Andrej Karpathy — OpenAI co-founder, former Tesla AI chief, and arguably the most influential AI educator alive — joined Anthropic on May 19 to lead a team with a specific mandate: use Claude to accelerate the research that produces the next Claude. It is the clearest signal yet that the race to automate AI research itself has moved from speculative blog posts to staffing decisions at frontier labs.
What actually happened
Karpathy started at Anthropic on May 19, working on the pre-training team under Nick Joseph — himself a former OpenAI researcher. Pre-training is the large-scale, compute-intensive phase that gives Claude its core knowledge and capabilities. It is also, by a wide margin, the most expensive part of building a frontier model.
"I've joined Anthropic," Karpathy posted on X. "I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D."
An Anthropic spokesperson confirmed to TechCrunch that Karpathy will start a team focused on using Claude to accelerate pre-training research. Nick Joseph welcomed the hire on X: "He'll be building a team focused on using Claude to accelerate pretraining research itself. I can't think of anyone better suited to do it."
Karpathy isn't arriving alone. Anthropic has been on a hiring spree: Ross Nordeen, a former Tesla engineer and founding member of Elon Musk's xAI, also joined earlier in May. On the same day as Karpathy's announcement, cybersecurity veteran Chris Rohlf — with over 20 years of experience including six years at Meta — joined Anthropic's frontier red team to stress-test advanced models against severe threats. This follows Anthropic's broader push into cybersecurity through Project Glasswing.
The autoresearch playbook
Karpathy didn't arrive at Anthropic with just a CV. He arrived with a working prototype.
In early March 2026, he released autoresearch, a 630-line open-source project that encodes a deceptively simple idea: give an AI coding agent a training script, a frozen evaluation metric, and a fixed five-minute compute budget per experiment. The agent proposes a change, trains, checks if the result improved, keeps or discards, and repeats — indefinitely, while the researcher sleeps.
"All LLM frontier labs will do this. It's the final boss battle," Karpathy wrote on X. "You spin up a swarm of agents, you have them collaborate to tune smaller models, you promote the most promising ideas to increasingly larger scales, and humans (optionally) contribute on the edges."
The structural principle is what his new team at Anthropic will apply at frontier scale: rather than a single researcher proposing one experiment at a time, a network of Claude agents runs parallel experiments with validated improvements accumulating in a ratchet that can only move forward. As TechTimes noted, if the bet is correct, "the lab that executes it best will compound its lead in each successive training run, because each generation of models will have been built with the help of a more capable predecessor."
Why this matters more than a talent acquisition
This is not merely a prestige hire. It is Anthropic staking a strategic position on a specific thesis: that AI-assisted research, not just raw compute, is how you win the frontier model race.
Anthropic co-founder Jack Clark laid out the intellectual framework two weeks before Karpathy's hire. In Import AI 455, published May 4, Clark wrote: "I now believe that recursive self-improvement has a 60% chance of happening by the end of 2028. In other words, AI systems might soon be capable of building themselves."
Clark was careful with caveats — he pegs the 2027 probability at just 30%, noting that AI research still requires "creativity and heterodox insights" that models haven't yet demonstrated at a transformative level. But the directional signal is unmistakable: someone with line of sight into the next two model generations believes we are approaching a threshold where each Claude generation materially accelerates the creation of its successor.
If they can use the current generation of Claude to make each training run even 5–10% more efficient, and do that repeatedly, they get compound returns. That's not science fiction. It's applied engineering with a measurable feedback loop.
Every lab is building the same loop
Anthropic isn't alone in this bet. Every major lab has now confirmed, directly or indirectly, that its models are participating in the construction of their own successors.
When OpenAI launched GPT-5.3-Codex in February 2026, the announcement included a sentence that deserved far more attention: "GPT-5.3-Codex is our first model that was instrumental in creating itself." Sam Altman posted on X: "It was amazing to watch how much faster we were able to ship 5.3-Codex by using 5.3-Codex." Altman has previously outlined two internal milestones: an intern-level AI research assistant by September 2026, and a true automated AI researcher by March 2028.
Google DeepMind is approaching the same territory from the scientific research angle. A Nature paper published May 19 — the same day as Karpathy's announcement — introduced ERA (Empirical Research Assistance), an AI system that discovered 40 novel methods outperforming the top human-developed approaches on a bioinformatics leaderboard and 14 models that beat the CDC ensemble for COVID-19 forecasting.
The competitive dynamic is clear: the lab that closes the loop first — where model N meaningfully accelerates the creation of model N+1 — gains a compounding advantage that grows with each generation. Karpathy is now the person most visibly tasked with closing that loop for Anthropic.
The business angle
If you're a business owner reading this and thinking "recursive self-improvement sounds like a problem for researchers, not for me," consider the practical implication: the AI tools you're evaluating today are about to improve much faster than you expect.
When models improve through human research alone, progress follows a roughly predictable curve. When models begin materially accelerating their own development, that curve steepens. The gap between "good enough" and "indispensable" AI tools could collapse from years to quarters — which changes the calculus on when to invest in AI integration.
This is particularly relevant for businesses already using Claude or considering agentic AI workflows. Anthropic is already on track for profitability and racing toward an IPO. The Karpathy hire signals that the company believes the competitive moat is shifting from "who has the most compute" to "who can make their models improve themselves fastest." If that thesis holds, Claude's capabilities could advance in step-changes rather than incremental updates.
What to watch
The key milestone is straightforward: does the next major Claude release arrive faster and perform meaningfully better than it would have without AI-assisted research? That's an inherently hard thing to measure from the outside, but watch for Anthropic's training efficiency metrics and release cadence over the next six to twelve months.
The safety dimension is equally important. As Clark himself noted, alignment techniques that work today "may break under recursive self-improvement as the AI systems become much smarter than the people or systems that supervise them." A technique that's 99.9% accurate drops to roughly 95% after 50 generations and to about 60% after 500, according to his analysis. Chris Rohlf's simultaneous hire to stress-test models suggests Anthropic is at least aware of the stakes.
For Australian businesses, the practical takeaway is this: the AI landscape isn't just evolving — it's learning to evolve itself. The window for cautious observation is narrowing. The tools you evaluate in Q3 may be substantially more capable than what you see today, and the organisations that build AI fluency now will be better positioned to absorb those step-changes when they arrive.
Sources
- OpenAI co-founder Andrej Karpathy joins Anthropic's pre-training team — TechCrunch
- Andrej Karpathy joins Anthropic to lead AI-accelerated pretraining research — AI Chat Daily
- Karpathy, Who Called Today's AI Agents 'Slop,' Joins Anthropic to Use Claude to Build the Next Claude — TechTimes
- Import AI 455: AI systems are about to start building themselves — Jack Clark, Import AI
- Andrej Karpathy Joins Anthropic: What Happens Next — The Algorithmic Bridge
- While You Were Sleeping, an AI Ran 700 Experiments and Improved Itself — The Nov Tech
- An AI system to help scientists write expert-level empirical software — Nature
