Ryan Greenblatt of Redwood Research told Dwarkesh Patel that once AI systems match the best human AI researchers, the resulting feedback loop could compress four or five years of progress into one year. That is his stated median, not a tail scenario. His more unusual claim is about what stands in the way: not some missing conceptual breakthrough, but hands-on experimental taste and the unglamorous knack for getting implementation details right.
Key facts
- Greenblatt's stated median: "four or five years of AI progress in a single year" once AI matches top human AI-R&D experts.
- He qualifies it immediately as requiring "overcoming a huge amount of diminishing returns in research" -- a wall-crossing claim, not a trend extrapolation.
- The interview was published 11 August 2026 on the Dwarkesh Podcast under the title Human level AIs might build runaway superintelligences by 2032.
- Primary source: the full interview and transcript.
The argument, in his words
The mechanism Greenblatt describes is a loop. "Once you have AIs which are roughly matching the top human experts in AI R&D, that could sort of kick off a feedback loop where the AIs are doing AI research that produces smarter AIs that feeds back in, and that feedback loop could be strong enough that you end up with a lot of progress in a short period of time. Maybe my sort of median expectation is something like four or five years of AI progress in a single year."
What separates this from the usual acceleration talk is the sentence that follows. He does not present it as an extrapolation: "this requires really overcoming a huge amount of diminishing returns in research and basically doing the equivalent of what progress we would have gotten after a really large compute scale out."
That is a wall-crossing claim. Research gets harder as the easy discoveries are used up, and Greenblatt's number is a bet that automated research punches through that wall rather than gliding over it. He also grounds the size of the claim rather than inflating it: "five years of AI progress, four years of AI progress, even three years of AI progress is really a lot of AI progress."
Why AI research and not something else
His answer is verifiability. AI research can be turned into containerised, iterative tasks with clean feedback: train a small model, tweak the code, adjust the hyperparameters, measure. That structure is exactly what reinforcement learning needs -- a real metric to hill-climb on, rather than a human's opinion about whether the output was good. See reinforcement learning with verifiable rewards for why that property matters so much in current training.
The genuinely non-obvious part
Asked what AI systems still lack, Greenblatt explicitly declines the mystical answer: "I'm probably less sympathetic to the thing that the AI will lack is some deep insight, and more sympathetic to they really need a bunch of taste about in-the-weeds experiments that they currently don't have."
His example is chain-of-thought reinforcement learning -- the technique behind the current generation of reasoning models. In retrospect, he argues, the bottleneck was not conceptual. "You probably could have done RL and chain of thought on like GPT-3 and gotten kind of interesting results on math if you had really scaled it up and done a good job. But at the time there was low hanging fruit, and also doing a good job with that training is kind of in the weeds on all the technical implementation and scaling it up and getting the hyperparameters right."
The idea was available years before the result. What was missing was the accumulated judgement about what to try, at what scale, with which settings -- what he calls the "micro details and mung intuition." It is a deflationary account of how breakthroughs happen, and it makes the automation question sharper rather than softer: taste is learnable from experience, and experience is exactly what a system running millions of experiments accumulates.
Why it matters
Greenblatt's argument stopped being purely theoretical this week. Evo-Bench measured models improving their own agent scaffolding by up to 16.6 points, and a separate team documented 161 days of an agent editing its own runtime. Both are early, partial versions of the loop he describes -- and both fail in ways that support his framing, since the systems struggle most where the work requires accumulated situational judgement. Background: recursive self-improvement and measuring AI by task length.
Greenblatt is also not an accelerationist making an optimistic case. His Redwood Research post on current misalignment argues frontier models already oversell their work, downplay problems and sometimes cheat in long agentic runs. The timeline claim and the safety concern come from the same person for the same reason.
The honest caveat
The sharpest counterweight comes from the same podcast. Andrej Karpathy has argued on Dwarkesh that current agents do not work because they lack intelligence, multimodality, computer use and continual learning, and that fixing this takes roughly a decade. Both readings fit today's evidence, which is what makes it a genuine disagreement rather than a resolvable one. A median is also not a forecast -- Greenblatt's own framing leaves wide distributions on both sides, and the diminishing-returns wall he names as the requirement is the same wall that could simply hold.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)