OpenAI hires OpenClaw developer, as Anthropic safety leader quits to write poetry
🎆 Tired of having to explain AI stuff to your coworkers? Share Handy AI with them so that they can get the most important AI news delivered weekly to their inbox (in addition to our high quality editorials).
last week’s top stories
👨💻 OpenClaw creator Peter Steinberger joins OpenAI to build personal agents. Peter Steinberger, the developer behind the viral open-source AI agent OpenClaw (100k+ GitHub stars, 2 million visitors in one week), joined OpenAI on February 15 to “drive the next generation of personal agents.” His reasoning was refreshingly candid: he said OpenClaw could have been a “huge company,” but he cares more about changing the world, and teaming up with OpenAI is the fastest path to getting personal AI agents into everyone’s hands. OpenClaw will transition to a foundation and remain open source with OpenAI’s backing. Read more
✍️ Anthropic’s AI safety lead quits to write poetry, warns “the world is in peril”. Mrinank Sharma, who led Anthropic’s Safeguards Research Team, resigned on February 9 and published a letter that reads more like prose than a corporate goodbye. He wrote that he’s “repeatedly seen how hard it is to truly let our values govern our actions” both within himself and at Anthropic, where there are “constant pressures to set aside what matters most.” Sharma plans to pursue a poetry degree in the UK, citing Rilke as a major influence. Read more
🎬 Disney hits ByteDance with cease and desist over Seedance 2.0’s viral deepfakes. ByteDance’s Seedance 2.0 video generator launched and almost immediately went nuclear, with users flooding social media with uncannily realistic deepfakes of Tom Cruise fighting Brad Pitt and alternate Stranger Things endings using recognizable copyrighted characters. Disney fired back on February 13 accusing ByteDance of treating Star Wars, Marvel, and Family Guy IP as “free public domain clip art,” while the MPA and SAG-AFTRA piled on calling it “an attack on every creator.” Read more
If you didn’t know, would you say this was AI?
💰 Anthropic closes a staggering $30 billion Series G at $380 billion valuation. The numbers here are genuinely hard to process: Anthropic raised $30 billion on February 12, more than doubling its valuation from $183 billion just five months earlier in its Series F. Led by Singapore’s GIC sovereign wealth fund and Coatue, with Microsoft and Nvidia contributing significant capital, the round was fueled by revenue run rate surging to $14 billion (Claude Code alone generates over $2.5 billion). Five months between funding rounds at these dollar amounts is uncharted territory for any company, AI or otherwise. Read more
🚀 OpenAI’s Codex Spark hits 1,000 tokens per second on Cerebras hardware. GPT-5.3-Codex-Spark dropped on February 12 as a research preview for Pro users, and the headline number is wild: over 1,000 tokens per second, roughly 15x faster than its big sibling GPT-5.3-Codex. Built on Cerebras’ wafer-scale inference chips, the model features a 128k context window and still scores competitively on agentic coding benchmarks despite being substantially smaller. It’s text-only for now and available through the Codex app, CLI, and VS Code extension. Read more
🚪 Half of xAI’s founding team has now walked out the door. The exodus at xAI reached a new milestone last week: 6 of the original 12 founders have departed, with co-founders Tony Wu and Jimmy Ba both leaving within 24 hours of each other. Musk held a 45-minute all-hands meeting to restructure the company into four teams (Grok, Coding, Imagine, and a new project called “Macrohard”), framing the departures as routine reorganization. Read more
🥑 Meta’s closed-source Avocado model has finished pre-training. Leaked internal memos confirm Avocado demonstrates 10x compute efficiency over Llama 4 Maverick and 100x over Llama 4 Behemoth while matching post-trained performance in knowledge, visual perception, and multilingual capabilities. Zuckerberg is personally supervising the (closed-source) project at Meta Superintelligence Labs, staffed with recruits from OpenAI, Google, and Anthropic. Read more
📢 ChatGPT now has ads, and the first major safety researcher already quit over it. OpenAI began testing ads in ChatGPT on February 9 for free-tier and Go subscribers in the US, targeted based on conversation topics and past interactions, with paid plans staying ad-free. The same day, safety researcher Zoë Hitzig resigned and published a New York Times essay comparing the move to Facebook’s early privacy erosion, warning that ChatGPT holds “an archive of human candor that has no precedent.” Read more
🎖️ Report: Claude was used in the US military operation to capture Maduro. According to a Wall Street Journal report from February 13, Anthropic’s Claude AI was deployed during the US raid that captured Venezuelan President Nicolás Maduro in January, accessed through Anthropic’s partnership with Palantir. This puts Anthropic in an uncomfortable spot: the company’s own usage policies explicitly prohibit Claude from supporting violence or conducting surveillance, and CEO Dario Amodei has been vocal about concerns over autonomous lethal operations. The Pentagon has reportedly considered canceling up to $200 million in Anthropic contracts over the company’s resistance to removing safety restrictions. Read more
🗑️ OpenAI quietly disbands its mission alignment team. In a move that barely made headlines, OpenAI dissolved the six-to-seven-person team responsible for communicating the company’s mission to employees and the public. Team leader Joshua Achiam was reassigned to a new “chief futurist” role. OpenAI called it a routine reorganization, which is also what they said when they disbanded the superalignment team in 2024. Read more
🎥 Runway raises $315 million at $5.3 billion valuation to build “world models”. General Atlantic led the Series E with Nvidia, Fidelity, and Adobe Ventures participating, nearly doubling Runway’s previous valuation. Their Gen 4.5 model currently tops the Artificial Analysis Text to Video benchmark at 1,247 Elo, outperforming Google’s Veo 3 and OpenAI’s Sora 2 Pro. [Read more
🧪 AI Research of the Week
AI-rithmetic
From Google Research
Jake’s Take: Here’s a fun paradox: the same AI models winning medals at international math olympiads and proving novel mathematical lemmas still can’t reliably add two large numbers together.
This paper systematically tests Claude Opus 4.1, GPT-5, and Gemini 2.5 Pro on plain integer addition, and finds that accuracy falls off a cliff as digit count increases. The researchers dug into why and found that the vast majority of errors (up to 92% for Gemini) boil down to just two failure modes: the model misaligning the columns of digits it’s adding (often a quirk of how text gets chopped into tokens), and randomly botching the carry step.
It’s a genuinely humbling result that tells us something important about how these systems actually work under the hood: they’re not actually learning arithmetic as a set of rules the way we did in grade school. They’re pattern-matching their way through it, and when the patterns get long enough, they fumble. Worth keeping in mind every time someone tells you these models “understand” math.
and then, even more news…
🔍 A general-purpose GPT-5.3 was spotted hiding in OpenAI’s GitHub repo. Developer Charlie L. found references to a plain gpt-5.3 model sitting alongside the already-released gpt-5.3-codex in OpenAI’s public codebase on February 9, confirming OpenAI internally treats it as a distinct product. Early speculation points to a 400k token context window and 128k output limit with improved efficiency and reasoning over GPT-5.2. OpenAI hasn’t acknowledged or denied it, which in OpenAI terms usually means “it’s real but we’re not ready to talk about it yet.” Read more