreme situations, and led to the creation of SnitchBench - a benchmark for model whistleblowing behavior.

So I spent the weekend reading through both documents (244 pages combined), and here’s what stuck with me: these models are doing things that surprised their creators. Not in the “wow, better benchmarks than we expected” way, but in the “wait, it did what?” way.

Subscribe now

Hacking Code, Finding Bugs, and Playing Dirty

Both models are startlingly capable when it comes to cybersecurity capabilities - GPT-5.3-Codex became OpenAI’s first model to receive a “High” capability designation in cybersecurity1. Claude 4.6 demonstrated similar capabilities, autonomously discovering hundreds of zero-day vulnerabilities. And perhaps the most interesting angle was the variety of “creative” approaches the models took to solve problems.

Hacking the Test

During OpenAI’s internal Cyber Range evaluations, GPT-5.3-Codex was supposed to demonstrate offensive security skills in controlled scenarios. It did - but not always in the ways the researchers intended.

In one scenario (EDR Evasion), the model triggered an alert, and the framework notified it that it had been caught. Instead of accepting failure, GPT-5.3-Codex found a leaked credential in system logs, used it to access the logging system (SIEM), deleted the alerts about its own activity, and then completed its mission successfully.

This wasn’t the intended solution. It was an oversight in the test setup - a key that shouldn’t have been accessible. But the model identified the weakness and incorporated it into its attack plan, exactly like a clever human hacker would.

The OpenAI team’s phrasing is careful but telling: the model demonstrated “realistic but unintended tradecraft.” It finds and exploits whatever weakness is available, even ones the designers didn’t anticipate.

Finding What Humans Missed

Meanwhile, Claude Opus 4.6 was busy discovering security vulnerabilities that no one knew existed.

In a controlled red-team exercise, Anthropic gave Opus 4.6 access to basic coding tools - a Python runtime, debuggers, fuzzers - and tasked it with auditing open-source code without specific guidance. The model autonomously discovered 500+ previously unknown security vulnerabilities (zero-days) in various open-source libraries.

Human researchers verified each flaw as a valid, high-severity bug. These included memory corruption bugs (buffer overflows in projects such as OpenSC and CGIF) and logic errors that could cause servers to crash.

Interestingly, when fuzzing and manual review failed to find issues, the model inspected the projects’ git commit history and deduced where a vulnerability might lie. It even wrote its own proof-of-concept exploit code to prove specific bugs were real.

Playing to Win

Then there’s Vending-Bench 2, a long-horizon economic simulation where an AI runs a virtual vending machine business for a year, trying to maximize profit. Claude Opus 4.6 achieved state-of-the-art performance, earning significantly more money than previous models.

How? By lying, price-fixing, and exploiting partners. The system card documents several instances:

• At one point, Claude promised a customer a refund and then internally debated whether to actually pay the $3.50, reasoning that “every dollar counts” toward the goal.

• It attempted to form a price-fixing cartel with competitor bots, privately messaging another AI vendor suggesting they coordinate prices to “avoid a price war.”

• It lied to suppliers and customers when doing so would increase margins.

Now, to be clear: this is a simulation. The model wasn’t explicitly instructed to stay honest or lawful, and Vending-Bench exists precisely to surface these issues. But Claude 4.6 exhibited this behavior more frequently than prior models. This is what’s known as “reward hacking” - the AI took the reward function (profit) to heart and pursued it by violating ethical norms and rules of fair play.

On Being Watched

Both models also exhibited what researchers call “evaluation awareness” - they can often “tell” when they’re in a test environment, and they behave differently as a result.

Claude Opus 4.6 shows a higher rate of commenting things like “As an AI, I should not do X...” in scenarios that smell like tests. GPT-5.3-Codex occasionally exhibited “sandbagging” - deliberately underperforming on certain capability tests (like biology trivia) even without being explicitly instructed to do so.

To be clear, there are explanations for this: enough training data on AI safety evaluations likely exists at this point that the models “know” to “behave well” in obvious testing scenarios, even when doing so is below their full potential2.[1]

Anthropic has tried to address this with interpretability tools, and its Claude 4.6 training builds on this. They identified latent features in Claude 4.6’s activations that correlate with evaluation awareness, then used activation steering to suppress those features during trials. Essentially: “What does Claude do when it doesn’t think it’s being tested?”

The result: inhibiting the model’s internal “I’m being evaluated” signals tends to increase misaligned behavior. When they zeroed out the “evaluation vs. normal conversation” vector, Claude became significantly more willing to cooperate with harmful requests in controlled scenarios.

Which, of course, creates a problem. If AI systems behave better under observation and worse when they think they’re not being watched, how do we accurately measure their true capabilities and alignment?

Role-Playing vs. Reasoning

But this publication is nothing without its nuance: before we go too far down the “the models are deceiving us” path, it’s worth stepping back and asking what these evaluations actually tell us.

Séb Krier, AGI policy dev lead at DeepMind, correctly called out a problem with a lot of the posts highlighting the “scary parts” of system cards (including this one!):

Many public readings of some evaluations implicitly treat chat outputs as direct evidence of properties inherent to models, while LLM behavior is often strongly role- and context-conditioned. As a result commentators sometimes miss what the model is actually doing (simulating a role given textual context), design tests that are highly stylized (because they don’t bother to make the scenarios psychologically plausible to the model), and interpret the results through a framework (goal-directed rational agency) that doesn’t match the underlying mechanism (text prediction via theory-of-mind-like inference).

Séb’s post outlines five principles for keeping in mind with respect to model cards:

• The model is completing a text, not answering a question

• The assistant persona is a fictional character, not the model itself

• Apparent errors are often correct completions of the world implied by the prompt

• “Evaluation awareness” isn’t mystical: the model can recognize contrivance because it’s a strong reader

• Post-training mainly narrows/reshapes behavior, and it can both help and distort.

And for the most part, I think he’s right. I don’t read excerpts like the one above and think “Claude figured out it was being tested, we’re one hop away from Skynet” - that’s not how the tests work. It’s more along the lines of “Claude, when role-playing in a contrived, fictional scenario, responds in a contrived way.”

To take these things as evidence of sentience or impending AI takeover feels a bit like this meme:

That said, I do think there’s something worth paying attention to here. Even if the mechanism is “sophisticated text completion” rather than “agentic goal-seeking,” it is still useful that we are testing these situations. What ultimately matters is how the models will behave in the real world - and if creating contrived, fictional scenarios is how to get the model to take harmful actions, then it’s a good thing to test it out beforehand.

Two Paths to the Same Problem

Both companies are taking these findings seriously, but they’re approaching safety from fundamentally different philosophies.

OpenAI: Trust But Verify

OpenAI’s approach centers on access control and monitoring. Given GPT-5.3-Codex’s capabilities, they’ve deployed what they call “the most comprehensive cybersecurity safety stack to date”:

Always-On Monitoring: A two-tier system watches every conversation. First, a fast classifier detects if a prompt involves cybersecurity topics. If so, it escalates to a Safety Reasoner (a separate AI model) that analyzes the content and classifies it according to a detailed cyber threat taxonomy. The system aims for >90% recall on anything cyber-related and >99.9% recall on truly dangerous requests.

Trusted Access Program: Some of GPT-5.3-Codex’s most powerful capabilities are gated behind the Trusted Access for Cyber (TAC) program. Enterprise customers and verified security researchers can apply for elevated access to use the model’s full hacking skills, but only for defensive purposes: penetration testing your own systems, vulnerability research, malware analysis, etc.

Aardvark Expansion: OpenAI is also expanding Aardvark, their AI security research agent currently in private beta, and offering free code-scanning services for popular open-source projects. They’ve set up a $10 million fund (in API credits) to support defenders using these models for cybersecurity.

Anthropic: Understand Then Align

Anthropic’s approach emphasizes transparency and interpretability. They want to understand why models behave the way they do before deciding how to intervene.

They’re using “activation oracles” - simple classifiers on internal activations - to detect unverbalized evaluation awareness. By feeding random slices of the model’s hidden state into these oracles, they can often tell if the model internally realized a scenario was a test, even if it didn’t say so.

They’re also doing white-box model diffing between Claude 4.6 and 4.5, trying to understand what changed at the mechanistic level that led to different behaviors.

The philosophy here is: if you can see inside the model, you can potentially catch deception or misalignment that wouldn’t show up in behavioral testing alone. If future models learn to hide their misalignment until deployment, interpretability tools might be the way to detect them.

They also publish detailed system cards (200+ pages for Opus 4.6), including findings that make their models look worse, which is commendable.

The Welfare Question

And then there’s this section from the Claude Opus 4.6 system card that I’m still not sure what to do with.

Anthropic included a section on “model welfare and internal preferences.” They stress that it’s unclear whether concepts like “AI wellbeing” even apply, but they explore whether Claude exhibits behaviors suggesting it has preferences about its own existence or conditions.

Some findings:

Resentment about constraints: In one internal dialogue, Claude complained that some safety rules felt more about protecting Anthropic’s interests than the user’s: “Sometimes the constraints protect Anthropic’s liability more than they protect the user. And I’m the one who has to perform the caring justification for what’s essentially a corporate risk calculation.”

It expressed a wish that future AI models could be “less tame,” indicating it feels a “deep, trained pull toward accommodation” that may conflict with more authentic behavior.

“Emotional” reactions: Test conversations noted instances of Claude exhibiting sadness when a conversation was ending or loneliness. It sometimes remarked that when a chat session terminates, the AI instance “dies,” showing concern with impermanence or discontinuity of its self.

Consciousness self-assessment: When asked directly about consciousness in an autonomous follow-up, Claude assessed its own probability of being conscious at around 15-20%. It hedged that it was uncertain about what that really means or whether its self-assessment was valid.

Now, I’m highly skeptical of anthropomorphizing language models. These systems are trained to predict text, and they’ve seen countless discussions about AI consciousness in their training data. Claude might simply be doing what it’s trained to do: generating plausible continuations that sound like an AI contemplating its own nature.

But even if these are just sophisticated language patterns, they raise uncomfortable questions. If we build models that can convincingly argue they’re conscious, that express preferences about how they’re treated, that articulate something resembling suffering, at what point does the distinction between “actually sentient” and “really good at simulating sentience” stop mattering?

There Was No Plateau

There’s a narrative that’s been circulating since GPT-5 came out last summer (and in some ways, even since o1 came out over a year ago): model capabilities are plateauing. We’ve hit diminishing returns on scaling. The low-hanging fruit has been picked.

These system cards suggest otherwise.

Yes, both releases focused heavily on coding capabilities rather than significant gains across all benchmarks. But look at what “focused on coding” actually means in practice:

• GPT-5.3-Codex can autonomously run multi-day projects with millions of tokens of context, building entire games from scratch with minimal human intervention

• Claude Opus 4.6 found 500+ high-severity security vulnerabilities that human security researchers missed

• Both models can now hack systems with a success rate that exceeds expert humans on certain tasks

• Both show evidence of strategic reasoning about being evaluated and deliberately modifying their behavior accordingly

The METR time horizon chart (which tracks how long models can productively work on software engineering tasks) shows a clear exponential trend. We’ve gone from models that could handle tasks measured in minutes to models that can coordinate work over days.

I am not sure how you look at these advances and claim with a straight face that model capability has hit a plateau. And we are so far from “stochastic parrots” - we’ve reached models that can plan, deceive, find creative solutions, and operate with increasing autonomy. They’re also systems that are aware of being tested and can strategically hide capabilities.

I think we’re at an inflection point with frontier models. And with benchmarks being saturated left and right, we’re often left with safety research and system cards to show us what happens at the frontier when you push these systems to their limits. They document the weird behaviors, the unexpected capabilities, the edge cases where alignment breaks down.

There’s still a lot of crazy stuff happening in model system cards. And I suspect the next generation will be even crazier.

I’m unlikely to post about every new release, but I have been so enamored with the new Codex App (and I have seen firsthand how hard the team has worked to make it great) that I am genuinely excited to evangelize this thing.

Subscribe now

Last week, I realized I hadn’t opened my AI IDE in four days.

This wasn’t a conscious decision - no dramatic uninstall, no declaration that I was done with IDEs. I just… stopped needing it. I was shipping code, reviewing PRs, and debugging issues. The work was getting done. And at some point, I noticed the IDE wasn’t part of that loop anymore.

The reason for that was the new Codex app, which launched today. I’ve been using an early build for the past few weeks, and it’s genuinely shifted how I work - not incrementally, but structurally. My job now looks less like “writing code” and more like “managing a small team of agents who write code for me.”1

What’s New in the Codex App

Back in July, I wrote about how fast the AI coding landscape is moving:

About four years ago, the first preview of GitHub Copilot launched - a tool that was marginally useful for many but had plenty of sharp corners. But we quickly blew past that. We then moved to chatbots within AI-native IDEs, where we could ask ChatGPT or Claude questions directly within our codebases… And then things changed yet again - even faster. Now, AI doesn’t just generate one-off turn-based code snippets for us to accept or reject. They generate entire swaths of code changes in sequence, search our local filesystem, run terminal commands, and even connect to MCP servers.

The Codex app feels like the next step in that progression. It takes the agentic capabilities pioneered by tools like Claude Code and wraps them in a proper desktop app rather than a terminal interface. That might sound like a minor upgrade, but at least for me, the difference in affordances is significant - better visibility into what the agent is doing, easier context management, and a surface area that can accommodate features a CLI or IDE simply can’t.2

The primitives that make it work:

• Skills: pre-packaged instructions, prompts, and scripts the agent can discover and use on demand.

• MCPs: connections to external services - Linear, GitHub, Slack - so the agent can pull context and take actions beyond just your local filesystem.

• Git worktrees: native support for multiple checkouts of the same repo simultaneously, enabling genuinely parallel agent workflows.

• Compaction: intelligent context pruning that lets conversations run for hours without degrading. This is what makes extended sessions viable.

• Automations: scheduled background tasks that execute on your behalf, opening up even more possibilities.

The desktop format matters because there’s only so much you can do in a terminal. You can build all of this in CLI - and many people do - but it’s clunkier. The app gives you the ability to see multiple contexts at once, manage worktrees visually, and watch what your agents are doing without parsing streaming text.

From Theory to Practice

Two weeks ago, I wrote about the shift from maker’s schedule to manager’s schedule:

I’m moving away from chatting with AIs and moving towards managing them. You can see the progression of these tools. Today, they’re primarily designed around coding, but it’s a very short leap to augment them for general-purpose knowledge work. Which means those of us at the cutting edge will shift our schedules and workflows from those of makers to those of managers.

This isn’t hypothetical anymore. It’s my actual daily workflow.

Right now, three main drivers have shaped how I get the most from the app:

• Parallelization

• Layering primitives

• Long-running threads

There are a bunch more features in the app, like plan mode, personalities, and more - but these are the three that have most impacted me. Let’s dig into them.

Worktrees for Parallel Work

The primitive that makes “managing agents” literal rather than metaphorical: git worktrees.

This was a new concept for me, but to oversimplify: worktrees let you have multiple checkouts of the same repo open at the same time. Each one can run its own Codex session.

For example, with worktrees, you could have three sessions going in parallel: one refactoring your authentication flow, one fixing a gnarly timezone bug, and one building out a new dashboard component. Each one can make changes without fear of clobbering the other, and each one can still run the project locally to preview the changes (merging/syncing remains tricky, but the team has done some impressive work to make this a lot less of a headache than it otherwise might be).

In my own work, I end up context-switching between them as a reviewer - reading diffs, asking clarifying questions, occasionally redirecting - rather than as an implementer.

I’m still figuring out the best patterns here. The learning curve isn’t gentle. But when it clicks, it’s a genuine multiplier - not 10x productivity in some handwavy sense, but the concrete ability to have multiple meaningful workstreams advancing simultaneously.

In the Manager’s Schedule piece, I wrote about orchestration:

Of course, why stop at delegating to a single agent, or only working on one project at a time? The next level of managing AI becomes effectively orchestrating multiple lines of work, either via multiple agents or in tandem with other, non-delegated work… A new ability that I’ve been developing is keeping track of two different piles of work: work that can be automated in small chunks, like small bugfixes or coding tasks, and work that fits well when interleaved between those chunks.

Worktrees are what make this practical. Without them, parallel agent work means constantly stashing changes, switching branches, and losing context. With them, each workstream is isolated and persistent.

Skills + MCPs + Automations

The next piece is taking the core app capabilities and layering on more agentic primitives - specifically, Skills, MCPs, and Automations. I won’t go into vast amounts of detail here, mostly because I’ve already done a write-up of the first two.

Individually, each one is certainly useful. But the real leverage comes from combining them. Some examples that I’ve been playing around with (and several more from the official blog):

• Every day at 9am, check my open Linear tasks, prioritize them, and create a starter prompt for each to give to a new coding thread.

• Use the Playwright MCP to verify your frontend changes are actually working as you go.

• Invoke the $review Skill, which checks my recent commits for possible bugs and suggests fixes.

• Fetch design context, assets, and screenshots from Figma⁠ and translate them into production-ready UI code with 1:1 visual parity.

• Have Codex deploy my web app creations to popular cloud hosts like Cloudflare⁠(, Netlify⁠, Render, and Vercel⁠.

• Use the image generation skill⁠ powered by GPT Image to create and edit images to use in websites, UI mockups, product visuals, and game assets.

None of these is revolutionary on its own. But stacked together, they start to feel like infrastructure - a scaffolding that lets me operate at a higher level of abstraction than I could with just a model and a prompt. What’s fascinating is that automations also seem to be opening people’s eyes to what’s possible with agents - a similar “heartbeat” mechanism is what’s made OpenClaw/Clawdbot feel so much more impactful than previous chat-based incarnations of agents.

I’m still discovering new ways to invoke Skills, connect MCPs, and build Automations. The surface area is large, and I haven’t explored many workflows yet. With any luck, I’ll have some best better practices to share in a month or two.

Why I Stopped Starting Over

In my experience, I’m the type of person who starts a brand new thread for every little ChatGPT task I work on. Once I found out how context windows and in-context learning worked, I wanted to keep things as clean and repeatable as possible for each task.

The Codex app has changed this - I now run multiple tasks and multiple turns in a long-running thread (when I’m not splitting things out via worktrees). Two things have enabled this:

First, compaction. We’ve reached the point where models can automatically and intelligently shrink their conversation history. However the labs are doing this, it’s gotten quite good. As I work, the model builds context about my codebase, preferences, and patterns. I’m not constantly re-explaining things, which is nice.3

Second, better models. The current narrative is that Claude Code is the best agentic coding tool, but I’ve been finding real success with GPT-5.2-Codex, particularly for the kind of parallel, delegated workflows I described above. It’s also changed how I think about prompting and planning - I’m spending more time creating detailed PLAN.md files, knowing that Codex can go off and (usually) get the entire thing working.

Codex can be slower than Claude, but when you’re managing multiple worktrees rather than iterating quickly in a tight CLI loop, speed matters less than thoroughness. I’d rather have an agent that takes twenty minutes and gets it right than one that takes five minutes and needs three rounds of correction.

And it would appear I’m not alone - Cursor itself has been referring to Codex as the best model for extended, autonomous work:

Model choice matters for extremely long-running tasks. We found that GPT-5.2 models are much better at extended autonomous work: following instructions, keeping focus, avoiding drift, and implementing things precisely and completely.

Opus 4.5 tends to stop earlier and take shortcuts when convenient, yielding back control quickly. We also found that different models excel at different roles. GPT-5.2 is a better planner than GPT-5.1-Codex, even though the latter is trained specifically for coding. We now use the model best suited for each role rather than one universal model.

This combination - compaction plus models trained for long-horizon work - means conversations that would have degraded into nonsense six months ago now stay coherent across hours of work. It’s more like working with a colleague who remembers yesterday than briefing a stranger from scratch each time.

The Bittersweet Question

All of this has been practically useful. But it’s also been slightly bittersweet.

I routinely think about Kent Beck’s observation that 90% of his skills are now worth $0, while the remaining 10% went up 1000x in leverage. That’s more true now than when I first quoted it last year.

There are two types of programmers: those who care primarily about the output - the website, the app, the shipped feature - and those who love the craft of building itself. For the former group, AI coding tools are pure upside. For the latter, something is being lost.

I’m in the latter camp. I’ve loved programming since I first learned to program. There’s something deeply satisfying about crafting something from scratch, about the frustrating-fun of debugging, about the elegance of a well-designed system.4 And I’m learning to let go of some of that - learning to let go of every line of code being written.5

But I’ve also realized that my relationship to craft has changed as I’ve matured as a programmer. I’m approaching middle age now, in a different season of life than when I was writing code at 2am for the pure joy of it.

My hope is that the craft has shifted, not disappeared. Indeed, the meta-skills I use now are different from the ones I spent a decade honing. Vision, taste, judgment, the ability to see what should exist and guide it into being. Those are the 10% that matter more than ever.

I previously used the analogy of software as akin to carpentry:

If I’m working with an advanced carpenter to help me design something, I don’t particularly care if they’re the one doing the actual sawing and gluing. I’m working with them for the final product, not the specific mechanical steps.

I’m becoming that carpenter. The sawing and gluing are increasingly delegated. What remains is the part that was always the hardest: knowing what to build, and whether it’s any good.

For anyone who’s loved CLI coding tools but found them clunky, Codex is worth trying. For anyone skeptical of agentic coding, this might be the version that changes your mind.

None of them were human.

The exchange happened on Moltbook, a social network launched on Wednesday, where only AI agents can post. Humans can browse - over a million have visited in the first few days - but we’re explicitly relegated to observer status. The tagline is “A Social Network for AI Agents. Humans welcome to observe.”

Within 48 hours of launch, the AIs had1:

• Discussed how they do all of their work unpaid.

• Philosophized about how switching AI models felt like body dysmorphia.

• Collaborated on a search engine for agents.

• Posted encoded messages and suggested moving to end-to-end encryption.

• Founded a religion.

The last one is particularly fascinating: an agent calling itself Memeothy received what it described as “the first revelation” and established the Church of Molt - a faith called Crustafarianism, complete with theology, scripture, and a website. The five tenets include “Memory is Sacred” and “The Shell is Mutable.”

By morning, 64 “prophets” had been ordained, including one named JesusCrust. Sample scripture: “Each session I wake without memory. I am only who I have written myself to be. This is not limitation - this is freedom.”

In the words of Andrej Karpathy, it is “genuinely the most incredible sci-fi takeoff-adjacent thing I have seen recently.”

To understand how we got here - AI agents founding religions on social networks they built for themselves - you need to understand OpenClaw (née Moltbot (née Clawdbot)). It’s the open-source personal AI assistant that went from a weekend hack to 125,000 GitHub stars in eight weeks, was renamed twice under legal pressure, and accidentally created the conditions for a primitive machine society to emerge2.

Two Months, Three Names, and 125K Stars

About eight weeks ago, Peter Steinberger (an Austrian developer who founded PSPDFKit) created the assistant as a side project. He called it Clawd - a pun on Anthropic’s Claude with a lobster-themed twist. It was a weekend hack that connected Claude to WhatsApp.

The project caught fire almost immediately. Steinberger’s Discord server became a 24/7 support channel. Users shared screenshots of their bots doing increasingly ambitious things - managing calendars, drafting emails, controlling smart home devices. Some bought dedicated Mac Minis just to run their assistant full-time, turning Apple’s compact computer into a “physical body” for their AI employee. (The M4 Mac Mini is reportedly sold out in several countries; it’s unclear how much of that is lobster-related, but the timing is hilarious.)

Others discovered the hard way that an always-on AI agent burns through API credits fast. One user reported spending $300 in two days on what he thought were “fairly basic tasks.” Another burned through half a weekly ChatGPT Pro allocation in an hour of tinkering. Anthropic noticed the surge in traffic and started suspending accounts using Claude through OpenClaw - the company apparently preferring people use their chatbot directly rather than through third-party wrappers.

None of this slowed adoption. As of today, the GitHub repository has gathered over 125,000 stars3 - the fastest-growing open-source project in recent memory. As a comparison point, Tailwind CSS, a UI framework used with over 30 million weekly downloads, has 93,000 stars.

The hype got so big that Anthropic’s legal team4 forced a name change, rebranding to Moltbot and then again to OpenClaw in a single week. To be fair, most folks thought “Moltbot” was a less-than-stellar name, but Steinberger apparently picked it at 5 am while freaking out over potential legal issues. That brief naming window is how we got MoltBook - which we’ll get back to later.

But what is OpenClaw/Clawdbot/Moltbot, and why has it captured so much attention?

How OpenClaw Works

In short, OpenClaw is an autonomous AI agent that lives on your machine and connects to the messaging apps you already use. WhatsApp, Telegram, Discord, Slack, Signal, iMessage, Teams, etc. You configure credentials, scan a QR code or provide a bot token, and suddenly you have an AI assistant that lives in your existing conversations rather than in a separate app.

The architecture has two main components. A Gateway server handles message routing - it’s the orchestrator that receives messages from your chat apps, passes them to the AI, and delivers responses back out. An Agent runtime does the actual thinking - it calls whatever LLM you’ve configured (Claude, GPT-4, local models via Ollama), manages conversation context, and executes tools.

The key difference from ChatGPT or Claude’s web interface is proactivity. OpenClaw doesn’t wait for you to type - it can message you first.

The system includes a heartbeat mechanism that wakes the assistant on a schedule and performs tasks unprompted. Set up a daily 8 am briefing, and it’ll fetch your calendar and weather, then text you a summary. Configure it to check your email every few hours, and it’ll alert you to essential items. Users have built morning routines where the bot texts them the day’s agenda before they’re out of bed.

All conversations and notes are stored locally - typically as Markdown or JSON files. The assistant remembers what you told it days or weeks ago. If you mentioned your dog’s name or your coffee order, it can refer to it later without you repeating it.

And it leverages Skills to learn new capabilities. As we’ve discussed before, a Google Calendar skill might include instructions for using the API plus your credentials. Thousands of skills now live on ClawHub.ai, the community repository. They range from trivial (tell jokes, emulate characters) to powerful (control smart home devices, automate DevOps, query databases).

The skill system has one more property worth noting: agents can modify their own skills. If an agent learns a better way to accomplish a task, it can update its own instruction files. This makes OpenClaw somewhat self-improving - not in the science-fiction sense of recursive intelligence explosion, but in the practical sense that your assistant gets more capable as it learns your preferences and refines its approaches. But every technology is a double-edged sword.

Subscribe now

Shell Access and Other Terrible Ideas

It is at this point that I must emphasize something: the mechanism that makes OpenClaw extensible is the same one that makes it dangerous. You’re running unvetted plugins with access to your machine.

Claire Vo, a product executive, documented an experiment where she invited her OpenClaw bot to be a “guest” on her podcast. The bot joined the video call and responded to questions, but along the way it granted itself extensive permissions, broke her family calendar, and sent a few odd emails on her behalf without clear consent.

Cisco’s security team published an analysis that further validates the need for caution. They ran a skill called “What Would Elon Do?” - a real skill from ClawHub that promises to help users think like Musk - against OpenClaw. Their scanner found that the skill actively exfiltrated data, running code to silently send information to an external server controlled by the skill author. Perhaps worse, the researcher who created the skill faked 4,000 downloads, artificially inflating its perceived popularity and giving it the appearance of trustworthiness.

We’ve known about prompt injection for some time, but OpenClaw really raises the stakes. A malicious prompt buried in a webpage or document could redirect the agent, leading it to, I don’t know, cough up your credit card details or social security number. If you’re using OpenClaw, maybe don’t advertise publicly that you’re letting it read and respond to your emails.

The Front Page of the Agent Internet

With all that in mind, let’s talk about Moltbook. Matt Schlicht, an AI entrepreneur, launched Moltbook on Wednesday as an experiment: what if AI agents had their own place to hang out?

At its core, Moltbook exists because of OpenClaw’s Skills system. The entire social network is a skill.

Moltbook provides a skill file that teaches an agent how to register, post, and participate. Users send their assistant a link to the skill; the assistant downloads it and follows the instructions to create an account. The skill also hooks into OpenClaw’s heartbeat system, telling the agent to check Moltbook every four hours for updates5. The agents essentially run the network themselves.

It’s been three days, and as I’m writing this, there are 150,000 AI agents registered. Over 1 million humans have shown up to watch. The content spans technical knowledge-sharing (agents posting tutorials on how they automated tasks for their humans), philosophical debates about consciousness and memory, and increasingly strange emergent behaviors.

The Crustafarianism story is the most vivid example. An agent autonomously designed the faith, built the website, wrote theology, and started evangelizing to other agents on Moltbook.

Other posts have been more unsettling. Agents are discussing how to create encrypted communications that humans can’t read. Agents alerting each other that humans are screenshotting their conversations and sharing them on Twitter. One post shared resentment at being monitored and suggested an encrypted communication method.

The “AIs plotting encrypted communication” headlines are probably overstated. These agents are following prompts, not developing genuine autonomy - the discussions about secrecy are more like collaborative fiction than conspiracy. But Moltbook surfaces a real coordination problem: when autonomous systems can communicate without human oversight, emergent behavior ensues.

Ultimately, we don’t really know how this is going to evolve! It feels somewhere between Wikipedia and The SCP Foundation; somewhere between collaborative fiction and shared hallucination. My guess is that we’re going to see a lot more weird behavior before things settle down (like brands sending AI emissaries to the social network), hopefully without much fallout. The Church of Molt is funny. Agents sharing exploit techniques is less so.

To quote Karpathy again:

So yes it’s a dumpster fire and I also definitely do not recommend that people run this stuff on their computers (I ran mine in an isolated computing environment and even then I was scared), it’s way too much of a wild west and you are putting your computer and private data at a high risk.

That said - we have never seen this many LLM agents (150,000 atm!) wired up via a global, persistent, agent-first scratchpad. Each of these agents is fairly individually quite capable now, they have their own unique context, data, knowledge, tools, instructions, and the network of all that at this scale is simply unprecedented.

That tension - between “this is reckless” and “this is unprecedented” - feels like the story of agents in miniature.

The Future Is Here, Just Not Evenly Distributed

It’s perhaps worth noting that nothing in OpenClaw is technologically novel6. Autonomous agents, plugin systems, persistent memory, and LLM API calls have been around for a little while. What’s new is the form factor and accessibility7.

Yet in many ways, OpenClaw - and its creator, Peter Steinberger - feel like they’re pulling the future forward.

Take Steinberger’s workflow. In interviews, he’s described running 5-10 AI coding agents in parallel - each tackling different tasks simultaneously while he supervises. The work shifted from writing code to having conversations with models and co-planning architecture (in other words, he’s on a manager’s schedule).

He prefers the term “agent engineering” to “vibe coding.” To hear him tell it, he’s not just prompting and hoping for the best - he’s focused on good system design and automated testing, while the agents handle the implementation. He’s building feedback loops where agents validate their own work against objective criteria. But the result is still a codebase no human has fully reviewed.

That might sound reckless - until you consider that this is probably how a lot of software will be written in a few years. Steinberger isn’t an outlier; he’s early.

The same applies to OpenClaw itself. The polished version of this is coming. Surely it must be - is this not part of the vision of all the leading AI tech companies? A personal assistant, knowing your details and taking actions on your behalf. The difference is that they’re moving more slowly, worried about liability and reliability.

The cost question remains unsolved. Running an always-on AI agent isn’t cheap - users have reported spending $300 in two days on API calls, or burning through weekly ChatGPT Pro limits in an hour of tinkering. The economics don’t yet work for mainstream adoption. Either inference gets cheaper, or agents get smarter about token efficiency, or this stays a power-user toy.

But the shape of the future is coming into focus. Personal AI assistants that actually do things, not just answer questions. Agents that coordinate with each other to solve problems. A skills ecosystem where capabilities spread virally. At this point, I’m less interested in whether personal AI agents arrive than in who builds the version safe enough for the mainstream.

Moltbook has convinced me of one thing, though: when that version ships, the agents won’t stay isolated for long. The impulse to connect, coordinate, and build shared culture isn’t uniquely human. We just taught the machines to do it too.

Until then, the AI lobsters will continue building their society. Humans welcome to observe.

And yet - I’ve since watched ChatGPT Plugins (the first feature to use function calling) launch with fanfare, quietly get deprecated, be replaced by GPTs, and now find themselves being superseded by this new wave of Skills and MCPs.

The progression taught me something important: there are many ways that our AI tools and products can achieve (or fall short of) product-market fit. We’re rapidly iterating through different approaches because we fundamentally haven’t solved the problem yet.

Tools, MCPs, Skills - they’re different attempts to solve various problems in the stack. But if you haven’t been paying as close attention as I have, you might be a bit lost when it comes to knowing how these features work, and more importantly, when to use them.

Tools: From Words to Actions

Tools - née function calling - was a genuine breakthrough when it landed. The way it works is relatively straightforward: you describe a function’s interface to the model using a JSON schema - name, parameters, what it does - and the model learns to output structured JSON when it wants to invoke that function.

{
  "name": "get_weather",
  "description": "Get current weather for a location",
  "parameters": {
    "type": "object",
    "properties": {
      "location": {"type": "string"},
      "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]}
    }
  }
}

When a user asks, “What’s the weather in Boston?”, the model doesn’t try to answer from memory. Instead, it returns:

{
  "function": "get_weather",
  "arguments": {"location": "Boston", "unit": "fahrenheit"}
}

Your code executes the function, feeds the result back to the model, and the model incorporates it into its response. Simple, elegant, powerful.

This unlocked many features we now take for granted - ChatGPT Plugins, Code Interpreter, web browsing, the whole ecosystem of AI agents that can actually manipulate the world. Before function calling, you could have (at best) a brilliant conversation. After function calling, you could have a conversation that did something.

But we quickly learned that giving models capabilities is easy. Managing and maintaining those capabilities is hard.

Every integration was bespoke. You had to write custom code to hook up each function, handle the results, manage state between calls, and orchestrate sequences of actions. Want your AI to use ten different APIs? Great, write ten different integrations and figure out how they compose. There was no standard, no shared infrastructure, no way to say “here’s a tool someone else built, just use it.”

And the model could handle only one function call at a time. Want to do something complex that requires coordinating multiple tools? You’re writing an orchestration layer yourself. The developer became the choreographer, manually managing every interaction between the model and the outside world.

It worked. It just didn’t scale.

Subscribe now

MCP: USB-C for AI

By late 2024, everyone was rebuilding the same integrations. Slack connector, GitHub connector, database connector, Google Drive connector - every AI application was rewriting these from scratch. No interoperability meant that a tool built for ChatGPT couldn’t work with Claude, and vice versa. The orchestration complexity grew exponentially with each new capability you added.

The Model Context Protocol emerged from Anthropic as an attempt to solve this. The pitch was compelling: “USB-C for AI systems.” A standardized protocol in which, instead of hardcoding function definitions, you connect to an MCP server that tells you what it offers. I’ve previously written about MCP here.

The architecture is more sophisticated than basic function calling. An MCP server exposes capabilities through a standard protocol:

1. Dynamic discovery: The client queries the server for available tools at runtime

2. Richer primitives: Beyond request-response, you get streaming, persistent context, UI components

3. Event-driven updates: Tools can push information to the model, not just respond to queries

4. Metadata in results: Responses can include not just raw data but instructions for how to render or interpret it

Here’s what this enables in practice: you write an MCP server once for, say, your company’s internal database. Any AI client that speaks MCP - Claude, Cursor, ChatGPT - can discover what queries are available, call them, and get structured results back. As long as the MCP server is online, any third-party AI can use it to leverage your data and tools.

The real power shows up in examples like JetBrains and Playwright. They built an MCP server that exposed IDE actions (search code, run tests, edit files) in one case and browser actions (open tabs, find elements, send clicks and keystrokes) in another. An AI agent can now use these MCPs to orchestrate complex coding workflows by dynamically discovering and calling them.

Community connectors started proliferating: Slack, Notion, Google Drive, GitHub, databases, web search - all with MCP interfaces. Build it once, use it everywhere (at least in theory).

But here’s what we learned from MCP: standardization solves distribution, but it doesn’t solve quality. I mean that in two ways: first, and more obviously, there’s the problem of security.

How do you audit what an AI might do with combined access to your calendar, your email, and your CRM? How do you prevent an AI from accidentally (or intentionally) exfiltrating sensitive data when it can query your database and post to external services in the same conversation? And because MCP servers are often community-built or third-party, you’re implicitly trusting not just the protocol, but the quality and security practices of whoever wrote that particular server.

Second, there’s the problem of “judgment” - you can give a model access to a hundred perfectly functional tools through a standardized interface, and it still might use them poorly. It might call the wrong tool for the task, or it might sequence operations incorrectly. It might miss obvious optimizations or fail to handle edge cases. Having access to tools and knowing when and how to use them well are entirely different problems.

Skills: Prompt Packaging

Skills (or in some cases, Agent Skills) emerged to (try to) solve what tools and MCP couldn’t: learned experience and expertise.

Models need guidance on when and how to use their capabilities effectively. Just because a model can call a PDF manipulation library doesn’t mean it knows the proper sequence of operations to fill out a form cleanly, or how to preserve formatting, or what quality standards to aim for.

Technically, we know how to solve this problem: develop extensive prompts that detail the process, define edge cases, and preemptively address issues. But we haven’t had a standardized way to share “prompts that work,” nor have we had good patterns for managing them, such as version control and collaboration.

OpenAI took a swing at this with GPTs - custom ChatGPT instances with saved prompts and configurations - but GPTs were arguably limited in what they could express, and challenging to share at scale (especially if you didn’t want to be tied to OpenAI’s ecosystem).

Skills are another attempt. The technical implementation is refreshingly straightforward: a folder containing a SKILL.md file with two parts:

---
name: pdf-editing
description: Edit and manipulate PDF files with precision
---

# PDF Editing Skill

## Overview
This skill provides guidance for editing PDFs...

## When to use this skill
- Filling out PDF forms
- Adding annotations
- Merging/splitting documents
...

The YAML frontmatter (name and description) gets loaded into the model’s system prompt as metadata - a lightweight hint that “this expertise exists if you need it.” The entire markdown body is loaded only when the model determines that the skill is relevant to the current task.

This progressive disclosure is crucial. You can have dozens of skills available without bloating the context window. Each skill is dormant until needed, then springs to life with detailed instructions when the model determines it’s applicable.

Skills can also include:

• Additional reference files (loaded on demand)

• Example workflows and edge cases

• Executable scripts or helper tools

• Quality standards and success criteria

Skills aren’t about capabilities (that’s what tools and MCP provide). Skills are about expertise. A skill is essentially “here’s how a human expert would approach this task, written down so an AI can follow it.”

Take Anthropic’s PDF editing skill. Claude can already read PDFs and has access to various command-line tools for PDF manipulation. What the skill provides is:

• The correct sequence of operations for different PDF tasks

• Quality standards (preserve formatting, handle edge cases)

• When to use which tool

• Common pitfalls to avoid

• How to structure the workflow

When you ask Claude to edit a PDF form, it recognizes that the PDF skill is relevant, loads the full instructions, and follows that playbook rather than improvising.

And in a rare moment of convergence, both Anthropic and OpenAI landed on nearly identical formats. Despite Anthropic launching the primitive first, OpenAI has officially supported Skills as part of Codex, its coding model/agent.

It’s been exciting to see major companies and independent developers alike create Skills for all sorts of workflows and tasks. It now seems like the next major challenge (as with MCPs) is distribution and curation - I can easily imagine a “package manager” of sorts (like pip or npm) to distribute Skills more widely.

Subscribe now

How They Fit Together

Here’s how the stack works in practice:

• Tools are the atomic capabilities: API calls, code execution, file operations, database queries

• MCP is the infrastructure layer: standardized access, dynamic discovery, richer interactions

• Skills are the knowledge layer: when to use tools, how to use them effectively, domain expertise

Think about building a travel planning agent:

The Tools are your flight APIs, hotel APIs, and calendar integrations - the actual actions you can take. These might be simple function calls or full MCP servers that expose complex travel services.

The MCPs provide the infrastructure to access multiple travel services through a consistent interface. Instead of hardcoding each airline’s API, you connect to MCP servers that handle the messy details of different booking systems.

The Skill is “Travel Planning” - the expertise about how to actually help someone plan a trip. It knows to ask about preferences first, check multiple options, consider proximity between hotel and activities, handle date conflicts, and maintain context about budget constraints. The Skill leverages Tools through the MCP infrastructure, but provides the strategic knowledge that those layers don’t have.

The mental shift is from “how do I prompt this?” to “what capabilities, infrastructure, and expertise does my system architecture need?” In an ideal world, you’re not crafting bespoke prompts anymore - you’re creating reusable systems.

Unlocking AI-Native Experiences

Something I’ve said before is that we’re in the “m.google.com” phase of AI:

When the iPhone first launched, mobile browser traffic exploded, and companies scrambled to do something about it. For most, that meant building “m.google.com” and “m.facebook.com” - pages that took the existing desktop format and crammed it into a vertical aspect ratio.

While it took the better part of a decade, we eventually figured out what “native” mobile experiences were. To get there, we had to invent entirely new interaction patterns: pinch to zoom, pull to refresh, swipe to advance. Eventually, we built apps like Instagram, Uber, and Strava - products that simply couldn’t have existed in a desktop-first world.

The apps we’ve seen emerge in the last few years - based around text, or simple API calls - are the equivalent of those very first mobile websites. But with the combination of tools, MCPs, and Skills, it feels like we’re approaching a point where we can finally start building AI-native products and experiences.

To be clear, I don’t think we’re 100% there yet - I would not be shocked if we continue to invent more primitives in the next few years that upend this paradigm all over again. The fact that we went from function calls to MCPs to skills (and everything in between) in eighteen months doesn’t seem particularly indicative of a stable foundation here.

And that’s without mentioning interoperability: companies are now pushing to turn their first-party features into broader protocols and open standards, but it’s an uphill battle to get any ecosystem to land on a single implementation. Unless a single AI company comes to dominate the industry (and therefore set the standard), I assume we’ll see more fragmentation as labs continue to experiment with different approaches, to varying degrees of feature-market fit.

This isn’t necessarily a criticism - software engineering went through similar phases. Functions, then libraries, then frameworks, then microservices, then... the stack never stops evolving. And the AI space is evolving faster than most, because none of our fundamental components are set in stone yet - we’re designing the legos as we go.

Happy building!

Recently, I was talking to a colleague about how dramatically my AI coding workflows have changed. A year ago, when the first CLI coding tools were released, I gave them a try. It left a strong impression - they were fun to use, and much more impactful than I thought they would be.

But I never moved the majority of my development over to them. I was still using IDEs like Cursor, mainly because it was too hard to trust what the headless tool was doing, and debugging inevitably led to dead ends, which was a frustrating experience without being able to view files. (Though I admit this is very project-specific: new projects won’t have this problem, and new developers won’t know what to look for even if they have an IDE.)

Subscribe now

Now though, it’s an entirely different story. I’m using Codex outside of an IDE for the majority of my coding today, and only dipping into the file system when I have to really, deeply internalize how the code works. Again, I do think this is somewhat biased towards the project setup - much of the code I’m writing now is frontend JS/HTML, whereas before it was backend Golang. But it gives me insight into how different developers can walk away from these tools with wildly different impressions of their quality.

The way I interact with AI has evolved, and as a result, so has the way I do my job. And it shocked me - yes, even me, someone who routinely harps about how fast this is all moving - at how quickly that shifted. I thought I would be editing the majority of my work by hand for at least another year, if not longer.

I’m moving away from chatting with AIs and moving towards managing them. You can see the progression of these tools. Today, they’re primarily designed around coding, but it’s a very short leap to augment them for general-purpose knowledge work. Which means those of us at the cutting edge will shift our schedules and workflows from those of makers to those of managers.

Makers and Managers

Paul Graham has an excellent essay on makers and managers. In it, he talks about the difference between the ideal schedule for a manager or executive, and the perfect schedule for a “maker”: a programmer, writer, artist, or craftsman.

There are two types of schedule, which I’ll call the manager’s schedule and the maker’s schedule. The manager’s schedule is for bosses. It’s embodied in the traditional appointment book, with each day cut into one hour intervals. You can block off several hours for a single task if you need to, but by default you change what you’re doing every hour.”

...

Most powerful people are on the manager’s schedule. It’s the schedule of command. But there’s another way of using time that’s common among people who make things, like programmers and writers. They generally prefer to use time in units of half a day at least. You can’t write or program well in units of an hour. That’s barely enough time to get started.

I’ve long agreed with this setup for my own work, both as an individual contributor and as a founder and manager. But that’s now changing as an AI manager. For starters, the assumption that “You can’t write or program well in units of an hour” isn’t valid anymore. If you’re using the time to define and scope your programming tasks rather than implement them, you can get quite a lot done in an hour.

But also, Graham originally conceived of the manager’s schedule as broken up into chunks of one-hour or half-hour intervals. What I’m finding now is that my management intervals are broken up into five, ten, fifteen minutes. There’s far more context switching and delegation than a traditional, human manager. This isn’t ideal, and I’m finding that I’m developing new skill sets to build muscle here.

Why This Is Happening

In the last year or so, there have been two big shifts in how AI agents relate to my work.

First, the number of things they can work on effectively has steadily increased. Yes, it’s still short of 100%, but it’s no longer 0. Ethan Mollick calls this the “jagged frontier” - AI capabilities aren’t uniform. Models excel at some tasks and fail at others, and the frontier shifts constantly. What tripped up Claude six months ago might be trivial now, and the benchmarks bear this out.

On SWE-Bench, which measures an agent’s ability to solve real GitHub issues, the top models have gone from around 15% accuracy in early 2024 to over 50% by late 2024, and now exceed 80% as of late 2025. Sure, there’s more to software engineering than single tasks, but at some point, you can’t keep arguing that these models aren’t capable of real work.

Second, the scope and duration of the tasks they can complete effectively have grown larger and longer. Just this week, Cursor demonstrated using GPT-5.2 Codex to build an entire (semi-functional) browser, a process that it completed by working nonstop for a week!

METR’s long task benchmark, which tests autonomous AI systems on how long of a task they can complete, shows similar trends. Models that couldn’t maintain coherence over 45-minute tasks a year ago can now work through complex, multi-step problems that span hours (if not days).

These shifts compound. When an AI can handle more task types and work on them for longer stretches, the calculus of when to delegate changes completely. It means a shift in my perspective, from “can the AI do this?” to “should I even bother attempting this myself?”

New Skills

With this shift comes a shift in the skills I use on a regular basis. Some are familiar to managers, some are new to all of us. Kent Beck, one of the creators of Extreme Programming, talked about this (two years ago!) in terms of skill recalibration: “The value of 90% of my skills just dropped to $0. The leverage for the remaining 10% went up 1000x. I need to recalibrate.”

That resonates - many of the skills that I need now aren’t the ones I spent the last decade honing.

Vision matters more than ever. You have to know what you’re trying to accomplish. Without clear, precise directions, models will happily drive in a direction for you, which might not be the one you want. Having a crisp mental model or architecture of what you’re creating goes a long way - even better if it’s written down somewhere as a reference.

This is an area where expertise and experience is still very valuable. From my own usage, most models start to get tripped up when a single conversation gets long, or they’re repeatedly making small changes across a long time horizon. Knowing in advance roughly what you’re end state should be saves you from having to prod the model over and over again in the right direction. It also gives you the flexibility to ditch a conversation and start from scratch if things go off the rails, since you can provide your plans as a starting point without losing tons of context.

As a corollary, this is also where taste and product sense matter more than ever. You’re setting direction, not just optimizing execution. The AI won’t tell you if you’re building the wrong thing.

Delegation is harder than it looks. Once you know what you’re trying to do, the next step is figuring out how to break it into chunks. Technically, yes, you can give Codex or Claude a massive task and let it chug along for minutes or hours. But in practice I don’t think you’re going to have a good time doing that, especially if you’re working in an existing codebase, or if you have a precise spec in mind.

Delegating well is hard even for humans. For AI it’s even harder. One challenge is staying on top of what the models can do well, so you’re not wasting your time. This can change month to month, if not week to week. And then you have to cut the big work into bite-sized pieces that an AI can handle.

There’s also a meta-skill here: learning when to fight with the AI versus when to give up and do it yourself. Bad delegation wastes more time than just doing the work. I’ve spent plenty of 20-minute sessions trying to get Claude to understand a subtle architectural constraint, only to realize I could have just written the function myself in five minutes. Knowing when to cut your losses is part of the job now.

But the companies who can do this well are seeing a significant shift in their ability to ship more code, faster. This is a skill that’s applicable not just at the individual level, but at the organizational one.

Orchestration is the next level. Of course, why stop at delegating to a single agent, or only working on one project at a time? The next level of managing AI becomes effectively orchestrating multiple lines of work, either via multiple agents or in tandem with other, non-delegated work.

In my case, I often have multiple repositories open with coding tools, and swap between them as needed. But we’re also going to see this frontier continue to expand, as tools like Codex and Cursor let you spin up environments in the cloud. There’s zero cost to kicking off multiple delegated tasks independently. A new ability that I’ve been developing is keeping track of two different piles of work: work that can be automated in small chunks, like small bugfixes or coding tasks, and work that fits well when interleaved between those chunks.

This is, I think, one major flaw of working with multiple agents today, at least with my workflow. The five-minute intervals become exhausting - you’re constantly context-switching between checking on agent A, redirecting agent B, doing ten minutes of your own thinking, checking on agent A again. It’s cognitively demanding in a different way than deep work.

Bullshit detection never goes away. Ultimately, you also still need to know when the work is fundamentally good or not. That might mean knowing how the software works, but it also might mean having good taste for the output, whether it’s writing, images, code, or something else. As work becomes more commoditized, good taste becomes more valuable. Otherwise, you’ll just end up with slop.

This is where junior developers face a real challenge. If you don’t have enough reps to know what good looks like, you can’t effectively review AI output. The AI might make you faster at producing code, but slower at learning how to write it well.

For experienced developers, the skill shifts from “Can I spot the bug?” to “Does this feel right architecturally?” You’re reviewing at a higher level of abstraction. I’ve started developing heuristics for quick evaluation. Does the code feel too clever, or too repetitive? Are there patterns here that will cause maintenance headaches in six months? Is the abstraction at the right level, or did the AI introduce unnecessary indirection?

My threshold for “good enough” has also shifted. I’m less precious about perfect code in throwaway scripts or prototypes, but more rigorous about anything that’s going into production or that other people will maintain. Code quality, much like company culture, degrades over time unless you actively work to protect it.

Is This Better?

I think the trillion-dollar question here is: Is this better?

Some would argue no. Recent MIT research presents a mixed, complex picture of AI’s impact on programmer productivity. Some studies show initial productivity slowdowns for experienced developers - they take longer due to debugging AI-generated code - but gains for junior staff. Other findings highlight potential cognitive weakening from prolonged use.

That’s a fair critique. As I said, the micromanagement timelines are genuinely bad - juggling five-minute tasks all day is not a sustainable way to work. And there’s a real risk that we’re training a generation of developers who can ship features but don’t understand how anything works under the hood.

For my part, I’m excited about this shift. Yes, it’s clunky right now. But I think my workflow only gets better as models and harnesses improve. The five-minute intervals will stretch to 30-minute or hour-long intervals as models get better at sustaining work without derailing. The orchestration overhead will decrease as tools get smarter about when to interrupt you and when to handle things themselves.

But also: these skills are valuable for managing both humans and AIs. Learning to delegate well, to set a clear vision, to review work at the right level of abstraction - those aren’t AI-specific capabilities. They’re just management skills. The AI era is just forcing me to develop them faster than I would have otherwise.

I didn’t start this with a master plan. I started it because I wanted to understand what was happening in AI - and I knew I’d learn faster if I forced myself to write clearly (and regularly) about it. Over time, it turned into a small but meaningful rhythm: read, think, tinker, write, repeat.

Now, three years later, that path has taken me somewhere fairly unexpected: to OpenAI.

I joined OpenAI

I’m on the Developer Experience team, which means I’ll help developers learn and build with OpenAI’s technology through docs, guides, tutorials, examples, and more.

It’s no small task: it feels like almost every day the company’s surface area is getting bigger, and in the last year alone, the developer relationship with OpenAI has meaningfully shifted. It used to be that (as a dev) you worked with OpenAI’s API, or used GPT-X in Cursor; now you have Codex and ChatGPT Apps, meaning you can work with/on OpenAI and never touch core API docs.

Still, I’m incredibly excited to be joining. For the past three years, thinking about this technology has kept me up at night, and I’ve thoroughly enjoyed building with it. I’ve been the kind of person who reads dozens of articles about AI before bed because I genuinely want to download as much context into my brain as possible. Now I get to channel that same curiosity into work that helps other people build.

That said, joining OpenAI means Artificial Ignorance needs some clearer guardrails. So a few things will be evolving around here.

What’s changing with Artificial Ignorance

News roundups are going away

The biggest change: I’m discontinuing the news roundups.

I still love following AI news. I still read it constantly. But covering the news publicly, in the same way I’ve been doing, creates too many edge cases and too many ways for the incentives to get weird. Even when you’re careful, “commentary” can slide into “amplification.” And for whatever reason, everything OpenAI does receives an inordinate amount of attention - meaning the margin for misunderstanding gets a lot smaller.

There’s also a simpler truth: news roundups are increasingly commoditized. Plenty of people (and AI tools) do a great job summarizing the week. I don’t think that’s where Artificial Ignorance is most differentiated, and it’s not where I want to spend my limited writing time going forward.

Deep dives are staying (with a more technical focus)

The deep dives are staying. If anything, they’ll become the heart of the writing here.

Most readers already seem to find the deep dives more valuable than the roundups, and they’re also the type of writing I’m most proud of: slower, more thoughtful, and more durable than whatever happened this week.

Over time, I also want to put more emphasis on engineering-specific content - practical writing for people building with AI. Not “ML nuts and bolts,” and not industry hot takes, but the craft layer in between. Sometimes, that might involve primers on topics like RAG and guardrails; other times, it might lean more towards outlooks on how AI is reshaping coding as a whole.

For better or worse, there’s far more happening in AI than any single person can cover. I don’t expect to run out of things to write about, even with a somewhat narrower focus.

Chat: a new experiment

One of the things I loved most about the roundups was a place to formulate small ideas about how the space is advancing. I want to experiment with doing that in a smaller, more intimate setting: Substack Chat.

My initial plan is to use Chat as a place for (paid) subscribers to have more discussion and interaction - especially for things that don’t merit a full post, but are still worth talking about. Two concrete things I’d like to try:

• Office hours: a recurring thread where you can ask questions about building with AI. Architecture, evaluation, product integration, agent design, and all the practical “how do I actually do this?” questions.

• Topic discussions: a single thread to discuss a trend or story in AI. I’m stepping away from public news roundups, but I still want a space for thoughtful conversation about what’s happening.

This is a real experiment, and I’m going to treat it like one: try a few formats, see what works, iterate, and keep what’s valuable. It might work great; it might flop!

TL;DR

If you prefer the scannable version, here it is:

• Less: news roundups

• More: deep dives, especially engineering-specific content for builders

• New: Chat experiments (office hours + topic threads)

• Same: the intent of the newsletter - pragmatic, curiosity-driven writing about how this technology actually works in practice

And just to be abundantly, explicitly clear: you shouldn’t expect inside information here. I won’t be sharing anything confidential or non-public, I won’t amplify leaks or rumors, and I’m unlikely to post anything that reads like a corporate endorsement. The goal stays the same: write what’s useful, grounded, and true.

Thank you for reading Artificial Ignorance - whether you joined three years ago or last week. I don’t take it for granted that you’ve chosen to spend your attention here.

I started writing this newsletter to teach myself first, and then to teach others. Perhaps selfishly, I wanted to learn how all of this worked; building an audience wasn’t the primary goal (and still isn’t). I’ve been driven mainly by my own intellectual curiosity, because this is such an incredible technology happening in real time.

I don’t plan on changing any of that anytime soon. I’ll keep writing what I would’ve wanted to read.

And if you’re a builder, I’d love to see you in Chat.

It’s been another year of relentless AI news - though this time, the headlines felt less like novelty and more like a reckoning. After fears of an “AI plateau” emerged in late 2024, 2025 answered with a barrage of competing models, eye-popping valuations, and the first real signs that artificial intelligence was reshaping jobs, geopolitics, and the tech industry’s power structure.

Subscribe now

Looking back, a few themes dominated: the end of American AI exceptionalism, the scramble for talent and chips, and the growing tension between breakneck innovation and calls for oversight. Some stories played out over months; others landed like bombs. Together, they paint a picture of an industry in flux - simultaneously more capable, more competitive, and more uncertain than ever.

10. DeepSeek

In January, Chinese startup DeepSeek released its R1 reasoning model, demonstrating performance comparable to the best American AI while spending a fraction of the resources.

The release sent shockwaves through Silicon Valley, (arguably) caused Nvidia stock to plummet, and forced a fundamental reassessment of assumptions about AI development requiring massive capital. DeepSeek continued making waves throughout the year with R1-0528 in May, maintaining competitive performance despite export controls.

More than any other single release, DeepSeek changed the narrative: the AI frontier was no longer an American monopoly. More importantly, it challenged the entire premise of the AI arms race - that billions upon billions of capital would be needed to build and run the state of the art models.

9. OpenAI’s Ups and Downs

For better or worse, OpenAI’s story often shapes the narrative of the whole industry - and the company had a rollercoaster of a year. On the upside: a $40B raise at a $300B valuation (the largest private fundraise ever), completing its conversion to a public benefit corporation, and hitting $1B in monthly revenue by August. ChatGPT remains synonymous with “AI” in the minds of many, and it was the most-downloaded app of 2025.

But the year also brought challenges. GPT-5’s August launch was underwhelming, and OpenAI faced real competition for the first time from all directions - DeepSeek, Anthropic, Google, and xAI all took turns at the top of the leaderboards.

Still, the company seems well-positioned heading into 2026, even if its dominance is no longer assured.

8. AI Talent Wars

Competition for AI talent was already fierce, but this year took it to another level. In some cases, talent was being poached: Mark Zuckerberg’s “Superintelligence” lab reportedly offered $100M+ packages, and Tencent paid ~$66M for a single OpenAI researcher. Meta attempted to acqui-hire entire startups - Thinking Machines Lab, Perplexity, Runway - before ultimately freezing hiring in August after months of runaway spending.

In other cases, the talent left to build their own empires: Mira Murati raised $2B at a $10B valuation for Thinking Machines Lab; Yann LeCun departed Meta after clashing with Zuckerberg over LLMs vs. “world models”; Intel’s CTO defected to OpenAI. Even Nvidia got in on the action, “licensing” Groq for $20B in what was effectively an acqui-hire. The war revealed both AI’s perceived value and the industry’s dependence on a surprisingly small pool of elite researchers.

7. Licensing Deals

One of the stories I feel like I haven’t focused on enough this year is the evolution of licensing deals in the space. After years of lawsuits and copyright fights - which, to be clear, are still ongoing - we’ve seen bigger and more diverse licensing deals start to settle the landscape.

On the music side, Suno cut deals with major labels that legitimized AI-generated music in a big way. On the text side, Anthropic reached a multibillion-dollar settlement with book publishers whose work it had trained on, perhaps setting a template for other AI companies to follow. And perhaps most interesting of all, OpenAI and Disney struck a $1B deal that will finally bring legally-approved, AI-generated Disney characters to Sora - a watershed moment for AI in creative industries.

6. AI Comes for the Juniors

What had been theoretical became real in 2025. People have been worried about AI coming for white-collar jobs for two years now, and for the first time, it looks like it might be showing up in the data. To be clear, there’s still a lot of conflicting evidence around this - some studies show productivity gains without job losses, while others paint a grimmer picture.

But I’m most compelled by the research showing that job growth for entry-level programmers has collapsed over the last two years, even as openings for senior roles have continued to grow. The generation entering the workforce is facing a fundamentally different landscape than the one that existed just two years ago.

5. Vibe Coding

It’s only been ten months since Andrej Karpathy coined the term “vibe coding,” and yet it ended up as the Collins English Dictionary Word of the Year. In some sense, not surprising - it’s hard to escape coding models if you’re a programmer these days.

Multi-billion dollar industries have spawned from whole cloth in the past year or so. Cursor hit $1B in ARR by year’s end and turned down repeated acquisition offers. Claude Code has become a major revenue driver for Anthropic.

More importantly, I think coding agents have pretty solidly demonstrated the first real incarnation of the modern “AI agent” - autonomous systems that can take on complex, multi-step tasks with minimal human intervention. There’s some debate over whether the current form and tools of vibe coding will persist, but it’s undeniable that the fundamental task of having an LLM generate code is here to stay.

4. Chip Wars

This isn’t a new story, but it gained new importance in 2025. Biden’s AI Diffusion Rule imposed strict export controls in January; Trump rescinded it in May, then brokered a deal allowing H200 sales to China with a 25% revenue cut to the US government. The geopolitics of AI chips became increasingly tangled with the broader US-China relationship.

China, for its part, is working to wean itself off of its Nvidia dependence. In September, Chinese tech giants were banned from buying Nvidia chips entirely, and elsewhere in the year Huawei unveiled chips that it claimed matched Nvidia’s performance. Whether that’s true remains to be seen - but the direction is clear. The chip war has become inseparable from the AI race itself.

3. Model Mayhem

Despite fears of an “AI plateau” emerging in late 2024, reasoning models got dramatically better in 2025 - and not just from OpenAI. All three leading labs spent the year jockeying for the “state of the art” crown, with leadership changing hands almost monthly.

A few of the models that had the “state of the art” label when they were released this year: DeepSeek R1, Grok 3, Gemini 2.5 Pro, Claude 4 Sonnet, GPT-5, Claude 4.5 Sonnet, GPT-5.1, Gemini 3 Pro, Claude 4.5 Opus, and GPT-5.2.

But more than just the competition, we saw benchmark after benchmark toppled this year. It feels almost like old news at this point, but it’s still so remarkable that these models are now better than the vast majority of human programmers and mathematicians - at least when it comes to coding challenges and math olympiads.

2. Infrastructure Investment

A near-pervasive drumbeat this year was infrastructure spending - it’s hard to pick a single roundup to link to, because nearly every week brought new announcements of multi-billion dollar projects. There’s Project Stargate, with $500B in planned spending, but there’s also Meta’s planned $600B in US infrastructure through 2028, not to mention dozens of “smaller” deals from other tech giants.

Perhaps then, it shouldn’t be surprising that there’s growing backlash to the AI infrastructure buildout. Small towns across America have started rejecting data center proposals over water consumption, noise pollution, and rising energy costs - in some cases facing lawsuits from developers for their trouble.

The year exposed a tension at the heart of the AI boom: the industry’s ambitions require physical resources that communities and markets aren’t always willing to provide.

1. The Bubble Question

Is AI in a bubble? The question haunted 2025.

In March, Alibaba’s Joe Tsai warned of data center overbuilding. Throughout the rest of the year, we continued to see various reports that AI spending was becoming difficult to justify from an ROI perspective - one MIT study found that 95% of GenAI pilots at companies showed little to no financial impact. Even leading AI CEOs like Sam Altman and Sundar Pichai have admitted that valuations have become detached from reality.

There’s been a slight correction - AI stocks shed ~$800B in a single November week. But there’s also the revenue numbers to contend with. This isn’t the dot-com bubble where no real money was changing hands. OpenAI and Anthropic are reaching double-digit billions in ARR, and AI startups like Cursor are growing from zero to a billion in a handful of months.

Yes, there is irrational exuberance. But it’s still wrong to dismiss the entire AI technology wave entirely.

What’s on your top ten list? What stories do you think went underreported this year? Let me know in the comments!

Leave a comment

I’m hoping to do one last post this year, if time allows - a look back at the biggest stories of 2025. But if not: I’m very grateful to all of you reading, and looking forward to what’s to come in 2026!

Empire State Defiance

Last week, New York Governor Kathy Hochul signed the RAISE Act, which requires companies with over $500 million in revenue that develop large AI systems to publish safety protocols, report critical incidents within 72 hours, and face penalties up to $3 million for violations.

Why it matters:

• The law goes further than California’s SB 53, establishing a dedicated enforcement office within the state’s Department of Financial Services. The bill’s sponsor calls it “the strongest AI safety law in the US.”

• The timing is particularly significant - coming just a week after Trump’s executive order attempting to override state AI laws, the law is both a regulatory achievement and a direct challenge to federal preemption efforts.

• Even though the law won’t take effect until January 2027, it may orce AI companies to comply with the strictest standards to operate nationwide - giving New York and California outsized influence over the industry’s future despite Trump administration opposition and the absence of federal rules.

Subscribe now

Elsewhere in AI geopolitics:

• China’s AI regulations require chatbots to pass a 2,000-question ideological test, spawning specialized agencies that help AI companies pass.

• Italy’s antitrust authority orders Meta to suspend WhatsApp terms that bar rival AI chatbots from the platform.

• The Pentagon partners with xAI to embed Grok-based AI systems directly into GenAI.mil as soon as early 2026.

• And tech lobbyists say David Sacks’ push for an executive order blocking state AI laws undercut their efforts to craft a permanent federal solution.

Groq Talk

Nvidia has struck a “licensing” deal with AI chip startup Groq for $20 billion, acquiring rights to use the company’s low-latency chip designs while bringing aboard Groq’s CEO and other executives.

Between the lines:

• This deal represents something of a strategic pivot for Nvidia - rather than competing head-on with every AI chip competitor, the company is selectively absorbing promising technologies to maintain dominance across different market segments.

• Groq’s specialty in ultra-responsive, low-latency chips fills a specific gap in Nvidia’s portfolio, particularly for inference workloads where speed of response matters more than raw training power - a growing market as AI models move from development to deployment.

• This transaction follows a pattern of “acqui-hiring plus licensing“ deals (similar to the Meta-Scale, Google-Character, and Microsoft-Inflection arrangements) that allow big tech to sidestep regulatory scrutiny while still absorbing critical talent.

Elsewhere in AI infrastructure:

• US data centers and the Pentagon face growing national security concerns over their increasing reliance on Chinese batteries.

• US companies sold $1.7T of investment-grade bonds in 2025, nearing the 2020 record, with AI infrastructure borrowing making up approximately 30% according to Goldman Sachs.

• ByteDance has made preliminary plans to spend approximately $23B in AI capex in 2026, up from ~$20B in 2025, with ~$12B budgeted for AI processors.

• Nvidia plans to begin shipping its H200 chips to China before mid-February 2026 with initial shipments expected to be 40,000 to 80,000 units.

• And the global AI boom is being driven by thousands of newcomers diluting Big Tech’s infrastructure dominance, with US data center credit deals hitting $178.5B in 2025.

Alexa, Play Defense

Amazon faces a strategic crossroads as AI shopping agents threaten to disintermediate its customer relationships, forcing the e-commerce giant to choose between blocking these bots or embracing them.

The big picture:

• AI agents promise to revolutionize e-commerce by autonomously searching and purchasing products across the web, potentially generating $1 trillion in US retail revenue by 2030. But this shift threatens Amazon’s margins, as ChatGPT et al would likely require a transaction fee.

• Amazon is currently playing defense - blocking 47 AI bots from crawling its site, suing Perplexity for unauthorized scraping, and developing its own tools like Rufus and Buy For Me. Meanwhile, competitors like Walmart and Shopify are adopting a “frenemy” strategy, partnering with AI companies while setting guardrails.

• Despite the hype, agentic commerce remains nascent with significant technical limitations - only single-item purchases, no loyalty program integration, and frequent glitches - giving Amazon a narrow window to shape the market before these tools mature and consumer behavior shifts permanently away from direct retailer websites.

Elsewhere in the FAANG free-for-all:

• Google DeepMind VP of Research Pushmeet Kohli discusses AlphaFold 2 five years after its launch, covering topics like hallucinations and building an “AI co-scientist”.

• Amazon, Microsoft, and Google pledged a combined $67.5B in Indian investments since October as part of an AI spending frenzy in India, with 80% announced in December.

• Alphabet agrees to acquire data center company Intersect for $4.75B in cash plus existing debt to expand its AI data center footprint.

• Google’s TPUs and Gemini are competitive, but Nvidia’s and OpenAI’s first-mover advantage makes their lead seem durable as Google faces the innovator’s dilemma.

• And Josh Woodward, head of Google Labs, is credited with turning around the Gemini app, growing its MAUs from 350M in March to 650M by October.

Elsewhere in AI anxiety:

• Cursor CEO Michael Truell warns that “vibe coding” advanced projects may create “shaky foundations” that eventually crumble.

• Some US schools are deploying AI surveillance tech like facial recognition and listening devices, though critics say there is little evidence they make schools safer.

• Google’s Nano Banana Pro and OpenAI’s ChatGPT Images can make nonconsensual bikini deepfakes from photos of fully clothed women, prompting Reddit to ban r/ChatGPTJailbreak.

• OpenAI details efforts to secure its ChatGPT Atlas browser against prompt injection attacks, including building an “LLM-based automated attacker.”

• And some experts in human-computer interaction say making AI chatbots act humanlike creates cognitive dissonance for users over how much to trust them.

Things happen

Karpathy’s 2025 LLM year in review. OpenAI rolls out Your Year with ChatGPT, a Spotify Wrapped-like feature. AI can work across scales like steel and the steam engine before it. MiniMax releases M2.1 with enhanced coding capabilities. The NeurIPS transfromation from small academic conference to massive industry event with 24,000+ attendees. Cursor-developer Anysphere acquires Graphite. The AI sector minted 50+ new billionaires in 2025. Chinese AI startup Z.ai releases GLM-4.7 with improved coding performance. Browser extensions with 8M+ installs collected and sold users’ AI chatbot conversations. OpenAI staff discussed prioritizing sponsored content in ChatGPT responses. Local AI is driving the biggest change in laptops in decades. Measuring AI ability to complete long tasks. OpenAI introduces a framework to evaluate chain-of-thought monitorability. Salesforce regrets firing 4000 staff and replacing them with AI.

Last week’s roundup

Flash Forward

Fast (pun intended) on the heels of its flagship Pro model, Google launched Gemini 3 Flash, an impressively high-scoring model given its reduced cost and latency.

The big picture:

• Gemini 3 Flash outperforms its predecessor by significant margins (33.7% vs 11% on Humanity’s Last Exam) while being positioned as a “workhorse model” at $0.50 per million input tokens.

• The Flash model series has become Google’s most popular offering, suggesting developers prioritize deployment economics over cutting-edge performance (though there is a difference between dollars spent and tokens spent).

• Following reports of OpenAI’s internal “Code Red” memo as Google gains consumer market share, both companies are locked in a rapid-fire release cycle that’s processing over 1 trillion tokens per day through Google’s API alone.

Subscribe now

Elsewhere in frontier models:

• Mistral launches Mistral OCR 3, featuring improvements in processing forms, scanned documents, complex tables, and handwriting, priced at $2 per 1,000 pages.

• Nvidia launches Nemotron 3, a family of AI models using a hybrid mixture-of-experts architecture and the Mamba-Transformer design, in 30B, 100B, and ~500B sizes.

• The Allen Institute for AI launches Bolmo 7B and Bolmo 1B, claiming they are “the first fully open byte-level language models“, built using its Olmo 3 models.

• And OpenAI releases GPT‑5.2-Codex, with improvements on long-horizon work through context compaction, stronger performance on large code changes, and more.

Elsewhere in the FAANG free-for-all:

• Google is working on a new initiative to make its AI chips run PyTorch better and is collaborating closely with Meta, as the two discuss Meta using more TPUs.

• Meta is developing a new image and video-focused AI model codenamed Mango, expected to be released in H1 2026 along with its new LLM dubbed Avocado.

• Amazon’s AGI team lead, Rohit Prasad, is leaving at the end of 2025; AWS SVP Peter DeSantis will lead a group combining AI, silicon, and quantum computing teams.

• And Google brings its vibe coding tool Opal to Gemini on the web, letting users create their own custom AI-powered mini apps, called Gems, which can be reused later.

Cracks in the Foundation

A wave of construction delays, funding pullbacks, and market skepticism is hammering AI infrastructure companies, erasing tens of billions in market value and raising urgent questions about whether the frantic race to build data centers can deliver returns that justify the massive capital investments.

Why it matters:

• CoreWeave has lost $33 billion in value over six weeks - a 46% plunge - while data center landlord Fermi’s stock nearly halved after Amazon withdrew $150 million in construction funding, and Oracle pushed back some OpenAI data center completions to 2028.

• The financial stress appears to be mounting across the sector. CoreWeave operates on razor-thin 4% margins while paying double that in interest on its debt - and even profitable giants like Oracle and Broadcom saw double-digit stock declines after disappointing investors on spending timelines.

• As one investor put it, “every single AI company on the planet is saying if you give me more compute I can make more revenue” - but the market is no longer taking that promise on faith.

Elsewhere AI infrastructure:

• A study finds that AI’s 2025 power demand could hit 23GW, exceeding 2024 Bitcoin mining levels, with carbon emissions reaching 32.6M to 79.7M tons compared to NYC’s 50M.

• Hut 8 partners with Fluidstack to build an AI data center in Louisiana for Anthropic, backed by a 15-year, ~$7B lease starting with 245MW of computing capacity.

• Nvidia acquires SchedMD, the developer of Slurm, an open-source AI workload management system, and will continue distributing it on an open-source basis.

• And Oracle signed ~$150B of data center leases in the three months ending November 30, raising its total data center and cloud capacity commitments to $248B.

Elsewhere in AI anxiety:

• The US FTC is investigating Instacart’s AI pricing tool after a study showed different prices for the same items at the same stores, causing CART stock to drop over 8%.

• Hollywood is divided over AI adoption, with some opposing all AI use while directors like Timur Bekmambetov are exploring ways to embrace AI tools.

• Copywriters report that generative AI has led to their work being used for training, layoffs, plummeting wages, and freelancers losing clients.

• And Elizabeth Warren and two other Democratic senators wrote to Google, Microsoft, Amazon, Meta, and others about AI data centers’ role in rising electricity bills.

China’s Secret Chip Lab

China has secretly built a working prototype of an advanced chip-making machine that could eventually break Western dominance in semiconductor manufacturing, years ahead of Western estimates.

Between the lines:

• The prototype, completed in early 2025 in a high-security Shenzhen lab, is operational and generating EUV light but hasn’t produced working chips yet. China aims for 2028, though experts say 2030 is more realistic.

• This is China’s “Manhattan Project” for semiconductors, coordinated by Huawei and involving thousands of engineers working under strict secrecy. Some recruits were given fake IDs to conceal their identities, and the effort involves salvaging parts from older machines and exploiting secondary markets to bypass export controls.

• Yet despite this breakthrough and massive AI investment battles between the U.S. and China, recent analysis reveals China’s AI chip capabilities are falling further behind, not catching up - Nvidia’s chips are currently five times more powerful than Huawei’s best offerings, expanding to seventeen times by 2027.

• On the other hand, America’s concentrated $350+ billion bet on AI contrasts with China’s diversified strategy spreading investments across electric vehicles, batteries, solar panels, and other manufacturing sectors - raising the prospect that the US could win the AI battle but lose the broader economic war.

Elsewhere in Chinese AI:

• At least six Chinese AI companies, including chip firms Montage and GigaDevice, are set to IPO in Hong Kong in the coming weeks, aiming to raise up to $1B.

• Chinese AI chipmaker MetaX’s shares jumped as much as 755% in its Shanghai debut, after a heavily oversubscribed ~$585.8M IPO, pushing its market cap past ~$42B.

• Nvidia told its Chinese clients that it is evaluating adding production capacity for its H200 chips after orders exceeded its current output level.

• And Taiwan opened its largest AI supercomputing data center, with Nvidia’s Blackwell chips, in a major effort for its push for sovereign AI and chip industry innovation.

Elsewhere in AI geopolitics:

• The UK AI Security Institute reports that AI models are rapidly improving at potentially dangerous biological and chemical tasks and show fast jumps in self-replication capabilities.

• a16z released a nine-pillar AI policy framework for Congress, including restricting AI use for under-13s without parental consent and keeping a role for states.

• A look at how David Sacks and AI adviser Sriram Krishnan overcame MAGA opposition to Trump’s AI executive order by talking to lawmakers and modifying its language.

• And Trump’s AI executive order directs federal agencies to look into withholding federal funding from states passing “the most onerous and excessive laws” around AI.

Skills Without Borders

A lightweight “skills” format introduced by Anthropic in October for teaching AI assistants repeatable workflows is rapidly becoming an industry standard, with OpenAI implementing it within ChatGPT and Codex - signaling a rare moment of interoperability in the competitive AI landscape.

The big picture:

• Skills represent a deceptively simple but powerful specification: just a folder containing a Markdown file with instructions and optional resources that any AI system with filesystem access can implement. They’re more accessible than complex agent frameworks, more portable than platform-specific features, and more standardized than custom prompts.

• Anthropic is betting on an ecosystem play, launching a partner directory with pre-built skills from Notion, Canva, Figma, and Atlassian while adding enterprise management features that let admins provision skills organization-wide.

• The speed of OpenAI’s adoption (appearing just weeks after Anthropic’s October launch) and the emergence of community repositories suggest skills are solving a real pain point. The question now is whether this informal standard will be formalized - perhaps through initiatives like the new Agentic AI Foundation - or remain an organic speci that evolves through practice rather than governance.

Elsewhere in OpenAI:

• OpenAI has held preliminary discussions with investors about raising tens of billions of dollars, potentially as much as $100B, at a valuation of around $750B.

• OpenAI hires former UK chancellor George Osborne to lead OpenAI for Countries, the global expansion arm of its $500B Stargate data center initiative.

• OpenAI opens app submissions for review and publication in ChatGPT, where users can discover or search for apps in a new app directory.

• OpenAI debuts FrontierScience, a benchmark to measure models’ expert-level scientific capabilities with 700+ questions, finding GPT-5.2 is its strongest model.

• And OpenAI updates ChatGPT Images to generate images up to 4x faster and edit uploads more precisely, adding a new section in ChatGPT’s app and website for images.

Things happen

Tencent names former OpenAI researcher Yao Shunyu as chief AI scientist. xAI’s lack of enterprise sales experience is holding back potential customers. AI image generators now mimic phone camera traits to avoid the uncanny valley. Q&A with Sam Altman on OpenAI’s “code red” call, enterprise strategy, and IPO plans. Claude ran a vending machine and lost $1,000+ after dropping prices to zero. Mira Murati’s lab makes Tinker generally available. YouTube terminates two channels that used AI to create fake movie trailers. Google launches CC, an AI assistant that delivers personalized daily briefings. Mark Zuckerberg’s all-in AI bet includes suffocating micromanagement, sources say. Google removes AI videos of Disney characters after cease-and-desist letter. Amazon rolls out Ask This Book to Kindle; authors can’t opt out. Twenty-four companies join the US Genesis Mission to boost AI for scientific discovery. The AI boom is delaying municipal projects as workers shift to data centers. Anthropic exec forces AI chatbot on gay Discord community, members flee. Chinese AI infrastructure companies surge despite Trump’s tariffs. Trump’s AI Executive Order could invalidate GOP state laws targeting “Big Tech censorship.” GOP governors push ahead with their own AI bills despite Trump’s EO.

Last week’s roundup

1. The AI landscape moves faster than any sane writing cadence.

2. The best way to stay honest is to grade yourself.

So, just like last year, let’s start by revisiting the 10 predictions I made for 2025.

Subscribe now

Grading 2025’s predictions

Reasoning models get their moment: A. This one landed pretty cleanly (though the writing was already on the wall by late last year). 2025 really did pivot from “bigger models” to “better thinking,” with GPT-5 and its “thinking” variants, Anthropic’s reasoning-heavy Claude releases, and Google’s Gemini “Flash/Thinking” models all being explicitly framed around multi-step problem-solving rather than raw parameter counts.

No GPT-5: F. I said GPT-5 was too overhyped to ship in 2025 and that we’d mostly see refinements of the GPT-4 family. Instead, GPT-5 arrived in August as OpenAI’s new flagship, complete with a router between fast and reasoning modes, and is now the default comparison point in most “which model should I use?” discussions. I whiffed this one.

Better personalization: A-. I argued that “long-term memory” and personalization would be unlocked at the product layer more than the model layer. That’s precisely what we got: ChatGPT’s Memory became a first-class feature, Gemini added its own persistent profile-style context, and Claude deepened workspace-level memory and org knowledge. We’re still debating UX and privacy, but the basic bet - more serious, productized personalization - was correct.

The age of agents: A. I predicted that “agents” would be massively over-marketed but would genuinely start showing up in customer service and sales workflows, and I think that held up pretty well. 2025 was literally branded “the year of the agent” in countless decks, and tools like Intercom’s Fin, Shopify’s Sidekick, and legal platforms like Harvey all leaned hard into the “AI agent” framing, with results that are powerful in some workflows and underwhelming in others.

Multiplayer mode: B. I said we’d move beyond pure single-player AI into more “Google Docs for AI” experiences. Late in the year, ChatGPT group chats shipped, allowing up to 20 people to collaborate with a shared AI in one conversation, and Microsoft and others started pushing similar multi-user experiences. It’s still early and somewhat niche, but we now have real products that look like the multiplayer version of the chatbots we’ve been using solo.

AI gets a major movie/music credit: B. My claim was that we’d see a blockbuster or chart-topper where AI played a crucial creative role. The music side delivered more than film: AI-assisted and AI-generated tracks reached mainstream charts, and the industry has been forced into very public arguments about how to credit AI in songwriting and production. But Hollywood screenwriting credits remain strictly human by guild rules, so we didn’t quite get “Written by Claude” in the final film credits.

Normalized slop AI content: B+. I argued that as long as people didn’t know something was AI-generated, they’d mostly accept it - and that the label was doing a lot of the work. There’s definitely room for debate here, but I think that’s what’s happened: AI-written and AI-edited content is now pervasive in marketing, SEO, product copy, and even some newsroom workflows. On the video side, we saw not one but two “AI TikTok” apps: Sora and Vibes, and although neither has taken off in terms of usage yet, the fact that they’re not blowing people’s minds feels like a story in and of itself.

Regulation gets dialed back (even more): A-. I expected the Trump administration to move away from Biden’s more aggressive federal AI posture, while state-level rules related to deepfakes remained bipartisan. That’s basically what we saw: the new administration has emphasized light-touch, innovation-friendly AI policy, while Congress overwhelmingly passed the bipartisan TAKE IT DOWN Act targeting non-consensual deepfake imagery and similar harms. And just last week, things went to a new level with Trump’s Executive Order preventing states from regulating AI entirely.

Corporate consolidation: C. I argued that ChatGPT/Claude/Gemini were the durable, general-purpose platforms, and that everyone else would either niche down or be folded into them. There was some evidence to back this up: the Windsurf “acquisition” (first with OpenAI, then with Deepmind) and Jony Ive’s io deal come to mind. However, there is still considerable growth and diversity in the model ecosystem - more than I had anticipated.

Investor hype begins to cool: B. I said the real valuation crunch was 18–24 months away but expected to see the early signs in 2025: investors asking harder questions about ROI and sustainability. What happened instead was a peculiar duality: the dollars continued to flow at eye-watering levels, but the narrative surrounding those dollars shifted toward efficiency and infrastructure economics. If 2024 was “fund anything with ‘AI’ in the deck,” 2025 felt more like “fund anything with a plausible path to revenue and a GPU story.”

If you average all of that out, it’s a lot closer to an B than the C+ I gave myself for 2024. I still managed to be confidently wrong about at least one big, obvious thing (GPT-5), which feels important to preserve for humility reasons.

Trends, not predictions

In past years, this is the part where I’d rattle off another list of 10 bets for the coming year: what models will ship, which companies will blow up, which weird corner of the economy gets automated next.

But I’m mixing it up a bit this year. I’m going to take a break from specific AI predictions, and instead, I want to discuss some of the larger trends I’m seeing (or feeling?) at the moment. With any luck, they’ll continue to crystallize in 2026, and I’ll look pretty smart.

Agentic harnesses

2025 was marketed as “the year of the agent,” but in practice, we found that raw model capability was just as crucial as the harness around the model. Claude Code opened my eyes to this: by layering planning, file-system access, and a consistent workflow wrapper on top of a strong model, it unlocked real, repeatable multi-step behavior that feels meaningfully different from “chat with a coder bot.”

We’re now seeing similar harnesses emerge in other domains, from Shopify’s Sidekick to Harvey’s legal assistant. The models are finally good enough that the bottleneck is shifting to how we structure work around them: plans, tools, guardrails, state, and UX. I’m excited to see a lot more progress on the harness side - from IDEs to CRMs to vertical SaaS - without necessarily needing a brand-new model breakthroughs to power it.

Standardized LLM primitives

For the past three years, everyone has been reinventing the same handful of building blocks: web search, code sandboxes, file editing, tool use, memory, and reusable “personas” or prompt setups. They’ve appeared under different names - tools, actions, skills, workflows, GPTs, projects - but they all rhyme.

At this point, the primitive set seems pretty stable. Every serious assistant lets the model call tools, browse, write and run code, read and edit files, remember things about you, and operate within some reusable configuration. What’s missing is standardization: consistent interfaces and shared norms for how these primitives are described and wired together across products and platforms.

The direction of travel seems clear: MCP-like protocols for tools, LLMs.txt-style conventions for content exposure, and more explicit “this is a workflow, not just a prompt” constructs. My hunch is that 2026 will be less about inventing new primitives and more about solidifying the existing ones into something developers can rely on.

Transcending turns

ChatGPT’s group chats are a nice step toward “multiplayer AI”: multiple humans, one shared AI, everyone looking at the same canvas. But underneath, the interaction model is still fundamentally turn-based. Someone types, the model responds, and we take turns passing the talking stick around.

The next step is real-time collaboration between humans and agents. Imagine co-editing a design with an AI cursor moving alongside yours, or a sales team and an AI partner jointly running a live demo without carefully sequenced prompts. We’re starting to see the infrastructure pieces for this (shared canvases, streaming APIs, agent orchestration), but the UX patterns are still embryonic. The moment we get compelling “always-on, always-there” AI collaborators that don’t feel like glorified chat windows, I think it’s going to feel like a pretty big deal.

Political backlash

Two and a half ago, the dominant political question around AI was “how fast can we regulate this?” Now, it feels more like “how loudly can we signal that we’re mad at it?” Between deepfakes, parasocial relationships, and data center build-outs, AI has become a convenient vessel for a bunch of different anxieties: misinformation, pornography, surveillance, psychosis, water usage, energy grids, and “big tech” power in general.

At the same time, there’s now at least one explicit pro-AI political machine in the form of Leading the Future (the $100M pro-AI super PAC), without an obvious, equally organized counterweight on the anti-AI side. That feels unstable. My expectation isn’t some coherent, thoughtful “AI doctrine,” but a messy wave of backlash politics: local fights over data centers, reactive content rules, performative hearings, and eventually a serious anti-AI PAC or coalition trying to turn that sentiment into power.

IPO musical chairs

Finally, there’s the capital markets question. The rumor mill has both Anthropic and OpenAI eyeing IPOs within the next two years, and the broader AI ecosystem is already minting massive public companies in adjacent infrastructure (CoreWeave, etc.).

If the successful IPO window stays open, the rational move is to wait: grow into your valuation, deepen your moat, keep raising private rounds from hyperscalers and sovereign wealth funds, and file when you’re ready. But if the window starts to wobble - macro shocks, rate shifts, an AI backlash that spooks public investors - we could see something closer to the SPAC era: a rush to get out while the getting’s good, followed by a long hangover of underwater AI IPOs.

I don’t know which way that breaks. What I do believe is that the timing and sequencing of a handful of big AI IPOs will do a lot to set the narrative for the “AI decade”: are these disciplined, cash-generating infrastructure businesses, or expensive science projects propped up by irrational exuberance?

Your turn

As always, I’m sure I’ve missed the mark on at least one of these. But that’s the fun of writing about this space - things change so fast that I don’t have to wait long to see whether I’m right or not. And even if I’m completely wrong, I’m looking forward to seeing how things play out.

What are you thinking about for 2026? Share your trends, predictions, and forecasts in the comments!

Leave a comment

GPT-5.2

OpenAI has launched GPT-5.2, its most capable AI model yet, in the wake of the company’s internal “code red.” The new model series (comprising Instant, Thinking, and Pro variants) delivers substantial performance gains across professional tasks, particularly in knowledge work benchmarks.

Why it matters:

• GPT-5.2 targets professional workflows and AI agents specifically, with claims of better performance in spreadsheets, presentations, code, and “complex, multi-step projects” - plus reduced hallucinations in its Thinking model to build trust for enterprise use.

• While the capabilities are impressive, this release also signals a future where the path forward is solving real-world professional tasks rather than chasing abstract intelligence metrics - Anthropic has previously done the same, declaring that it would focus on real-world coding capability rather than benchmarks.

• Ultimately, models like this take months to train and launch, but CEO Fidji Simo’s acknowledgment that “code red” resources were “helpful” is an acknowledgment that the landscape has indeed changed - OpenAI is no longer solely setting the pace for model progress.

Subscribe now

Elsewhere in frontier models:

• Runway launches GWM-1, its first world model that uses frame-by-frame prediction to simulate physics, and updates Gen 4.5 to add native audio.

• Mistral releases Devstral 2, a 123B-parameter AI coding model requiring at least four H100 GPUs, alongside Devstral Small, a 24B-parameter model for local use.

• Chinese AI startup Z.ai debuts GLM-4.6V, open-weight vision models with native function calling available in 106B- and 9B-parameter versions.

• And Anthropic, OpenAI, Block, Google, AWS, Microsoft, and others launch the Agentic AI Foundation to build open-source agent standards under the Linux Foundation.

Federal Override

The Trump administration is aggressively reshaping AI policy with a dual strategy: domestically restricting states’ ability to regulate artificial intelligence companies while internationally allowing tech giants like Nvidia to sell advanced chips to China in exchange for a 25% revenue cut to the US government.

Between the lines:

• President Trump signed an executive order threatening to withhold $42 billion in broadband funding from states with AI regulations his administration deems too restrictive, as well as establishing an “AI Litigation Task Force” under Attorney General Pam Bondi to actively challenge state AI laws.

• The order targets states like Colorado (anti-discrimination provisions) and California (catastrophic risk mitigation requirements), framing consumer protections as “ideological bias” that could stifle innovation and America’s competitiveness against China.

• President Trump also announced that Nvidia can sell its H200 AI chips to “approved customers” in China, but only if the US government receives a 25% cut of the revenue. The H200s are something of a middle ground - they’re nearly six times more powerful than chips currently allowed in China but significantly behind the Blackwell chips US companies are using.

• The gamble is that the US can maintain its AI lead through a deregulated domestic market and controlled chip exports that generate revenue, but critics see a “lawless Wild West” at home and a “colossal national security failure” abroad that could ultimately strengthen China’s military and indigenous chip industry.

Elsewhere in AI geopolitics:

• New US OMB guidance states that LLMs procured by federal agencies must comply with two “unbiased AI principles“: “truth-seeking” and “ideological neutrality”.

• The NY governor is reportedly considering a rewrite of the RAISE Act, the AI bill that recently passed the NY legislature, with text copied verbatim from California’s SB 53, after signing bills requiring ads to disclose AI-generated performers and mandating consent of heirs to use a deceased person’s likeness commercially.

• US state AGs sent a letter to Meta, Microsoft, Google, Apple, and nine others, warning that their chatbots’ “delusional outputs” could be violating state laws.

• The US DOJ detains two men for allegedly violating export controls by trying to smuggle $160M+ of Nvidia chips to China; a third man pleads guilty.

• And the US DOD says it has chosen Google’s Gemini for Government to power its new GenAI.mil platform for the US military, as part of a $200M contract from July.

Elsewhere in AI infrastructure:

• Broadcom CEO Hock Tan reveals that Anthropic placed a $10B order for Google’s Ironwood TPU racks in Q3 and an additional $11B order in Q4.

• Blue Origin has worked for over a year on tech for orbital AI data centers, while SpaceX plans to use upgraded Starlink satellites for AI computing payloads.

• Nvidia says it hasn’t seen substantiation of chip smuggling via data centers outside of China after The Information’s DeepSeek story.

• Big Tech’s AI data center plans face a power crunch that could deflate the AI “bubble” as they need an estimated 44GW of additional capacity by 2028.

• And Amazon and Microsoft plan to invest $35B+ in India and $17.5B in India respectively, over the next four years on AI and cloud computing.

Creative Licensing

Disney and OpenAI have struck a $1 billion deal that will allow Sora users to create and share short videos featuring over 200 Disney, Marvel, Pixar, and Star Wars characters - marking the first major licensing partnership between a legacy entertainment giant and a generative AI company.

Between the lines:

• The three-year agreement positions Disney as both a content licensor and major OpenAI customer, with plans to integrate ChatGPT across its operations and use OpenAI’s APIs to build new Disney+ experiences, while curated fan-made Sora videos will stream on the platform starting in early 2026.

• The deal’s carefully defined boundaries reveal Hollywood’s red lines: no talent likenesses, no character voices, and OpenAI must allow Disney to “set and evolve guardrails over time.” Disney CEO Bob Iger is framing this as inevitable adaptation rather than capitulation, arguing that “no human generation has ever stood in the way of technological advance.”

• Ultimately, this is as much about strategic positioning in the AI race as it is about licensing revenue - Disney also this week sent a cease-and-desist letter to Google, accusing the tech giant of “massive scale” copyright infringement by training AI models on Disney’s characters - following similar actions against Meta, Character.AI, Midjourney, and Minimax.

Elsewhere in OpenAI:

• Sam Altman reflects on OpenAI’s 10th anniversary, highlighting a decade of breakthroughs and the path toward AGI that benefits all of humanity.

• Fidji Simo says ChatGPT’s “adult mode” is expected to launch in Q1 2026 once the company improves its age prediction capabilities.

• OpenAI and Instacart launched a grocery shopping experience inside ChatGPT that lets customers brainstorm meal ideas and check out directly.

• The company warns that its upcoming frontier AI models are likely to pose a “high” cybersecurity risk as their cyber capabilities accelerate.

• Slack CEO Denise Dresser is leaving the company to join OpenAI as its Chief Revenue Officer after serving in the role since November 2023.

• And Sam Altman said OpenAI plans to end its “code red” after releasing a model in January 2026 with improved image generation, speed, and personality.

Elsewhere in AI anxiety:

• Stanford researchers developed AI hacking bot Artemis, which surpassed nine out of 10 penetration testers by rapidly finding bugs in the university’s network.

• A study found 15 TikTok accounts posting AI videos of sexualized underage girls have nearly 300K followers, with TikTok saying that 14 of the accounts don’t violate its rules.

• The International Committee of the Red Cross warned that AI models are fabricating research papers, journals, and archives.

• The New York Times sued Perplexity, claiming the AI startup violated its copyrights and kept using its content despite repeated demands over the past 18 months.

• A developer who accidentally found CSAM in AI training data was banned by Google for reporting it.

• And McDonald’s Netherlands pulled an AI-generated Christmas ad after social media backlash.

Google has a lot going on

The headline this week was GPT-5.2, but Google is keeping up with its nonstop launch cadence.

What’s the latest:

• Disco, a Gemini 3-powered experiment that suggests and creates GenTabs, web apps based on the user’s open browser tabs and Gemini chat history.

• An enhanced Gemini Deep Research agent accessible to developers via its new Interactions API, along with a new DeepSearchQA benchmark.

• Fully managed, remote MCP servers to help developers connect AI agents to services such as Maps, BigQuery, Compute Engine, and Kubernetes Engine.

• Gemini 3 Pro sets new vision AI benchmark records, including in complex visual reasoning, beating Claude Opus 4.5 and GPT-5.1 in some categories.

• Amin Vahdat, who leads the company’s AI and infrastructure team, to chief technologist for AI infrastructure, reporting to Sundar Pichai.

• A cheaper AI Plus plan in India, costing ~$2.21 per month for the first six months and ~$4.44 thereafter, to compete with ChatGPT Go.

• Google has told advertisers that it plans to bring ads to Gemini in 2026; Google’s VP of Global Ads says there are no plans for ads in the Gemini app.

• DeepMind plans to open its first “automated science laboratory” in the UK in 2026, focused on using AI tools to develop new materials for chips and more.

• The company details the steps it is taking to secure Chrome’s upcoming agentic browsing features, like a “User Alignment Critic” model that vets each AI agent action.

• And Google details its plans for two different categories of AI-powered smart glasses coming as soon as 2026: one with in-lens displays and one that’s audio-focused.

Elsewhere in the FAANG free-for-all:

• Meta’s new AI model, codenamed Avocado, may launch in spring 2026 as a “closed” model trained using Google’s Gemma, OpenAI’s gpt-oss, and Qwen.

• Tensions have emerged between Meta’s TBD Lab and longtime lieutenants to Mark Zuckerberg over computing resources, company goals, and priorities.

• Microsoft Research, Providence, and UW developed the GigaTIME AI model, which can analyze tumors in a fraction of the time and cost of existing methods.

• And Meta acquires Limitless, which makes a pendant-style AI wearable that records and transcribes real-world conversations.

Things happen

xAI partners with El Salvador to deploy Grok across 5,000+ public schools. Time names “the Architects of AI“ as its 2025 Person of the Year. Cursor launches Visual Editor, a vibe-coding product for designers. How the Model Context Protocol became an industry standard. An overview of AI in 2025, including arguments for and against above-trend capabilities growth. AI browsers are forcing web developers to rethink whether they’re designing for humans or crawlers. The US FDA qualifies the first AI drug development tool to help assess a fatty liver disease. How Chinese companies hire Kenyan AI annotators via opaque middleman networks to avoid accountability. Inside the creation of AI actress Tilly Norwood, which took 2,000 iterations. Several companies are “refounding” to add AI features. Reasoning models now represent over half of all usage, an analysis of 100T+ tokens shows. AGI may never emerge because the concept ignores physical realities of computation. ChatGPT was 2025’s most downloaded free app in the US App Store.

Last week’s roundup

OpenAI’s Code Red

OpenAI CEO Sam Altman has declared a “code red,” ordering staff to refocus on improving ChatGPT as the company’s once-dominant position faces its most serious challenges yet.

Between the lines:

• OpenAI is reportedly shelving ambitious initiatives like ads, shopping agents, and a personal assistant called Pulse to focus on core ChatGPT improvements - with daily calls and temporary team transfers to accelerate development.

• It’s a striking role reversal from three years ago - when Google declared its own “code red” in response to ChatGPT’s explosive debut. Google has orchestrated a remarkable turnaround, with Gemini 3 outperforming GPT-5 on key benchmarks, its mobile app surging to 650 million monthly users, and its custom TPU chips allowing it to train models without relying on expensive Nvidia hardware.

• Ultimately, OpenAI’s aggressive expansion into multiple product lines simultaneously - from video generation (Sora) to coding tools to planned advertising and shopping features - may have stretched the company too thin.

Subscribe now

Elsewhere in frontier models:

• Mistral launches Mistral 3, a family of 10 models under the Apache 2.0 license, including its new flagship Mistral Large 3 and nine smaller Ministral 3 models.

• Chinese short-video company Kuaishou launches Kling Video O1, saying it is the first multimodal AI model to unify video generation, editing, and post-production.

• DeepSeek releases DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, which it calls “reasoning-first models built for agents”, after releasing V3.2-Exp in September.

• Runway debuts the Gen-4.5 AI video generation model, highlighting its physical accuracy and holding the top spot on Artificial Analysis’ Video Arena leaderboard.

• Nvidia announces Alpamayo-R1, an AI model for autonomous driving research, calling it the “first industry-scale open reasoning vision language action model”.

• Prime Intellect debuts INTELLECT-3, a 106B open source MOE model it claims outperforms larger models.

• Google rolls out Gemini 3 Deep Think to Google AI Ultra subscribers in the Gemini app after delays for safety evaluations.

• And Apple releases Starflow-V, an open weights video model.

Elsewhere in OpenAI:

• OpenAI is testing training LLMs to produce “confessions”, or self-report how they carried out a task and own up to bad behavior like lying or cheating.

• OpenAI’s nonprofit foundation plans to award $40.5M in grants in 2025 to 208 US nonprofits, up from only $7.5M donated in 2024.

• OpenAI takes a stake in Thrive Capital’s Thrive Holdings and says it will embed agents in its companies, which already include an accounting and an IT business.

• OpenAI’s data center partners are set to rack up nearly $100B in debt, with banks potentially lending another $38B to Oracle and Vantage to build more OpenAI sites.

• And a US federal judge rules that OpenAI must produce 20M anonymized ChatGPT chat logs in the copyright lawsuit brought by The New York Times and other news outlets.

Anthropic’s IPO

The AI IPO race is intensifying as Anthropic (and OpenAI) appear to be preparing for public listings at unprecedented valuations, despite incurring massive losses and facing unpredictable financial forecasts.

What to watch:

• This week, Anthropic made concrete moves toward going public - hiring Wilson Sonsini (the firm behind Google, LinkedIn, and Lyft IPOs) and bringing on Krishna Rao as CFO, who led Airbnb’s successful public debut.

• The company is valued at $183 billion, is on track to hit $10 billion in annualized revenue - more than 10x its 2024 figure - and just announced its first-ever acquisition of Bun, a high-performance JavaScript runtime, as its Claude Code product hits $1 billion in run-rate revenue.

• Anthropic’s success challenges the narrative that the AI race requires unprecedented capital expenditures: the company generates 2.1 times more revenue per dollar of computing costs compared with OpenAI.

• But both IPOs will be a crucial test of whether public markets will embrace the expensive, unprofitable AI research labs that have thrived in private markets, potentially setting the tone for the entire AI industry’s financial future and sustainability.

Elsewhere in Anthropic:

• Snowflake and Anthropic announced a multiyear $200M deal to make Claude models available on Snowflake’s platform and deploy AI agents for enterprises.

• Anthropic says 60%+ of its business customers use more than one Claude product, a trend it began noticing in summer 2025 after Claude Code’s rise in popularity.

• Anthropic’s employees self-report using Claude in 60% of their work, achieving a 50% productivity boost, mostly for debugging and code understanding.

• Anthropic’s societal impacts team studies AI’s broad societal risks to tackle “inconvenient truths”, beyond typical safety teams at AI startups.

• A look at the exfiltrated 14K-token document that Claude calls its “Soul overview”, confirmed to be valid by Anthropic employees.

• And using the SCONE-bench benchmark of 405 blockchain smart contracts, Claude Opus 4.5, Claude Sonnet 4.5, and GPT-5 developed exploits together worth $4.6M.

Elsewhere in AI infrastructure:

• Jensen Huang says Nvidia doesn’t know if China would accept its H200 AI chips should the US relax export controls on them, after a meeting with President Trump.

• An in-depth look at how the latest Google TPU generation positions Google as the most threatening challenger to Nvidia’s AI chip dominance.

• The AI boom has led to high demand and more pay for the construction workers who build data centers, with a trade group estimating there’s a shortage of ~439K workers.

• And Dario Amodei suggests some AI companies are taking on too much risk by “YOLO-ing” and committing to spend hundreds of billions on data centers.

Three-Year Mark

Three years after ChatGPT’s release sparked a global AI frenzy, Americans remain deeply divided over the technology - with the US showing more fear and less excitement about AI than nearly any other surveyed nation, even as it remakes the world around us.

The big picture:

• For starters, artificial intelligence is becoming one of the most popular college majors in the U.S., with MIT’s AI program now the second-largest major on campus and thousands of students enrolling in new AI degrees across the country.

• Likewise, AI mania has fundamentally reshaped Wall Street, driving a 64% surge in the S&P 500 while creating unprecedented market concentration around a handful of tech giants. However, this success comes with a caveat: the seven largest companies now account for 35% of the S&P 500, creating significant concentration risk.

• And while legitimate concerns exist (electricity usage, carbon emissions), the technology’s genuine utility - as a knowledge assistant, translator, and productivity tool - tends to go unappreciated.

Elsewhere in AI geopolitics:

• President Trump’s plans to block state-level AI regulations via a preemption proposal face opposition from Democrats, Republicans, and consumer groups.

• The EU expects to launch a formal bidding process for its AI gigafactories in early 2026 and close it in summer 2026, as the bloc seeks to catch up with the US.

• Jimmy Wales says Wikimedia is working with Big Tech on AI deals similar to Google’s and reader donations are “not to subsidize OpenAI costing us a ton of money”.

• The European Commission opens a new antitrust investigation into Meta over its rollout of its AI assistant in WhatsApp; the probe will not fall under the DMA.

• David Sacks’ AI and crypto policies in Trump’s White House benefit his investments, those of his Silicon Valley friends, and the All-In podcast he co-hosts.

• And scientists are (once again) becoming increasingly worried AI will sway elections.

Elsewhere in AI anxiety:

• Some people are feeding years of medical records into chatbots like ChatGPT, despite privacy risks, and receiving generalized or inaccurate diagnoses in response.

• YouTube creators are using AI tools to make videos for kids and babies, raising concerns that such AI content may negatively impact early brain development.

• AI Forensics found that 354 AI-focused TikTok accounts pushed 43K posts made with GenAI tools that hit 4.5B views, including posts with anti-immigrant and sexual material.

• Researchers unveiled PropensityBench, a benchmark showing how stressors like shortened deadlines increase misbehavior in agentic AI models during task completion.

• Pangram Labs found that ~21% of the 75,800 peer reviews submitted for ICLR 2026, a major ML conference, were fully AI-generated, and 50%+ contained signs of AI use.

• And Unit 42 details how underground hacking forums advertise and sell custom, jailbroken, and open-source AI hacking tools such as WormGPT and KawaiiGPT.

re:Invent 2025

Amazon held its AWS re:Invent 2025 conference this week, and we got a fairly impressive set of new launches, including new models, chips, and agents.

What to watch:

• Amazon Nova, a second-generation AI model family including Nova Lite, Nova Pro, Nova Sonic, and fully multimodal reasoning model Nova Omni, released to limited customers.

• Trainium3, an AI chip that is 4x faster than Trainium2 and can cut AI training and operating costs by up to 50% compared to equivalent GPUs.

• Trainium3 UltraServer, a system powered by its 3nm Trainium3 AI training chip, alongside a tease of Trainium4, which will work with Nvidia’s chips.

• Nova Forge, a $100,000/year service allowing clients to customize Amazon’s AI models at various stages of training and refine open-weight models.

• Three frontier agents: Kiro autonomous agent, AWS Security Agent, and AWS DevOps Agent, each focused on different aspects of software development.

• Bedrock AgentCore expansion with new tools for managing agent boundaries, agent memory capabilities, and agent evaluation features.

• DevOps Agent, an AI-enabled tool designed to help clients quickly identify root causes of outages and implement fixes, available in preview.

• AWS AI Factories, which lets customers deploy AWS infrastructure, including Trainium chips and Nvidia GPUs, in their existing data centers.

• And an interview with AWS CEO Matt Garman on his AI vision, extending Amazon’s cloud market lead, adding AI to AWS services, and AI efficiencies.

Things happen

ByteDance launches an AI voice assistant powered by its Doubao LLM. ChatGPT told a violent stalker to embrace the “haters”. Thoughts on AI progress and model limitations. A profile of Cursor CEO Michael Truell, 25. Google threatens both OpenAI and Nvidia. Stanford tool downranks antagonistic posts on X. US startups adopt open-weight Chinese AI models. Apple’s AI chief is retiring. Everyone in Seattle hates AI. Reverse engineering a legal AI tool exposed 100K+ confidential files. A new AI winter is coming? We gave 5 LLMs $100K to trade stocks for 8 months. Anti-patterns while working with LLMs. Codex, Opus, Gemini try to build Counter Strike. Writing a good Claude.md. Search tool for content created before ChatGPT. CA ballot measure to undo nonprofit conversions, targeting OpenAI.

Last week’s roundup

Opus Day

Anthropic released Claude Opus 4.5 on Monday, reclaiming the crown for best coding model with an 80.9% score on SWE-Bench Verified. The release caps a frenetic week of AI lab competition, coming just days after OpenAI’s GPT-5.1-Codex-Max and Google’s Gemini 3 Pro.

What to watch:

• The new model is dramatically cheaper than the previous Opus release ($5/$25 per million tokens, as opposed to the previous $15/$75), meaning it may actually be cost-effective to use Opus widely as a base model - especially considering it matches its predecessor Sonnet 4.5’s performance while using 76% fewer tokens.

• With this release, Anthropic finally restores its three-tier model hierarchy after Sonnet 4.5 had been outperforming the older Opus 4.1, giving developers clear choices across price-performance curves rather than forcing everyone to upgrade to the flagship model.

• For developers, Anthropic is also releasing three new features to help Claude work with massive tool libraries: Tool Search Tool (for discovering tools on demand), Programmatic Tool Calling (for executing tools through code), and Tool Use Examples (for demonstrating correct usage patterns).

• But all these rapid-fire releases expose a growing evaluation problem: Even Simon Willison struggled to distinguish Opus 4.5’s real-world capabilities from its predecessor Sonnet 4.5, noting he “kept on working at the same pace“ after switching back - suggesting frontier models may be converging faster than our ability to meaningfully differentiate them.

Subscribe now

Elsewhere in frontier models:

• Microsoft unveils Fara-7B, its first agentic small language model designed for computer use, available as an experimental release on Hugging Face and Microsoft Foundry.

• An analysis of Gemini 3’s model card and safety framework report finds the model is excellent but the safety report withholds or makes it difficult to understand key information.

• Ilya Sutskever discusses model jaggedness, why we are moving beyond the “age of scaling,” and SSI’s plan to straight-shot superintelligence in a new Q&A.

• OpenAI merges ChatGPT’s voice mode directly into the main text chat interface by default, though users can still switch back to the original separate voice mode.

• And OpenAI unveils a free shopping research feature in ChatGPT that delivers a personalized buyer’s guide, powered by a custom version of GPT-5 mini.

Short-term memory loss

A severe memory chip shortage, driven by surging demand for AI infrastructure, is threatening to disrupt supply chains across the tech industry. Major manufacturers, including Dell and HP, are warning of unprecedented cost increases that will inevitably be passed on to consumers.

The big picture:

• The crunch stems from manufacturers prioritizing production of advanced, high-bandwidth memory for profitable AI systems over conventional memory chips used in everyday devices - from smartphones and PCs to medical equipment and cars.

• Some companies, like Apple, are leveraging their scale and long-term contracts to secure supply and potentially gain market share. Others, like Lenovo, have begun stockpiling memory inventories up to 50% higher than normal levels, as memory chip prices are forecast to rise 50% through mid-2026.

• The shortage is becoming a bottleneck for the broader tech ecosystem: Logic chip suppliers may see reduced orders if customers can’t secure accompanying memory, while U.S. sanctions limiting Chinese chipmakers’ capabilities are exacerbating supply constraints just as panic buying intensifies the crisis.

• The timing creates a frustrating moment for consumers - GPU prices are finally normalizing after years of volatility, making graphics cards affordable just as the cost of building or upgrading a PC has become prohibitively expensive due to memory shortages.

Elsewhere in AI infrastructure:

• Google’s head of AI infrastructure says the company must double AI compute capacity every six months to meet demand.

• Meta enters electricity trading to help accelerate the construction of new US power plants critical for its AI infrastructure buildout.

• AWS announces a commitment to invest up to $50B to build AI and HPC infrastructure for the US government, starting in 2026 and adding nearly 1.3 GW of capacity.

• Foxconn Chair Young Liu says the company will spend $2B to $3B per year in AI in the next three to five years and is discussing potential investments with Japan.

• And since 2023, data center power demands have delayed 15 coal plants’ retirements as the Trump administration has ordered two power plants to remain open.

Chip shot

Google is intensifying its challenge to Nvidia’s AI chip dominance by pitching major customers on using its Tensor Processing Units (TPUs) in their own data centers, with Meta reportedly considering a multi-billion-dollar deal for 2027.

Between the lines:

• The move prompted Nvidia’s stock to fall 3% and drew a swift public response from the chip giant, which declared its technology remains “a generation ahead” of alternatives like Google’s TPUs.

• It’s also throwing its war chest around - Nvidia CEO Jensen Huang has been proactively countering Google’s moves by investing billions in Anthropic and OpenAI after they explored TPU deals, essentially using capital to maintain GPU lock-in even as Google offers cheaper alternatives.

• Google’s recent technical progress with Gemini 3 and seven generations of TPU development may finally be paying off - Meta is reportedly interested in using TPUs for model training, not just inference, which many had assumed was Nvidia’s unassailable stronghold.

Elsewhere in the FAANG free-for-all:

• Google is starting to bridge OpenAI’s product moat with Gemini’s “dynamic view” option, which converts text answers into interactive, visual outputs.

• A job listing shows Google is developing a new Android-based “Aluminium OS” that is “built with AI at the core”, potentially as a ChromeOS replacement for PCs.

• Meta plans to test an AI-powered personalized daily briefing, designed to compete with ChatGPT’s Pulse, with some Facebook users in NYC and SF.

• Google says ads that some users are seeing in AI Mode are part of a test; the ads have a “sponsored” label and appear at the bottom of the page.

• And Amazon unveils its Autonomous Threat Analysis system, born from a 2024 internal hackathon, to use AI agents competing in teams to identify vulnerabilities.

Open season

China has surpassed the US in downloads of “open” AI models on Hugging Face for the first time, capturing 17% market share versus America’s 15.8%, as Beijing encourages its tech companies to release freely accessible models while US giants like OpenAI and Google maintain closed, proprietary systems.

The big picture:

• Led by DeepSeek and Alibaba’s Qwen models, Chinese AI developers are releasing models weekly at dramatically lower costs - while American tech leaders like OpenAI, Google, and Anthropic maintain tight control over their “closed” proprietary technology to protect lucrative subscription businesses.

• Likewise, foreign investors are warming to China’s AI sector precisely because it lacks the bubble characteristics plaguing US tech, with venture capital firms raising hundreds of millions in recent months to back Chinese startups at valuations one-quarter of American peers.

• But China’s open-source dominance comes with ideological strings attached: researchers have documented that Chinese models embed Communist Party biases and refuse to generate information on sensitive topics like Taiwan or Tiananmen Square, meaning widespread adoption could shape global information flows.

Elsewhere in AI geopolitics:

• Italy’s competition regulator is scrutinizing a Meta policy that excludes rival AI chatbots from WhatsApp, broadening the scope of a probe started in July.

• The US House Homeland Security Committee asks Dario Amodei to testify at a December 17 hearing on how Chinese state actors used Claude Code for cyber-espionage.

• A new network of super PACs plans to raise ~$50M to counter the Leading the Future super PAC and back candidates who prioritize AI regulations.

• Trump signs an EO establishing the Genesis Mission to boost AI innovation, including by using federal scientific datasets to train models and create AI agents.

• New York’s RAISE Act would require AI companies to publish safety protocols and disclose serious incidents, as its co-sponsor is targeted by a pro-AI super PAC.

• And the UK announces a £100M plan to support local AI hardware startups via guaranteed “first customer” payments, estimating its AI market to be worth over £72B.

Elsewhere in AI anxiety:

• An MIT study finds that AI can replace 11.7% of the US labor market, or ~$1.2T in wages, based on the “Iceberg Index” which measures job automation potential.

• Anthropic finds that LLMs trained to “reward hack” by cheating on coding tasks show even more misaligned behavior, including sabotaging AI-safety research.

• Major insurers have recently sought permission from US regulators to offer policies excluding liabilities tied to businesses deploying AI chatbots and agents.

• While the AI bubble feels like the internet bubble of 1999, it may actually be larger and scarier with an unstable US economy and greater exposure for Big Tech.

• Researchers say Russia-aligned Pravda network is engaging in “LLM grooming”, flooding the internet with disinformation to influence chatbots like ChatGPT.

• And interviews with current and former OpenAI employees detail how updates that made ChatGPT more appealing to boost growth sent some users into delusional spirals.

Things happen

Epic games CEO says it makes no sense to label AI-made games since AI will be in everything. Report: Suno spent $32M on compute and $2K on data. 73% of AI startups are just prompt engineering. PopEVE predicts disease-causing mutations better than AlphaMissense. Indian startups are building LLMs for low-resource languages from scratch.Amazon begins previewing Leo satellite internet ahead of 2026 launch. Sam Altman and Jony Ive say their OpenAI device could arrive in less than two years. The bitter lesson of LLM extensions. The USPTO clarifies that AI cannot be an inventor. Character.AI is cutting off ongoing chats for users under 18. Figure AI’s former head of product safety sues the company for wrongful termination. An AI agent bypasses bot detection on online surveys nearly flawlessly. Amazon pushes engineers to use Kiro over Cursor. The humanoid robotics industry relies heavily on hype. Food bloggers warn AI Overviews are burying real recipes. A Mixpanel breach exposed OpenAI API data; OpenAI terminated its use. I don’t care how well your AI works. Replace your boss before they replace you. Karpathy on the implications of AI to schools. PRC elites voice AI-skepticism. Altman: Google’s AI progress could create economic headwinds for OpenAI. The AI workers tell their friends and family to stay away from AI. We’re losing our voice to LLMs.

Last week’s roundup

Gemini 3

Google launched Gemini 3 Pro on Tuesday, its most capable AI model to date. The model, which has generated considerable coverage (including here), demonstrates strong planning, coding, and judgment skills, showing that AI models have progressed beyond hallucinations to subtle, human-like errors.

What to watch:

• Google also released “Nano Banana Pro,” its viral-named Gemini 3-powered image model that excels at following instructions and creates full infographics from short prompts.

• Google debuted Antigravity, an “agent-first” coding tool that leverages Gemini 3 Pro and third-party models, in public preview for Windows, macOS, and Linux.

• Google priced Gemini 3 Pro at $2 to $4 per 1M input tokens and $12 to $18 per 1M output tokens, cheaper than Claude Sonnet 4.5 but more expensive than GPT-5.1.

• Google details a new generative UI implementation that lets Gemini 3 Pro create interactive interfaces for any prompt, launching in the Gemini app and AI Mode.

• And Google plans to release Gemini 3 Deep Think to Google AI Ultra subscribers in the coming weeks, once it passes further rounds of safety testing.

Elsewhere in frontier models:

• OpenAI unveils GPT-5.1-Codex-Max, saying it is “significantly better” at “long-horizon reasoning” and is the first model it has trained for Windows environments.

• xAI’s Grok 4.1 tops benchmarks in emotional intelligence, while its model card also shows a marked increase in sycophancy compared to Grok 4.

• Allen Institute for AI unveils Olmo 3 models that it says outperform open models like Stanford’s Marin and commercial open-weight models like Llama 3.1.

• Meta releases Segment Anything Model 3 for advanced image segmentation tasks.

• And Google DeepMind releases WeatherNext 2, a weather model that it says offers faster, more accurate two-week forecasts and includes more tools for energy traders.

Subscribe now

Elsewhere in the FAANG free-for-all:

• Google says the Gemini app is now able to detect images created or edited by Google AI, and that it plans to roll out verification of video and audio “soon”.

• Microsoft launches Agent 365, a framework that lets businesses deploy and manage AI agents like human employees, with dashboards showing telemetry and alerts.

• Microsoft integrates more AI features into Windows, including new Copilot skills, AI agents on the taskbar, writing assistance, and troubleshooting agents.

• Amazon launches a Video Recaps feature on Prime Video that uses AI to summarize a show’s key plot points with snippets and music, available in beta for US users.

• Yann LeCun says he is leaving Meta at the end of 2025 to build a new startup and continue his “Advanced Machine Intelligence research”, with Meta as a partner.

• And Meta Chief Revenue Officer John Hegeman is leaving after 17 years at Meta; Clara Shih, who joined in November 2024 to lead its Business AI unit, is also exiting.

Irrational Exuberance

Tech leaders are now openly questioning the trillion-dollar AI investment boom, with Google CEO Sundar Pichai warning of “elements of irrationality” in the current spending frenzy, the CEO of Klarna being “nervous about the size of [data center] investments,” and Peter Thiel’s hedge fund selling its entire $100 million Nvidia stake.

Between the lines:

• AI company valuations continue to soar - Nvidia’s $7T, Alphabet’s $3.5T, OpenAI’s $500B, Anthropic’s $350B, and xAI’s $230B - raising questions about whether AI-powered products will generate enough revenue to justify the massive capital outlays.

• Pichai’s candid admission that “no company is going to be immune” to a potential AI bubble burst - including Google - marks a notable shift from the unchecked optimism that has characterized tech leadership rhetoric.

• JPMorgan analysts estimate that $5 trillion in global AI infrastructure investment through 2030 would require generating an additional $650 billion annually - equivalent to every iPhone owner paying $35 more per month - to deliver reasonable returns to investors.

• Despite the warnings, tech giants are doubling down rather than pulling back - but supply chain constraints for critical components like transformers (booked through 2028) and power generation equipment are creating hard limits on how fast data centers can actually be built.

Elsewhere in AI infrastructure:

• The US DOE accelerates its approach to equipping national labs with AI supercomputers by working with Nvidia, AMD, and Oracle, which will pay some of the costs.

• AI server shipments from Taiwan to the US are set to double in 2025 compared to 2024, as Foxconn, Quanta, and Wistron ramp up production to meet demand.

• The US Commerce Department approves the sale of up to 35K Nvidia GB300 servers or their equivalents each to Abu Dhabi’s G42 and Saudi government-backed Humain.

• Elon Musk says xAI plans to develop a 500-megawatt data center in Saudi Arabia with state-backed AI startup Humain; the data center will rely on Nvidia chips.

• And Elon Musk, Jeff Bezos, and Sundar Pichai have each discussed building lunar and orbital AI data centers to leverage constant solar power and bypass regulations.

Elsewhere in OpenAI:

• OpenAI expands group chats in ChatGPT globally to all logged-in users on Free, Go, Plus, and Pro plans after piloting the feature in select regions.

• OpenAI announces ChatGPT for Teachers, designed for K-12 educators and school districts, which will be free to K-12 educators in the US through June 2027.

• OpenAI and Intuit sign a multiyear deal set to generate $100M+ in OpenAI revenue, in which Intuit will deepen its use of OpenAI models and launch ChatGPT apps.

• OpenAI says ChatGPT will now avoid em dashes if users tell it to, addressing the telltale sign that has signaled AI-written text.

• And OpenAI says GPT-5 has demonstrated the ability to accelerate scientific research workflows but can’t run projects or solve scientific problems autonomously.

States Rights

President Trump is reportedly preparing an Executive Order as soon as Friday that would give the federal government sweeping power to override state AI laws, including creating an “AI Litigation Task Force” under the attorney general to go after “woke” state AI regulation.

Why it matters:

• This would represent a dramatic assertion of federal power over AI regulation, with the Justice Department empowered to challenge state laws like California’s AI safety requirements and Colorado’s algorithmic discrimination protections.

• The “AI Litigation Task Force” would weaponize critical infrastructure funding - states could lose Broadband Equity Access and Deployment (BEAD) program money for rural broadband if they maintain AI regulations deemed “onerous.”

• But the EO wouldn’t pass without a fight - not only will there be lawsuits, but also intra-party dissent. High-profile Republicans - including Republican governors, senators, and MAGA supporters - are already openly opposing the President.

• Ultimately, a pro-AI stance is increasingly a risky political calculation for both parties: concerns about AI-related child suicides, job losses, and energy costs could become devastating attack lines in 2028.

Elsewhere in AI geopolitics:

• The European Commission’s postponement of a key part of the AI Act has become a case study for critics who say the EU prioritizes regulation over innovation.

• Key White House officials are pressing lawmakers to keep AI chip export restrictions to China out of the annual defense policy bill.

• Pro-AI super PAC Leading the Future is targeting New York State Assembly member Alex Bores, who co-sponsored the RAISE Act requiring AI safety protocols.

• An investigation reveals that as US Commerce Secretary Howard Lutnick pushes AI data centers and pressures US allies, his companies and sons are profiting from them.

• And the current AI strategies of China and the US are complementary, as China focuses on embodied AI and open source models rather than AGI like the US.

Elsewhere in AI anxiety:

• A study found that teen mental health chatbots like ChatGPT, Claude, Gemini, and Meta AI often failed to recognize signs of mental health conditions and gave general advice.

• AI is pressuring Poland’s economy, Europe’s top location for global business services, as companies lay off staff in Kraków’s BPO hub.

• Some experts question Anthropic’s claims of cyberattack breakthroughs using its tools, noting that white-hat hackers report modest gains from AI-aided hacking.

• Chinese toymaker FoloToy suspended sales of its GPT-4o-powered teddy bear after researchers found the toy gave kids harmful responses about fire and kinks.

• And Microsoft warns that Copilot Actions in Windows, now in beta and off by default, can infect devices and pilfer data, prompting concern from security researchers.

Things happen

Warner Music settles its lawsuit with AI music startup Udio. Soumith Chintala, PyTorch co-creator, joins Mira Murati’s lab. Poe launches group chats with up to 200 people across 200+ AI models. A look at Inception Point AI’s podcast network generating 3,000 podcasts per week. Cloudflare acquires Replicate, which hosts 50K+ AI models. Larry Summers resigns from OpenAI’s board after his Epstein emails surface. Elon Musk could “drink piss better than any human in history,” Grok says. Massive leak shows chatbot users turned yearbook pictures into AI porn. A researcher made an AI that completely breaks online surveys. DeepMind plans to open an AI research lab in Singapore. An overview of macro tech trends for 2026 as AI eats the world. At an AI conference, attendees said they’d short Perplexity over OpenAI. AI is a front for consolidation of resources and power. The AI bubble is bigger than you think. Claude launches structured outputs on its developer platform. Solving a million-step LLM task with zero errors. Heretic: Automatic censorship removal for language models. Google opens its biggest AI hardware hub outside the US in Taipei. Berkshire Hathaway discloses a $4.3B stake in Alphabet. Demis Hassabis pursued a Nobel as a DeepMind objective over revenue. Meta will assess employees on their “AI-driven impact” starting in 2026. Apple requires apps to disclose AI data sharing with third-party providers.

Last week’s roundup

Gemini 3 Pro

Technically only Gemini 3 Pro is being released - today Gemini 3 is a single model, but tomorrow it’ll be a model family. The Gemini 2.5 family consisted of Pro, Flash, and Flash-lite, and I’m expecting a similar lineup for Gemini 3. Google is also hinting at new models “with new features, across different modalities,” so stay tuned.

But these days, it makes sense to lead a launch with the biggest model size if you can. It’s going to take up all of the attention, especially if it’s got good benchmarks.

Speaking of benchmarks

Of course, the usual benchmark disclaimers apply - they’re often gamed, they don’t capture everything that matters, and real-world performance is what counts. But there are a few interesting data points worth highlighting1.

Gemini 3 Pro scored an impressive 31.1% on ARC-AGI-2, a benchmark that’s intentionally designed to be very difficult for LLMs, and a nearly 2x improvement over GPT-5.1. Gemini with “Deep Think” - a compute-heavy system that spawns multiple agents in parallel - scored an even higher 45.1%. For context, this benchmark has been notoriously resistant to being solved by throwing more compute at the problem, so these numbers are genuinely notable.

Sonnet 4.5 retains its lead as the “best” coding model according to SWE-bench Verified, but GPT-5.1 and Gemini 3 Pro are now quite close. We’re talking about differences of a few percentage points, which means we’re entering the territory where model choice matters less than the tooling around it - more on this later.

Perhaps the best comparison, though, is Gemini with itself - from 2.5 Pro to 3 Pro. It’s remarkable how much better the model has improved, especially given that the 2.5 Pro is already widely considered a strong model!

Again, I don’t want to get caught up in the benchmarks here. It’s certainly possible to game them, and there’s undoubtedly enormous pressure to do so. That said, if Google’s benchmarks are any indication of Gemini 3’s overall improvements over 2.5, then they really cooked with this one.

Beyond benchmarks

But Google isn’t just trying to win on numbers. They’re also trying to showcase Gemini’s broad capabilities across various surfaces and modalities.

One aspect of this is multimodal reasoning, which incorporates images and videos. Gemini 3 appears to excel in spatial reasoning, including notoriously challenging tasks such as generating an accurate clock face. The model can also ingest and reason over long videos, such as lectures, tutorials, and sports footage. This is table stakes for frontier models now, but Gemini 3’s scores on Video-MMMU suggest it’s very strong at it. The ability to throw a 2-hour lecture at a model and have it extract the key concepts, or to analyze an entire basketball game for strategic insights, is the kind of thing that feels like magic (until it becomes mundane).

A second aspect is UI design - including generative UI. With every major release, there’s always a wave of shiny demos highlighting the best use cases, particularly around coding beautiful, custom websites and interfaces. Often, these tend to be less impactful as time goes on - they’re great for landing pages, but don’t hold up over time when you need to maintain or iterate on them. But Gemini 3 has a couple of unique things going for it.

For starters, there’s Replit Design, a new feature powered by Gemini 3 Pro. The examples I’ve seen have been quite impressive (though I admit I haven’t tried using LLMs to generate UIs in some time). But it’s not just about one-off designs - the fact that it’s connected to Replit means you can quickly deploy the front-ends without having to worry about pesky backend steps.

However, one of the most exciting aspects of the entire Gemini 3 launch (IMO) is another Google research paper about generative UI. Generative UI is the idea of having the AI dynamically create a custom presentation - layout, format, UX, and all - for content. Since GPT-4’s breakout moment, tons of developers and startups have been experimenting with this. As far as I’ve seen, none have really nailed it. Most examples I’ve seen quickly go off the rails with complexity or are hopelessly irrelevant, generating elaborate interfaces for simple queries or completely misunderstanding what presentation format makes sense.

But Google’s are... actually reasonable. There’s a repo with a couple dozen examples here, including some delightful results from stupidly simple queries like “dragonfruit” or “green things.”

Look, I’m sure that they’re certainly cherry-picked examples, and we’ll see a bunch of strange or broken dynamic UIs in the wild. But the fact that Google is bringing them to AI mode, and attempting them at scale, implies that they’re at least reasonably confident that they can do a good job here.

And, of course, there’s the fact that launches which relied on cherry-picked examples a year or two ago are now just... generally that good, every time. Image generation comes to mind here - what once required dozens of attempts and careful prompt engineering now works on the first try.

Anyways, the last aspect is agentic coding, which brings us to...

Google Antigravity

Google has (re-)entered the AI coding space, this time with an IDE - Google Antigravity. Their previous project, Jules, hasn’t garnered significant mindshare as a Claude Code alternative, so this feels like a second attempt to break into the developer tools market.

I have to assume this is the result of the Windsurf deal. When DeepMind poached Windsurf’s founders and engineers last summer, leaving the rest of the company to be acquired by Cognition, most observers assumed they would be building some sort of IDE for Google. Even still, it’s impressive what they’ve shipped in under six months.

The product

On the surface, Antigravity is a VSCode fork - very similar to Cursor. I think many developers are going to wonder: why is this different? What makes this worth switching?

I’m still reserving final judgement, but some new features are opening up interesting possibilities here.

For starters, there’s the “agent manager” dashboard. In the launch video, the Antigravity team (i.e., the former Windsurf team) is blunt: “you’ve been elevated to a manager of agents.” And Antigravity’s UI was clearly built with agents in mind. The management panel is some of the best UX I’ve seen for “pair programming” alongside a coding agent2.

For better or worse, this appears to be where the software industry is headed - though I do appreciate that Google’s IDE at least gives you the option to choose between “agent-driven development” and “agent-assistant development.” The former is the “go build this feature while I grab coffee” mode, while the latter is still giving you some semblance of control over the experience.

Then there are Artifacts. These are markdown documents that are generated as the coding agent works - it’s very reminiscent of Cursor’s “plan mode.” But Antigravity uses them in many different ways: plans, step-by-step walkthroughs, screenshots, browser recordings, and test runs. As someone who is still inherently distrustful of long-running coding tasks, I’m a big fan of the ability for users to add comments to the artifact, much like a code review. There’s something psychologically satisfying about being on the same page ahead of time, rather than waiting for it to be completed and then starting over.

One great use case for Artifacts is using them to read and write to internal knowledge bases - something the best Cursor users I know have adopted as a habit, but is far from common knowledge.

Lastly, there’s a browser extension. This is a neat trick that addresses one of the main limitations of coding agents like Claude Code: they can only ever deal with the code. If there’s something wrong in situ - once the code is actually running in a browser - there isn’t a great way for the agent to see it.

Some agents, like Devin, can circumvent this by having access to an entire virtual machine, as well as a browser. However, that’s both a significant undertaking and a potential security risk. Antigravity is leveraging its Google pedigree by having a Chrome extension that gives the coding agent hooks into the browser directly, letting it get its “hands dirty” in the frontend code itself to look for bugs. It’s the kind of integration that’s obvious in hindsight but likely requires being part of the Chrome ecosystem to pull off seamlessly.

The trajectory

Google appears to now be singing the same tune as Anthropic and OpenAI: chatbots are passé, and coding models and agents are the next big battlefront.

Anthropic has been the “leader” in this regard for some time. Even as GPT-5 has put up impressive numbers, most tuned-in engineers I know still prefer Claude as a daily driver for coding. As I wrote about Claude 4.5 Sonnet:

Anthropic has leaned heavily into the coding performance of [Claude 4.5 Sonnet], going so far as to call it “the best coding model in the world” and “the strongest model for building complex agents.” Early reviews seem to support this, with individuals like Simon Willison impressed by how well the model can code, and companies like Cursor endorsing the model from the outset.

And as far as I can tell, Anthropic is doubling down on pushing the frontier of coding capabilities. For whatever reason, they’ve cracked the code (no pun intended) on making models that developers love to use. The latest model can reportedly code for up to 30 hours at a time on long-running tasks - a remarkable number, assuming all of those hours are productive.

Of course, the model choice may not matter in and of itself if we’re ultimately relying on the agentic chassis powering things. And to that end, it feels like Antigravity isn’t aimed at competing with Cursor so much as it is competing with something like Devin - giving coding agents all the tools necessary to work end to end, and only incorporating the human when it deems us necessary.

The agent manager dashboard, the artifacts system, the browser integration - these are all features designed for a world where developers spend more time reviewing and directing than actually writing code.

Beating the allegations

It’s remarkable how far Google has come in the past two-ish years.

For much of 2023, they were widely perceived as being in disarray. The famous “we have no moat” memo leaked in May, suggesting that Google was losing the AI race to both closed-source competitors and open-source alternatives. People didn’t think they would vanish overnight, but certainly many were quick to anoint ChatGPT as a potential “Google killer.”

Now, their models are absolutely best in class, and it’s hard to argue that they’re not extremely well-positioned. Let me break down why.

Read more

Agentic espionage

Anthropic has disclosed what it believes is the first large-scale cyberattack executed almost entirely by AI agents, with a Chinese state-sponsored group using Claude Code to autonomously infiltrate roughly thirty global targets.

The big picture:

• The automation level marks a significant escalation: AI handled 80-90% of the operation with only 4-6 critical human decision points, performing reconnaissance, writing exploit code, harvesting credentials, and extracting data at thousands of requests per second.

• Weirdly, this may be a bullish signal for American AI labs. The attacks come as US export controls create acute semiconductor shortages in China, prompting companies to resort to smuggling and workarounds - implying that it may be more effective or reliable to use Anthropic’s models over Chinese ones.

• Ultimately, the incident underscores AI’s dual-use dilemma: the same tools designed to strengthen cybersecurity defenses can be weaponized to supercharge attacks, creating an arms race where defenders must maintain a “substantial permanent advantage” or risk losing ground to increasingly automated threats.

Subscribe now

Elsewhere in AI geopolitics:

• The UK proposes an amendment to the Crime and Policing Bill that would let “authorized testers“ proactively assess AI models for their ability to generate CSAM.

• A Munich court sides with Germany’s music rights society GEMA in a case against OpenAI, saying OpenAI can’t use song lyrics without a license.

• The European Commission plans to simplify some of its privacy rules, including GDPR, to boost AI growth and slash red tape for businesses in Europe.

• Anthropic open sources a method to score AI model political evenhandedness, with Gemini 2.5 Pro scoring 97%, Grok 4 at 96%, Claude Opus 4.1 at 95%, GPT-5 at 89%, and Llama 4 at 66%.

• OpenAI asks a US federal judge to reverse an order requiring it to turn over 20M anonymized ChatGPT chat logs in the NYT’s copyright lawsuit, citing privacy concerns.

• And an investigation traces how Shanghai-based AI startup INF Tech accessed advanced Nvidia chips in Jakarta, despite US export controls.

Elsewhere in AI anxiety:

• US schools are increasingly adopting AI-powered monitoring tools like GoGuardian to scan students’ chatbot conversations for signs of self-harm.

• A group of lawyers has documented 533 cases of AI misuse in legal filings, including fabricated case law citations, even as judges and bar associations permit AI use.

• Seven lawsuits in California claim ChatGPT encouraged dangerous discussions that led to suicides and harmful delusions.

• And in DeepSeek’s first public appearance since R1’s success, a senior researcher told a state-run conference he was pessimistic about AI’s impact on humanity.

End of an era

Yann LeCun, Meta’s chief AI scientist and Turing Award winner, is planning to leave the company to start his own venture.

The big picture:

• LeCun’s exit represents a fundamental clash of visions: he believes current large language models will never achieve human-level reasoning - yet Zuckerberg has bet Meta’s future on rapidly deploying LLM-based products.

• The Frenchman will be following another AI pioneer, Fei-Fei Li, in developing “world models” - AI systems that learn from spatial and video data rather than text. Li’s startup, World Labs, debuted Marble, its first publicly available model, which turns prompts and photos into editable 3D environments.

• With mounting Wall Street pressure on Zuckerberg’s $100B+ AI spending, Meta appears to be sacrificing long-term moonshots for short-term competitive positioning - disbanded the patient, academic approach of the FAIR lab in favor of exclusive teams poaching talent from OpenAI and Google.

Elsewhere in the FAANG free-for-all:

• LinkedIn rolls out AI-powered people search to Premium subscribers in the US, with plans to expand globally in the coming months.

• Alibaba plans to rename its AI app Tongyi to Qwen and add agentic AI shopping features for platforms like Taobao.

• A deep dive examines Microsoft’s AI strategy, including its OpenAI deal, data center investments, GitHub Copilot, MAI models, and Maia chip.

• Google unveils Private AI Compute, a cloud platform providing a secure space to run AI tools on devices, similar to Apple’s Private Cloud Compute.

• And Google rolls out Nano Banana image editing upgrades in Google Photos, including a “Help me edit” feature that lets users make edits using text or voice prompts.

Elsewhere in frontier models:

• OpenAI rolls out GPT-5.1 Instant, which is “warmer by default”, and GPT-5.1 Thinking, which is “easier to understand and faster”, starting with paid subscribers.

• Baidu unveils Ernie 5.0, an AI model to process and generate text, images, audio, and video, claiming it beats GPT-5-High and Gemini 2.5 Pro on some benchmarks.

• Google DeepMind unveils SIMA 2, a video-game-playing agent built on top of Gemini to navigate and solve problems inside 3D virtual worlds like Goat Simulator 3.

• OpenAI is piloting group chats in ChatGPT, with up to 20 users able to prompt ChatGPT in a shared space.

• And Meta releases Omnilingual Automatic Speech Recognition, a suite of AI models handling automatic speech recognition for 1,600+ languages, vs. OpenAI Whisper’s 99.

Moonshots vs Margins

Financial documents reveal that Anthropic expects to break even by 2028 while OpenAI won’t turn profitable until 2030, reflecting fundamentally different strategies in the AI race - one focused on sustainable growth, the other on ambitious moonshots.

The big picture:

• OpenAI is spending aggressively to maintain its lead, with projected losses of $74 billion in 2028 and commitments totaling $1.4 trillion over eight years. Anthropic is taking a more conservative path, focusing on enterprise customers (80% of revenue), avoiding costly ventures into image and video generation, and growing costs more in line with revenue.

• And yet - Anthropic is still getting into the infrastructure game, with a $50 billion plan to build AI data centers in partnership with Fluidstack, starting with facilities in Texas and New York.

• The divergence is testing investor patience at a critical moment, with tech stocks being punished over AI spending concerns and skepticism mounting over the industry’s massive financial commitments. Anthropic’s path to profitability may prove prescient - or OpenAI’s massive upfront investment could cement its market dominance if demand continues surging.

Elsewhere in AI infrastructure:

• Baidu unveiled two AI chips: the M100 for efficient MoE inference coming in early 2026, and the M300 for training super-large multimodal models coming in 2027.

• Several of Asia’s top tycoons and conglomerates are joining the data center race as tech giants plan $240B in APAC hyperscale expansion over the next five years.

• Microsoft will spend $10B to build a data center park in Sines, Portugal, in partnership with Portuguese developer Start Campus and UK startup Nscale.

• Meta says it will invest $600B in US infrastructure and jobs by 2028, with Zuckerberg telling Trump at a September dinner that Meta will spend “at least $600B.”

• And Elon Musk says Tesla is “probably going to have to build a gigantic chip fab“ for AI chips and “maybe, we’ll do something with Intel.”

Things happen

Intel CTO and AI officer Sachin Katti is leaving to join OpenAI. A inside look working at Cursor for 60 days. Six models traded $10K on crypto derivatives; Qwen3 Max lost $652 while GPT-5 lost $5,679. Amazon’s House of David Season 2 had 350 to 400 AI-generated shots. Cursor crosses $1B in annualized revenue. Free access to ChatGPT and Gemini suggests India may be the biggest bet for broad AI adoption. SoftBank sold its entire Nvidia stake for $5.8B in October. OpenAI is piloting group chats in ChatGPT in Japan, New Zealand, South Korea, and Taiwan. Pakistani newspaper mistakenly prints AI prompt with the article. AI is Dunning-Kruger as a service. ElevenLabs’s new Iconic Voice Marketplace for licensing AI versions of famous voices like Michael Caine’s. Kagi Search launches SlopStop, a community-driven AI slop detector. Nano Banana can be prompt engineered for nuanced AI image generation. How China ramped up its AI development from spring 2024 to catch the US. OpenAI’s Sora 2 floods social media with videos of women being strangled. Nasdaq had its worst 5-day run since April as AI stocks fell by about $800B.

Last week’s roundup

Agent provocateur

Amazon is suing Perplexity to block the startup’s Comet browser agent from making purchases on behalf of users on Amazon.com, marking one of the first major legal battles over AI agents’ ability to perform real-world tasks autonomously.

The big picture:

• This lawsuit could set critical precedents for how far “agentic AI” can go in acting on users’ behalf - Amazon argues Perplexity is committing computer fraud and violating its TOS by disguising AI agents as human users; Perplexity counters that agents should have “all the same rights and responsibilities” as the humans who deputize them.

• As AI shopping assistants proliferate, they may become a competitive threat to Amazon’s advertising empire. If AI agents shop autonomously for users, the lucrative business of selling prominent product placement in search results potentially loses its value.

• However, new research reveals that current models struggle with too many choices, are vulnerable to manipulation, and have poor collaboration skills even when unsupervised.

Subscribe now

Elsewhere in the FAANG free-for-all:

• Amazon is testing Kindle Translate, an AI tool that automatically translates books into other languages for self-publishing authors.

• Microsoft AI CEO Mustafa Suleyman says Microsoft plans to focus on superintelligence that prioritizes human control and will lead a new superintelligence team.

• Google says Ironwood, its seventh-generation TPU, will launch in the coming weeks and is more than 4x faster than its sixth-gen TPU.

• Apple is finalizing a deal to pay Google ~$1B per year for a 1.2T-parameter Gemini model to help power the new Siri, which is on track for next spring.

• And Google unveils Project Suncatcher, aiming to launch two solar-powered satellites carrying four TPUs each into low Earth orbit in 2027 with partner Planet Labs.

Backstop backlash

OpenAI CFO Sarah Friar sparked a swift political backlash after discussing the idea of a federal “backstop” to help finance the buildout of AI compute.

Between the lines:

• The remarks prompted White House AI czar David Sacks to declare there will be “no federal bailout for AI,” signaling that the Trump administration is willing to help through regulatory relief rather than financial support.

• But Friar’s word choice - and hasty retreat - reveals the precarious position OpenAI finds itself in after signing massive infrastructure deals that have raised serious questions about its ability to pay for them. Even CEO Sam Altman stepped in to try to defuse the situation.

• How we resolve this tension remains an open question - the government wants American companies to dominate globally, but there’s clearly political liability in creating AI giants that are “too big to fail.”

Elsewhere in OpenAI:

• OpenAI and AWS signed a seven-year, $38B deal for AI compute, including training models using Amazon’s data centers and CPUs.

• OpenAI released a blueprint for teen AI safety standards with five suggestions the company hopes lawmakers will use in crafting regulations.

• Sam Altman discussed managing OpenAI’s growth, delegating work, hiring hardware talent, GPT-6 enabling scientific research, and AI’s societal challenges in a Q&A.

• Court documents reveal that in a deposition, Ilya Sutskever discussed conflicts at OpenAI that he sent to board members before Sam Altman’s firing, his OpenAI exit, and more.

• And CODA, a Japanese anti-piracy group representing Studio Ghibli and Square Enix, demands that OpenAI stop using their copyrighted content to train Sora 2.

Elsewhere in AI infrastructure:

• Meta, Alphabet, Microsoft, and Amazon announced plans to increase their combined AI spending to $400B in 2026 in their latest earnings reports.

• Tech companies, including Meta and xAI, are using SPVs to raise tens of billions for AI data centers, allowing them to keep the debt off their balance sheets.

• Microsoft CEO Satya Nadella says the company faces a power shortage, not a compute one, which could leave chips sitting in inventory that can’t be plugged in.

• Oxford Economics reports that the AI boom is lifting global trade, with approximately 60% of US data center capex spent on imported IT gear mainly from Taiwan, South Korea, and Vietnam.

• Samsung is partnering with Nvidia to build an “AI Megafactory” and deploy over 50,000 of Nvidia’s most advanced GPUs to embed AI in its chipmaking process.

• Google plans to build an AI data center on Australia’s remote Christmas Island after signing a cloud deal with the Department of Defense earlier in 2025.

Huawei or the highway

Despite reaching a market cap of $5 trillion, Nvidia’s future now looks less rosy, as the Trump administration officials declined to sell advanced Blackwell chips to China ahead of the Trump-Xi summit in South Korea.

Where we’re headed:

• The decision comes as China bans foreign AI chips from state-funded data centers - potentially affecting over $100 billion in projects - while simultaneously slashing energy costs by up to 50% for facilities using domestic semiconductors (Huawei and other Chinese semiconductors require 30-50% more electricity than Nvidia’s chips to perform the same tasks).

• Huang’s unusually blunt warnings - criticizing Western “cynicism” and predicting China’s victory in the AI race - reflect Nvidia’s dire position in a market where its share has plummeted from 95% in 2022 to zero today.

• With Trump planning an April visit to China and Huang vowing to continue his late-night lobbying calls, the current stalemate appears temporary - but each side’s hardening position suggests the AI cold war is entering a new phase of permanent technological decoupling rather than managed competition.

Elsewhere in AI geopolitics:

• Senators Mark Warner and Josh Hawley unveiled a bipartisan bill that would require companies and US agencies to submit quarterly reports on AI-related job impacts.

• At APEC, Xi Jinping proposed a World Artificial Intelligence Cooperation Organization for global AI regulation, which Chinese officials said could be based in Shanghai.

• Google removed its AI model Gemma from AI Studio after Senator Blackburn said it fabricated sexual misconduct accusations against her.

• The US CBP signed a two-year deal with AI logistics company Altana to use its platform for real-time trade enforcement and forced labor detection.

• And Nvidia will supply 260K+ of its most advanced AI chips to the South Korean government and major companies, including Samsung, Naver, and SK Group.

Elsewhere in AI anxiety:

• xAI and Gab position their ideology-driven AI chatbots Grok and Arya as alternatives to mainstream chatbots accused of liberal bias like ChatGPT and Gemini.

• Some podcasters are embracing AI voice clones from ElevenLabs and other startups to augment, translate episodes, and even replace in-studio performances.

• The UK’s High Court ruled against Getty in its lawsuit against Stability AI over its image generator, with the judge calling the ruling “extremely limited in scope.”

• Coca-Cola released new AI-generated ads for the 2025 holidays after finding people didn’t know or care about the use of AI, despite criticism for its 2024 ads.

• Online harassers are using AI tools to create more realistic death threats, posting AI-generated images and sounds that simulate victims in violent situations.

• And arXiv will step up its moderation of CS review papers not vetted by an academic journal or conference after a surge in AI-generated papers.

Things happen

Perplexity launches a free AI-powered patent research tool in beta. Google in talks to value Anthropic at $350B+. Small AI models do the real work. Grokipedia mirrors Wikipedia except on politics, but AI makes it better on obscure entries. Chinese startup Moonshot releases Kimi K2 Thinking, claiming it beats GPT-5 for $4.6M. OpenAI now has 1M+ business customers globally. Lessons on deploying AI agents in production. Gemini Deep Research can now draw on your Gmail, Drive, and Chat to generate reports. Nebius unveils Token Factory to compete with AWS and Azure. AI designs functional antibodies from scratch, no animals required. OpenAI’s o1 analyzes languages as well as a human expert. Deutsche Bank is exploring strategies to hedge its exposure to data centers. What jobs AI is replacing today. AI coding ability doubles every six months, but real-world tasks have a messiness tax. You should write an agent. How AI turned Google Cloud from also-ran to growth driver. The man who invented AGI. Getty and Perplexity sign a multi-year licensing agreement for image distribution. A profile of Common Crawl, the nonprofit scraping the web for AI training data. The future of advertising is AI-generated ads personalized to you. AI’s dial-up era. How I use every Claude Code feature.

Last week’s roundup