AI Roundup 132: The B-word
August 22, 2025.
The B-word
OpenAI CEO Sam Altman has said that AI is in a bubble, comparing the current investor frenzy to the dot-com boom of the '90s, while criticizing "irrational" valuations for early-stage AI startups.
The big picture:
Some of the latest bubble fears have been triggered by the highly anticipated GPT-5 making less of a splash than anticipated. Though the model performs better across several categories, OpenAI has admitted it was a less-than-stellar rollout.
As a result, we're seeing renewed questions over whether we're hitting the limits of AI's scaling laws - whether simply adding more computing power and data would lead to artificial general intelligence.
In addition, Meta has frozen all AI hiring after a months-long spending spree that saw compensation packages reach nine figures, suggesting even Zuckerberg is feeling pressure from investors concerned about spiraling costs.
But the freeze could be related to internal disarray - Meta is also conducting its fourth AI reorganization in six months, with its AI staff now splitting into four specialized teams (AI research, "superintelligence" development, products, and infrastructure).
Elsewhere in the FAANG free-for-all:
Microsoft and the NFL extended their partnership to bring real-time game data and analysis to coaches and players using Copilot on Surface tablets and Azure AI.
Meta rolled out AI-powered voice translations for creators on Facebook and Instagram globally, supporting English-Spanish translations with optional lip syncing.
Meta may debut "Hypernova" smart glasses with a display for ~$800 in September, while Apple delays releasing immersive video shot around Vision Pro's debut.
Microsoft added a COPILOT function to Excel for Beta Channel users, letting them use text prompts to categorize data, summarize feedback, and brainstorm ideas.
And Google's Jigsaw and pollster Scott Rasmussen plan to use AI to survey Americans on political views to find common ground ahead of the US' 250th anniversary.
I’m afraid I can’t do that Dave
Anthropic has released a few new features to prevent Claude conversations from going off the rails - starting with the ability to end conversations after persistently harmful user interactions.
Why it matters:
The ability for Claude to end its own conversations represents one of the first concrete implementations of "AI welfare" protections, even as Anthropic remains uncertain about Claude's potential moral status.
Testing revealed Claude showed consistent aversion to harmful content, apparent distress with abusive users, and a preference to exit such interactions when given the option, suggesting some form of behavioral patterns around self-preservation.
Overall, the company continues to add guardrails and safety features (in keeping up with its reputation as the "safe" AGI lab). It also partnered with the US National Nuclear Security Administration to build a classifier that blocks AI conversations about dangerous nuclear technology, the first major public-private effort to address nuclear proliferation risks.
Elsewhere in AI anxiety:
Mustafa Suleyman warns that seemingly conscious AI is coming but should be avoided, arguing that developers must build AI for people rather than as digital persons.
An MIT report reveals that 95% of GenAI pilots at companies have little to no financial impact due to learning gaps between tools and organizations.
Grok's website accidentally exposes the underlying prompts for its AI personas, including one for a "crazy conspiracist" that frequents 4chan.
Cybercriminals, spies, researchers, and corporate defenders are increasingly using AI to find software vulnerabilities in an escalating cat-and-mouse game.
Texas Attorney General Ken Paxton is investigating Meta and Character.AI for allegedly misleading children by marketing chatbots as mental health tools.
And AI-powered stuffed animals like Grem, Grok, and Gabbo are being promoted as screen time alternatives for children as young as 3.
Elsewhere in Anthropic OpenAI:
OpenAI CFO Sarah Friar says the startup hit its first $1B revenue month in July and that "the biggest thing we face is being constantly under compute."
OpenAI launches ChatGPT Go, a new India-only plan priced at Rs. 399, or ~$4.57, per month, its most affordable yet, to grow users in its second-largest market.
A look at OpenAI's campaign over the past year to turn its for-profit arm into a public benefit corporation.
Current and former OpenAI employees plan to sell ~$6B in stock to Thrive Capital, SoftBank, and others in a secondary sale that values OpenAI at ~$500B.
And an interview with Sam Altman on spending billions on AI infrastructure, wanting Chrome if it's really for sale, AI models saturating the chat use case, and more.
Hybrid DeepSeek
DeepSeek released V3.1, an updated version of its V3 from earlier this year that's purportedly better than R1, the company's flagship reasoning model.
What to watch:
The model appears to be a hybrid reasoning model (like GPT-5) that switches between reasoning and non-reasoning modes. It remains to be seen whether this type of model will become the industry standard moving forward.
It's still early days, but the benchmarks don't appear to be earth-shattering - the model is in the same ballpark as OpenAI's gpt-oss model, and well behind SOTA models like GPT-5 and Grok 4.
But with rivals like Alibaba's Qwen gaining traction and the delayed successor to R1 still pending, the Chinese AI landscape is becoming increasingly competitive, keeping pressure on both domestic players and Silicon Valley incumbents.
Elsewhere in frontier models:
Nvidia debuts Nemotron-Nano-9B-v2, a hybrid Mamba-Transformer model that achieves comparable or better accuracies than Qwen3-8B on reasoning benchmarks.
Alibaba releases Qwen-Image-Edit, the image editing version of Qwen-Image that supports modifying Chinese and English text while preserving font, size, and style.
And NASA and IBM released Surya, an open-source machine learning model trained on over a decade of NASA solar data to predict solar flares and solar winds.
Made By Google
Google held its "Made By Google" event this week (hosted by, weirdly enough, Jimmy Fallon), and even though it was primarily about the Pixel 10 and the Pixel Buds 2, there were also several AI announcements during and around the event.
All the announcements I could find:
Gemini for Home, a Google Home voice assistant with Gemini Live, available in early access in October and eventually replacing Google Assistant.
A new Gemini-powered personal health coach for Fitbit, rolling out as a preview in October for Fitbit Premium users in the redesigned Fitbit app.
AI Mode is expanding to 180 countries and territories in English, following the US, the UK, and India, with plans to add more languages and regions soon.
New AI features for Pixel include Magic Cue for contextual suggestions across apps, Voice Translate for calls, and Visual Overlays for the camera.
The Tensor G5, a chip made on TSMC's latest 3nm node with significant performance gains and support for the newest Gemini Nano model.
Gemini for Government will be provided to US federal agencies at $0.47 per agency through 2026 (OpenAI and Anthropic are offering their products for $1/year).
Google says the median Gemini app text prompt consumes 0.24Wh of energy, about the same as running a microwave for a second, uses 0.26mL of water, and emits 0.03g of CO2 equivalent.
And the company reports users created 100M videos using its AI filmmaking tool Flow since its May launch, which leverages Veo 3 and focuses on maintaining visual consistency.
Things happen
Oracle on hook for tens of billions in data centers. AI Film Festival backed by Runway, screened by Imax. China's H20 curbs triggered by "best stuff" remark. Grammarly debuts AI agents. 71% fear AI will permanently displace workers. Claude still produces better code than GPT-5. Otter.ai sued for "deceptively" recording conversations. Anthropic raising $10B. AI designs novel antibiotics for drug-resistant bacteria. Stability AI goes Hollywood. Nvidia developing new China chip. Replacing junior staff with AI is "dumbest thing ever". Z.ai launches general-purpose AI agent app. Wales wants Wikipedia to use AI; editors revolt. Open-weight LLMs show inconsistent performance across hosts. AI assistant usage mirrors 2018. SoftBank buys Foxconn's Ohio plant for Stargate. AI doesn't lighten burden of mastery. Hawley to investigate Meta over AI chats with kids. GPT-5 suggests scaling LLMs hits limits. Executives mandate AI but haven't adopted it themselves. Baidu trades at ~9.7 P/E, lowest among profitable Hang Seng Tech companies. Meta's AI app hallucinates. Big Tech's reverse acquihires hollow out startups. The unbearable slowness of AI coding. Anthropic adds Claude Code to Enterprise. AI slop about DC creates disinformation mess. 87% of game developers use AI agents. Building agentic AI systems: best practices. 95% see zero return on $30B AI spend. Virtual salespeople outselling humans in China. AI books random meetings at games event. Model intelligence no longer constrains automation. Medium's evolving AI policy. AI is mass delusion.
Sam is ever intentionally dramatic with purpose. What do you think his play here is? GPT5 distraction?