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.
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.
