The avalanche that hit math is coming for AI development
Nov 13, 2026 · 44m
More Technology podcasts · More Science podcasts
Summary
Dwarkesh Patel argues that AI is only a couple of years away from automating AI research, a view updated by recent breakthroughs in mathematical problem-solving and increased AI usage in R&D. He predicts that by 2028, automated researchers will match top human experts, compressing years of progress into a single year through faster, smarter, and more numerous agents. The episode explores whether this leads to a rapid "software-only singularity" or a slower "Spark" scenario, while addressing objections regarding verifiable proxies and continual learning.
Topics discussed
Introduction: Updating views on superintelligence timeline
Evidence of AI-accelerated R&D at Anthropic and OpenAI
Rapid progress in AI mathematics and verifiable domains
Why solving well-scoped problems matters more for ML than math
Could AI have invented the Transformer architecture sooner?
Predicting 2028 for automated AI researchers matching humans
Implications of doubling AI progress rates on takeoff timing
How speed, intelligence, and parallelization accelerate progress
Software-only singularity vs. Spark scenario probabilities
Capabilities of superintelligence in verifiable domains by 2027
Challenges in training AI for real-world, non-verifiable tasks
Evidence for skill transfer from verifiable to general tasks
Rapid ramp-up of AI in complex roles like corporate leadership
Continued compute scaling after software progress plateaus
Appendix: Verifiability as the key to algorithmic progress
Automating pre-training research and efficiency gains
Verifiable proxies for post-training and inference efficiency
Addressing objections: Continual learning and paradigm shifts
Data synthesis as a major driver of AI progress
Experiment compute bottlenecks and efficient agent swarms
More from Dwarkesh Podcast
Listen ad-free on Castria