← Anthropic
Wes Roth
The REAL Reason Andrej Karpathy Joined Anthropic
⏱ 25 min video · 3 min read24 May 2026Worth watching
TL;DR
Andrej Karpathy has joined Anthropic's pre-training team to use Claude to accelerate Claude's own training — a direct implementation of recursive self-improvement (RSI). The video argues this hire signals Anthropic is making an all-in bet on AI-automated AI research, not just acquiring a celebrity name.
Key points
1
Karpathy joined Anthropic on May 19, 2026, reporting to Nick Joseph (head of pre-training), specifically to build a team using Claude to accelerate pre-training research.
2
His open-source project Auto Research (released March 2026) demonstrated a working AI agent research loop: propose code change, run 5-minute experiment, evaluate metric, commit or revert — 700 experiments in 2 days yielded ~20 stackable improvements and an 11% training speed-up on a GPT-2 benchmark.
3
Anthropic co-founder Jack Clark has publicly forecast a 60%+ chance of fully automated AI R&D (no humans involved) by end of 2028, and Karpathy's hire appears to be the execution plan for that forecast.
4
Anthropic is simultaneously massively scaling compute through commitments with Google Cloud, SpaceX/xAI Colossus, and potentially Microsoft, making pre-training efficiency gains worth tens of millions of dollars.
5
There is a visible strategic split in the industry: most major AI figures (Sam Altman, Dario Amodei, Karpathy, Elon Musk, Ilya Sutskever, Sergey Brin) are betting on RSI via coding-capable LLMs, while Demis Hassabis appears to favor world models, and Yann LeCun rejects LLMs as a path to AGI entirely.
Key arguments
→
Karpathy did not join for status or pay — he joined because the technical challenge (using Claude to improve Claude) was uniquely compelling, suggesting RSI is now considered a concrete near-term engineering problem, not a theoretical one.
→
The Auto Research loop (propose, test, evaluate, commit/revert) is already working at small scale; the strategic question is what happens when run at frontier compute scale with Claude-level models — results within 6-12 months should confirm or deny the thesis.
→
Anthropic's focus discipline (no music, voice, video, or image models) combined with this RSI bet represents a high-conviction, concentrated strategy that has repeatedly paid off — worth tracking as a signal of where AI capability jumps will come from next.
Notable quotes
“Karpathy didn't join Anthropic to work on Claude or work on safety. He joined specifically to use Claude to make Claude better — to set that flywheel of recursive self-improvement to start it spinning.”
“Anthropic is not buying status. They are trying to compress the research cycle to produce the next iteration of Claude and all of the future Claude models.”
“This bet on recursive self-improvement might be the most impactful bet that humanity has ever made.”
Worth watching?
✅
Worth watching the full video?
Watch if you want the live commentary and Karpathy clips — the core analysis and all key details are captured here, but the video adds useful enthusiasm and pacing for a story that genuinely matters.
Saved you some time? The creator still deserves a like.
Want this for your own channels?
Add the channels you follow. Every new video summarised and in your inbox the moment it drops. From £4/month.
Try it free