AI Research Agents by 2028: Dwarkesh Patel's Forecast

Dwarkesh Patel has published a forecast arguing that AI's trajectory in mathematics, from failing basic arithmetic to solving a Millennium Prize problem in 88 hours, is a preview of recursive self-improvement in AI research itself. He predicts systems will match the best human researchers by 2028 and reach superintelligence by 2030, with a 25% chance the timeline slips past 2032. The essay, published November 13, 2026, draws on internal lab metrics showing Anthropic's progress rate on well-scoped tasks has doubled in the past year, with the majority of acceleration attributed to AI-driven R&D.

The Math Trajectory as Template

Patel traces a three-year arc. In 2023, models could not reliably add four-digit numbers. By 2024, high school competition problems remained out of reach. In 2025, AI won gold at the International Mathematical Olympiad. Earlier in 2026, systems tackled open Erdős problems. A month before Patel's essay, one solved a Millennium Prize problem in 88 hours, though the specific problem remains unnamed in public reporting.

His core claim is that AI research will accelerate faster than pure mathematics. Pure math, Patel argues, demands the capacity to ask powerful questions, a capability Toby Ord notes current systems lack. AI research, by contrast, rewards grinding through empirical conjectures that improve training efficiency. Patel quotes Chris Olah's observation that the transformer architecture emerged from a natural combination of existing ideas taken to aggressive lengths, a process he believes pattern-matching systems can replicate.

Lab Metrics Signal Acceleration

Patel cites METR evaluators reporting that Anthropic's overall rate of progress on well-scoped tasks is now twice what it was a year ago, with the majority of this acceleration attributable to extensive AI use in R&D. He writes that Claude now leads more than 26% of Anthropic's AI R&D work, up from under 1% in February 2026.

At OpenAI, the heaviest 10% of researcher users increased their daily spend on internal coding agents from around $1,000 at the beginning of summer 2026 to over $7,000 by late August. Patel interprets these figures as evidence that labs are already deploying AI systems to accelerate their own capabilities research, creating a feedback loop.

Expected Timeline and Uncertainty Bounds

Expected to release on an unspecified 2026 timeline, Patel places his median estimate for AI matching the best human researchers at 2028. He forecasts superintelligence by 2030, though he notes it could arrive as early as next year under aggressive assumptions. He assigns a 25% probability that superintelligence does not materialize until after 2032.

Patel references Ajeya Cotra's observation that if AI is nearly doubling the rate of AI capabilities progress and the acceleration remains smooth, extreme superintelligence could arrive within a year. He presents this as an upper bound rather than a central forecast. The essay acknowledges uncertainty but frames the lab-sourced metrics as sufficient to justify aggressive timelines.

Context and Open Questions

The essay does not specify which Millennium Prize problem was solved, leaving a gap in verification. Separately, Anthropic CEO Dario Amodei recently described compute as the company's deepest business dilemma, a constraint that could slow the acceleration Patel forecasts. The tension between reported internal progress rates and public statements about resource bottlenecks remains unresolved.

Patel's forecast hinges on the claim that empirical AI research is more tractable for current architectures than pure mathematics. If pattern-matching systems can efficiently explore the space of architectural and training improvements, the recursive loop could compress timelines. If deeper conceptual breakthroughs are required, the math analogy may not hold.

What This Means

Patel's essay synthesizes lab-reported metrics into a concrete timeline, putting median odds on top-human AI researchers by 2028 and superintelligence by 2030. The argument rests on the premise that AI research is more amenable to recursive self-improvement than pure mathematics, and that current acceleration rates at Anthropic and OpenAI provide early evidence of that loop. The forecast carries high uncertainty, with a quarter probability of timelines extending past 2032, but frames the next two years as a critical window for observing whether lab-internal AI use continues to compound progress at the rates reported in late 2026.