Dwarkesh Patel: AI That Does AI Research Is Months to Years Away

Podcast host Dwarkesh Patel told viewers on Nov. 13 that AI systems able to match the best human AI researchers could arrive "in a couple of years, and possibly only a couple of months." His video essay argues that recursive self-improvement — each generation of AI building a more capable successor — has already begun inside leading labs.

The claim rests on measured acceleration at Anthropic, not only on argument. METR, an independent evaluation group, estimated on Nov. 5 that Anthropic is improving its models on well-specified tasks at roughly twice the rate it did between March 2024 and October 2025. METR flagged wide error bars: the true figure could be closer to 1.3 or 3 times. The group attributed a majority of that acceleration to Anthropic's use of AI for its own R&D.

Vendor data points run in the same direction. Anthropic reported that Claude "leads" 26% of its AI R&D work as of August. OpenAI said in September that the 90th-percentile user in its research organization consumes more than $7,000 of tokens per day — a budget line R&D managers can read as a proxy for how heavily internal researchers now lean on models.

Who is making the forecast?

Patel, 26, hosts the Dwarkesh Podcast, a long-form interview show whose guests have included Anthropic CEO Dario Amodei, OpenAI co-founder Ilya Sutskever, Microsoft CEO Satya Nadella, Mark Zuckerberg and, in April, NVIDIA CEO Jensen Huang. TIME named him to its 100 most influential people in AI in 2024; The Economist called him Silicon Valley's favorite podcaster in 2025. He co-wrote "The Scaling Era: An Oral History of AI, 2019-2025" with Gavin Leech.

His timeline has moved sharply. Earlier this year, he placed superintelligence "at least the better part of a decade" away. Three months of debates on and off his podcast, plus rapid AI gains in mathematics and evidence that AI is already accelerating AI research, changed his position.

What are the concrete numbers in his scenario?

Even in his slower scenario, Patel expects three to six years of AI progress compressed into one. Two variables govern the outcome, he said: how reliably machines can test their ideas, and how much computing power labs can supply to keep experiments running.

His median year for AI matching the best human AI researchers is 2028 — conditional on labs going "full steam ahead." He assigns "at least a 25% chance that it doesn't happen until after 2032."

How could a data-center system outperform experienced staff?

Patel addressed the standard objection that expertise requires years on the job. He cited Napoleon and John D. Rockefeller, who "needed experience, yes, but only a couple years' worth, not decades." Deployed systems, he argued, can learn fast because they can absorb context at machine speed.

"If you let [a future AI model] shadow that CEO, on its first morning, it could read every email and Slack message and document that the company's ever produced," he said. "By the end of its first day, copies of it have interviewed every employee and major client in parallel."

He conceded one limit: feedback that arrives over months, such as the next quarterly earnings, still takes months. The systems that follow the compressed year of progress should reach "top human performance in all domains within months of being developed," which he places in 2028 or 2029 in his median scenario.

What should R&D managers weigh?

The forecast comes from an interviewer, not a lab, and the supporting METR figures carry uncertainty bands a factor wide. But Patel's argument ties directly to portfolio decisions: if token-intensive internal AI use is already doubling effective research velocity at one frontier lab, benchmarking an organization's own AI-assisted R&D throughput becomes a near-term budget question rather than a speculative one. Patel's next testable marker is whether model-driven R&D acceleration holds through 2026.