On May 19, Andrej Karpathy posted seven sentences on X. Co-founder of OpenAI. Former Tesla AI Director. The man who taught more people how AI actually works than anyone alive.
"I've joined Anthropic."
Every newsletter in the world covered it as a talent war story. OpenAI vs Anthropic. KD joining the Warriors. Famous researcher switches sides.
They all missed what he was actually hired to do.
Presented by Atoms.dev
The founders in this issue don't wait for the next model. They build now.
Tell Atoms your idea. A team of AI employees - PM, engineer, SEO specialist, data analyst that builds it, deploys it, and helps you acquire customers. 165,000 GitHub stars. #1 Product Hunt of the Week. Trusted by Amazon, Microsoft, and NVIDIA.
What Karpathy was actually hired to do
Nicholas Joseph, Anthropic's Head of Pretraining, confirmed the role within hours of the announcement. The exact language Anthropic gave to TechCrunch, Axios, and VentureBeat was identical:
"Karpathy will start a new team focused on using Claude to accelerate pretraining research itself."
Not using humans to improve Claude. Using Claude to improve Claude. Read that one more time.
Claude code written by AI
100%
Per Anthropic CPO, Feb 2026
Karpathy's mandate
RSI
Recursive self-improvement
SWE-Bench score today
93.9%
Was 2% in late 2023
Time to get there
26 mo
2% → 93.9% in 26 months
AI coding capability - SWE-Bench score 2023 to 2026 (%)
The self-improving loop: Claude 3 helped train Claude 4 → Claude 4 is helping train Claude 5 → Karpathy's job is to make that loop go faster.
Why this hire is different
Anthropic could have hired a hundred competent infrastructure engineers. It hired the person whose career has been a tour of the exact problem. Tesla (neural nets at production scale), OpenAI 2023 (mid-training and synthetic data), Eureka Labs (compressing complex ideas into teachable form). That combination is unusually well-suited to making a model help train the next model.
The number Jack Clark said out loud
Three weeks before Karpathy was hired, Anthropic co-founder Jack Clark published Import AI Issue 455. One sentence should have stopped the industry cold.
Jack Clark, Oxford University, May 20 2026
"My prediction is by the end of 2028, it's more likely than not that we have an AI system where you would be able to say to it: 'Make a better version of yourself.' And it just goes off and does that completely autonomously."
Jack Clark's public RSI probability estimates
Eliezer Yudkowsky's response was four words: "Then you'll die with the rest of us."
That exchange is worth sitting with. Clark is not a doomer posting from the sidelines. He co-founded Anthropic, one of the two or three labs most likely to build the system he is describing. When someone in that position assigns 60% odds on a public three-year timeline, that is not a casual take.
How we got here, faster than anyone predicted
(Mar 2025) Dario Amodei predicts AI will write 90% of code within 6 months. Industry calls it impossible.
(Oct 2025) Amodei confirms the prediction came true: "70, 80, 90% of the code written at Anthropic is written by Claude."
(Feb 2026) CPO Mike Krieger at the Cisco AI Summit: "Right now for most products at Anthropic, it's effectively 100%. Claude is being written by Claude."
(May 4 2026) Jack Clark publishes Import AI 455: 60% probability AI trains its own successor by 2028.
(May 19 2026) Karpathy joins Anthropic specifically to build the team making the self-improvement loop run faster.
Presented by Atoms.dev
The founders who understand this timeline are building AI-native businesses right now - not waiting. Atoms gives you the full team to do it without hiring.
Three things to do this week
Here is where every other Karpathy piece stopped. They told you what happened. This is what to do about it.
Shorten your roadmap window to 6 months.
Most founders build 12-month roadmaps assuming roughly linear capability improvement. That assumption is wrong. SWE-Bench went from 2% to 93.9% in 26 months. If the self-improvement loop accelerates that curve further, the AI tools available to you in 18 months will be unrecognisable from today. Build in 6-month windows, not 12.
Audit your moat right now.
If your advantage is "we use AI better than competitors," that moat has a half-life measured in months. AI capability is a floor that rises under everyone. The founders who survive this are building in places AI cannot easily reach: proprietary data, distribution, regulatory depth, deep customer relationships. Check which category your moat falls into today.
Run one workflow through AI completely this week.
Not assisted. Not co-piloted. Pick one function in your business and replace the human output entirely with AI output for one week. Klarna replaced humans entirely and it failed. The founders winning right now found the specific functions where full replacement works. The only way to find yours is to test it.
Karpathy's seven sentences were the most covered AI story of the week.
Almost nobody read the eighth word.
"Using Claude to accelerate pretraining research itself" — that is the sentence that matters. The man who taught the world how AI works just took a job making AI teach itself. The co-founder of the company he joined puts 60% odds on that running without humans by 2028.
This is not a talent war story. It is a story about the timeline every founder is operating on. Most of the assumptions baked into your current roadmap were made before this week. Some of them no longer hold.
One founder you know is still planning like it's 2023. Forward this before their next roadmap session.
📧 Forward this to 3 entrepreneur friends who need to see this opportunity
