Recursive self-improvement has started
Dwarkesh argues recursive self-improvement is already starting: AI automating AI research could match the best human researchers by about 2028, then compress years of progress into one and reach superhuman expertise in every field soon after.
Watch on YouTube ↗Charlie's summary
- ▶ 0:59
AI is already speeding up AI progress at the frontier labs
Dwarkesh cites METR, which spent three months inside Anthropic and found its rate of progress on well-scoped tasks is 2x what it was a year ago, mostly because of AI use in R&D. OpenAI's heaviest 10% of researcher users went from about $1,000 to over $7,000 a day on coding agents. Claude now leads 26% of Anthropic's AI R&D work, up from under 1% in February.
- ▶ 2:03
Math is the preview: AI research is mostly unglamorous, verifiable grinding, which AIs are disproportionately good at
Dwarkesh says math went from failing four-digit addition in 2023 to solving a millennium prize problem in 88 hours. Math cares about theory, but AI labs care only about what empirically works, so grinding through conjectures matters far more in ML. He thinks the transformer was a combination of known ideas plus many small experiments, so Claude 7 could have invented it.
- ▶ 8:12
Automated researchers as good as the best humans arrive around 2028, perhaps 2027
Dwarkesh's median for AI matching the best human AI researchers is 2028, assuming no slowdown. He gives at least a 25% chance it slips past 2032. Models generalize beyond what they were trained on: the millennium problem solver was never trained on millennium problems. He also notes that METR's doubling, taken naively, implies 2027.
- ▶ 11:15
Faster, smarter and more numerous automated researchers speed up AI progress a lot, despite diminishing returns
Dwarkesh says AIs work at about 10x human speed around the clock, roughly 50x the work per calendar week. About 80% of a lab's algorithmic progress comes from its 50 top researchers, and every automated researcher would match them. AI research, like formalizing a theorem, is modular and checkable, so swarms parallelize well. Compute for experiments still caps the gain.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.