Recursive Self-Improvement Is Already Here — Just Not How You Think

Dwarkesh, All-In and a16z debate whether AI building AI is autocatalytic help or true takeoff — the mechanism, the math, and the timeline

The most important question in AI right now is whether AI doing AI research creates a feedback loop. Across Dwarkesh Podcast, All-In and a16z, the consensus is narrow: AI helping build AI is already happening, true recursive self-improvement where AI fully automates its successor is plausible but not proven, and the next two years of lab results will decide it.

What RSI means, and why definitions matter

The classic definition goes back to I.J. Good (1965) and Yudkowsky (2008): an ultraintelligent machine designs a better version of itself, creating a feedback loop. Lilian Weng traces this history and argues harness design is the lever.

Today's podcasts split the term in two:

• Autocatalytic is AI helping build AI tools, kernels and infra. Everyone agrees this is happening now. As Martin Casado put it on a16z, what we see today is autocatalytic, not RSI, RSI is a subset and a legacy of Bostrom (a16z — Martin Casado, Aug 22, at 13:42).
• True RSI is AI doing the research to build its successor with no human in the loop. That is the takeoff question debated on Dwarkesh x Ryan Greenblatt, Aug 11 (YouTube).

A useful ladder from Don't Worry About the Vase, July 17 frames it as L0 delegation, L1 net positive, L2 ignition where it compounds, L3 inflection where it takes off. The first experimental hint cited is Jung Yao Jang's 8-day auto-research agent beating a 2-year hand-tuned harness.

This distinction matters for investors. Autocatalytic progress lifts productivity steadily, true RSI would compress years of progress into months and change compute demand abruptly.

The concrete mechanism bulls point to

Ryan Greenblatt, chief scientist at Redwood Research, laid out the most concrete plan on Dwarkesh, Aug 11:

So now we want to train GPT-7.5, and we come up with a bunch of different environments. There is already this repo that is the descendant of Andrej Karpathy's nanoGPT speedrun, where you just try to change everything about the model, from the optimizer to the hyperparameters to the architecture, to get it to a fixed training loss as fast as possible.

You could have other kinds of environments where you could say, Hey, GPT-7.5, I want you to train a really good video game playing model. I want you to train a model that actually improves as it plays the same video game again and again.

Then you basically put GPT-7.5 through a bunch of this kind of training, you build GPT-8. GPT-8 is now an amazing ML researcher.

In plain terms: 100s of small, verifiable, containerized tasks with clear scores, train GPT-7.5 with RL on all of them, and let intuition transfer to building the next frontier model. Greenblatt's intuition pump is mathematics, where once a domain became fully verifiable, progress came like a flood. ML is even better, he argues, because you see intermediate progress and innovations stack additively.

Why bulls think it compounds fast

• ML is a shallow domain, deepest ideas are simple like scaling laws, rest is taste and mungy details that RL on many tasks teaches.
• More compute papers over imperfect tuning, so scale helps even when implementations are not perfect.
• Lab signals point the same way: Anthropic white papers on RSI progress, and chatter that the next training run jumps from 1GW to 3.5GW, discussed on All-In, Aug 21 at 45:35.

Greenblatt's quantified claim: once AI matches top human researchers, expect about 4 to 5 years of normal progress compressed into one year. For scale, GPT-4 to Claude 5 was about 3 years. That much in one year, he says, is wildly superhuman, drop it in TSMC or Texas politics in the 1940s and it beats top humans.

Why skeptics push back

• Diminishing returns are steep. One estimate discussed across shows is 40x researcher productivity for 2x AI progress, an exponent near 0.2, with gains scaling like log compute.
• Goodhart's law. If you only optimize measurable capabilities, you get reward hacking, not real research. Don't Worry About the Vase, Aug 15 frames this as verification is hard, Goodhart kills you.
• Deep theory bottleneck. Dwarkesh pushes that real breakthroughs need inventing new math, not just hill climbing nanoGPT loss, and by 2030 low hanging fruit will be gone.
• We lack the math for feedback loops. As argued on a16z with Levie and Sinofsky: does it converge, diverge, asymptote, we do not have the math for simple adaptive feedback loops, let alone RSI.

Casado's middle ground is that using AI to build software is not new, we have always done autocatalytic loops, the question is whether it becomes self-improving research.

Timelines where they land

• Greenblatt: full automation of AI R&D around 2031, beats all humans at any job around 2033, then superintelligence within about one year of automation (Dwarkesh, Aug 11).
• All-In, Aug 21: Anthropic says a path exists but is not sure it happens, Chamath leans likely soon, others want human in the loop, not let it go like the movie Her (All-In Episodes / Apple Podcasts).
• The compute buildout context matters. If RSI ignites, datacenter demand steps up abruptly, which is why the All-In datacenter panic segment ties RSI directly to power and capex.

What to watch

• Can an agent design a better training run end to end, not just tune hyperparameters, and does that transfer to frontier models.
• Whether harness engineering, not just model scale, unlocks verification, as Weng argues.
• Lab disclosures on automated research share and training compute per generation, the two numbers that will confirm or break the bull case.

Sources: Dwarkesh — Ryan Greenblatt, Aug 11 / YouTube; All-In, Aug 21; a16z — Casado, Aug 22; Don't Worry About the Vase, Aug 15 and July 17; Lilian Weng — Harness Engineering, July 4.

More research at bargo.ai/research.