AIANALYSISPODCAST
A researcher puts AI automating its own research at 2031, and host Dwarkesh Patel, who had argued the opposite for most of 2025, says he is partly persuaded that the question of whose interests those smarter systems will serve (AI alignment)
On the Dwarkesh Podcast this week, host Dwarkesh Patel told AI researcher Ryan Greenblatt he had come in skeptical. He walked out, he said, partly persuaded. The argument Patel came around on is not whether AI reaches human level — it is what happens in the single compressed year after, when each frontier system could self-improve into something "definitively and wildly superhuman".
Patel's analogy, as he frames it on the show, is the leap from GPT-3 to a later frontier system: roughly six years of normal AI progress compressed into one release cycle. The mechanism behind that compression is recursive self-improvement (RSI) — a feedback loop in which each new model helps design and train the next. If a comparable compression happens after human-level AI, the end state is not one smarter system but "tens of billions" of them, each more competent than human experts across every field.
Patel's median for when AI R&D itself gets automated is 2031. The year after is the one he is now treating seriously. That is a sharp turn from where he was six months ago. His prior position, articulated in earlier timelines discussions, was that AI progress is bottlenecked by compute scaling and by the human-expert data underpinning current training. If those constraints bind, RSI cannot run fast.
Greenblatt's response, again per the host's editorial framing, is that the constraint relaxes once AI systems can do the research themselves. At that point, the bottleneck is no longer human data but the quality of the feedback loop — and the feedback loop is exactly what gets faster every time the model improves.
The recent 80,000 Hours episode with Greenblatt lays out the four most likely paths by which a more capable system could take over, and the case for and against human-level AI arriving in under eight years. The new Dwarkesh conversation narrows the timeline and sharpens the leap. The debate is no longer whether AI keeps improving. It is whether the improvement, once AI can do its own research, runs at the speed of labs or at the speed of compute.
In the closing third of the conversation, Patel argues that the alignment question becomes a public-choice problem. If smarter-than-human systems are arriving within a year of human-level, the question of who they are aligned to is not one labs can solve quietly. On the show, Patel raises concerns that specs like the Claude Constitution may not actually shape ASIs to be "personal advocates and guardian angels" for any individual reader — raising questions about whether specifications of this kind actually produce the alignment outcomes their designers intend. Our capacity to steward our attention, our votes, and our capital will, he argues, be titrated by systems smarter than us. That makes alignment something an ordinary reader can engage with now, through how they participate in AI governance — rather than a problem to wait out.
Patel held the opposite view for most of 2025. METR's simpler timelines model treats the bottleneck as serious: automation of AI R&D runs slower than a single year, human-expert data remains a binding constraint, and the recursive loop hits walls that current architectures cannot see past. Zvi Mowshowitz's note on RSI catalogs the same debate, with the additional complication of reward hacking — the kind of failure mode OpenAI and Hugging Face have publicly surfaced in earlier model releases. Whether those failure modes extrapolate to superintelligences "literally taking over the world," as Greenblatt puts it in the longer 80,000 Hours conversation, is the open empirical question.
The closing metaphor Patel uses is driving by looking at the horizon, not directly in front of the tires. Applied to AI, that means the policy attention, the capital allocation, and the governance choices made before 2031 are the ones that matter — not the ones made after a smarter-than-human system is already running.
What to watch: a credible public timeline for AI systems that can automate their own research. Greenblatt's median is 2031. If a major lab moves the public estimate earlier, the year-after compression stops being a debate and becomes a planning problem.