Epoch After Hours Forecasting AI progress until 2040
68 snips
Sep 5, 2025 Yafah Edelman, Head of Data and Analysis at Epoch AI, dives into the transformative future of AI over the next 15 years. She discusses the potential for AI to solve the Riemann Hypothesis and predicts significant advancements in cognitive automation by 2030. The conversation touches on the economic ramifications of rapid automation, the evolving role of software engineers, and the challenges of scaling robotics. Yafah also speculates on a future dominated by self-replicating machines and exponential growth in AI capabilities.
AI Snips
Chapters
Transcript
Episode notes
Compute Scaling Will Slow But Stay Substantial
- Current training-compute scaling likely slows from ~5x/yr to ~2.5–3x/yr due to longer R&D and diminishing training-duration gains.
- Firms could still push faster by using entire clusters, so compute growth remains substantial in the near term.
Larger Runs Yield Competent Agents And Scientific Gains
- By end of decade a ~10^3× larger training run could yield highly competent agents and much fewer simple reasoning failures.
- Such models may begin producing novel scientific and mathematical discoveries and replace much coding work.
Math And Rigorous Fields Are Low-Hanging Fruit
- AI-assisted math and rigorous fields are especially promising for early major breakthroughs.
- Jaime and Yafah consider an AI-assisted solution to a famous problem plausible within five years.