Andrej Karpathy announced today that he is joining Anthropic. If that name does not mean anything to you, here is why it matters. Karpathy is a founding member of OpenAI. He helped build the company Sam Altman now runs. He left in 2017 to lead AI at Tesla, where he ran the Full Self-Driving and Autopilot programs. He went back to OpenAI for a year. He left again to start an AI education company. And today he announced he is joining the one lab that exists because other people left OpenAI before him. Anthropic was founded by OpenAI exiles. Today it added the most famous one yet. Karpathy posted the news on X himself: "Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative." Confirmed by TechCrunch, CNBC, Axios, and Reuters within hours. This is not a routine hire. This is one of the most respected researchers alive walking from the company he started to the company trying to beat it. And the job he is taking is the part that should make you sit up.

Karpathy is joining Anthropic's pre-training team under team lead Nick Joseph. Pre-training is the massive, expensive, compute-intensive phase that gives Claude its core knowledge and capabilities. It is the foundation everything else is built on. But that is not the headline.

The headline is what Anthropic told reporters about his specific mission. Karpathy is building a new team focused on using Claude to accelerate pre-training research. Read that slowly. He is being hired to use the current version of Claude to make the next version of Claude better.

That is recursive self-improvement. AI improving AI. If Anthropic can use today's Claude to make the next training run even 5 to 10 percent more efficient, and then do it again with the model that produces, the gains compound. Each generation of the model becomes a slightly better researcher for building the next one. The industry has a dramatic name for this. Take-off. The road to the singularity. In practice it is applied engineering with a measurable feedback loop, and Anthropic just hired one of the few people on earth who can bridge LLM theory and large-scale training to lead it.

Tapping Karpathy for this specific role is a deliberate signal. Anthropic is betting that AI-assisted research, not raw compute alone, is how it stays ahead of OpenAI and Google. Everyone has compute now. Anthropic locked in commitments from Amazon, Google, and SpaceX over the last month. The bottleneck is no longer chips. The bottleneck is the small pool of researchers who can actually push the frontier. Karpathy is at the very top of that pool.

While the Karpathy news broke, Google held its biggest event of the year. The I/O 2026 keynote shipped a tidal wave of AI. Here is what actually matters.

Gemini 3.5 Flash dropped today. It beats Gemini 3.1 Pro on coding, agentic, and multimodal benchmarks while running 4x faster on output than other frontier models. Live right now in the Gemini app, Search, Antigravity 2.0, and the API. Gemini 3.5 Pro is in testing and ships next month.

Gemini Spark is Google's new 24/7 AI agent. It anchors a new $100 per month AI Ultra plan aimed at developers, creators, and power users. This is Google's direct shot at the agent category that Anthropic and OpenAI have been fighting over.

Gemini Omni is real. The video model that leaked two weeks ago shipped. Omni Flash takes image, audio, video, and text as input and outputs editable video grounded in real-world knowledge. Rolling out to AI Plus, Pro, and Ultra across the Gemini app, Google Flow, and YouTube Shorts. The Sora vacuum just got filled.

Android XR smart glasses. Google previewed Samsung-built Gemini glasses launching this fall, demoed live with Nano Banana photo editing on stage.

Code Mender. Demis Hassabis announced an AI tool that automatically finds and patches security vulnerabilities. Sound familiar? It is Google's answer to Claude Mythos. Hassabis closed his segment by saying "AGI is now on the horizon."

One keynote. Three shipped models, a new agent, smart glasses, and an AGI declaration. Google is not playing catch-up quietly anymore.

If you use Claude, today is good news long-term. Hiring Karpathy to make Claude train better Claude means the model you rely on is about to improve faster. The pre-training gains compound over every future version.

If you use Gemini, you have real new tools today. Gemini 3.5 Flash is live and fast. Gemini Omni brings editable AI video into apps you already use. If you are a power user, the $100 Ultra plan with Gemini Spark is Google's most serious agent offering yet.

If you work in AI, the talent war just became the main event. Compute is commoditized. Capital is flowing freely. The scarce resource is people who can build frontier models, and the flow of those people is pointing toward Anthropic. If you are choosing where to build your career, the market is sending a loud signal.

If you watch the race, today was a split screen. Google shipping a tidal wave of product on one side. Anthropic quietly collecting the people who build the products on the other. Both strategies are valid. The next year tells us which one compounds faster.

One section, one prediction, stated plainly.

Here is what we think happens next, and how confident we are.

Karpathy's hire is not really about Karpathy. It is about what Anthropic is signaling: that the next leap comes from AI improving AI, not from buying more chips. If that bet is right, the gap between the labs stops being about who has the most compute and starts being about who has the best feedback loop between their current model and their next one.

Our call: within 12 months, at least one frontier lab announces a model where a meaningful chunk of the training improvements came from its own previous model doing the research. Anthropic just hired the person to make that happen first. Confidence: high.

The counter-case, so you have both sides: recursive self-improvement has been promised before and the compounding gains have a habit of flattening out faster than the hype suggests. Karpathy himself spent part of this year warning about AI hype before deciding he could not sit out the frontier. If the loop stalls, this becomes a great hire that did not bend the curve.

Either way, you heard the framing here first. Bookmark this one. We will tell you in a year if we were right.

Andrej Karpathy, a co-founder of OpenAI and former Tesla AI lead, joined Anthropic today. He is building a team to use Claude to accelerate the training of the next Claude, which is recursive self-improvement in practice. He joins a wave of elite hires including ex-xAI founding member Ross Nordeen and Meta cybersecurity veteran Chris Rohlf. It lands the same week Anthropic closes in on a $900 billion valuation and days after Musk lost his case against OpenAI. Meanwhile Google I/O shipped Gemini 3.5 Flash, the Gemini Spark agent on a new $100 Ultra plan, the Omni video model, Android XR glasses, and an AGI-is-on-the-horizon declaration from Demis Hassabis. The model war is loud and public. The talent war is quiet and decisive. Today Anthropic won the quiet one, and the quiet one might matter more.

Google I/O Recap For AI 👉 Here

Stay building. 🤖