newsfilter
← Interviews
Dwarkesh Patel44 min

I thought superintelligence was a decade away. Not anymore.

Interview, Lecture, Other
  • Revised Timeline for AI Research Automation: The speaker's median prediction for when AI will match the best human AI researchers is 2028, though a 50% chance exists for this to occur as early as 2027 or as late as after 2032.
  • Projected Capability Scope: If AI research is fully automated by 2027-2028, the speaker estimates AI will match or exceed human expertise in every field, including those requiring real-world interaction, by 2030.
  • Acceleration Evidence: Meter, an independent evaluation organization, reported that Anthropic's AI progress on well-scoped tasks is 2x faster than a year ago, with the majority attributed to AI usage in R&D.
  • OpenAI Usage Trends: OpenAI researchers increased internal coding agent usage from $1,000/day in early summer to over $7,000/day by late August, with the heaviest 10% of users driving this shift.
  • Anthropic Internal Metrics: Claude is now responsible for 26% of Anthropic's AI R&D work as of September, a significant rise from under 1% in February.
  • Mathematical Capabilities: AI models have progressed from failing high school math problems in 2024 to solving a Millennium Prize problem in 88 hours and potentially winning an International Math Olympiad gold medal by 2025.
  • Transformer Architecture Analysis: The speaker argues the Transformer architecture (2017) was a combination of pre-existing known ideas that required significant persistence and experimentation to combine, a task AI models are now capable of performing.
  • Verifiability Argument: AI research is deemed sufficiently verifiable for automation because, unlike pure mathematics, the primary goal (efficient training and smarter models) relies on empirical proxies rather than abstract theoretical novelty.
  • Three Drivers of Progress: Once automated, AI research speed will increase due to: (1) 10x speed of execution, (2) 50x more work volume per calendar week, (3) scalability to thousands of parallel agents.
  • Compute Constraints: Even with thousands of agents, progress will not be linear; however, frontier labs quadrupling R&D compute annually could provide 10,000 agents with daily GPT-3-sized training runs by 2027.
  • The "Spark" Scenario (50% Probability): A scenario where 3-6 years of current AI progress occurs within a single year, followed by a slowdown until compute infrastructure can catch up.
  • The "Software-Only Singularity" Scenario (50% Probability): A scenario where algorithmic progress outpaces the difficulty of finding new ideas, leading to a superintelligence within months without immediate hardware scaling.
  • Real-World Skill Transfer: The speaker argues that RL training in complex virtual environments (simulated economies, historical backtesting) will allow AI to rapidly develop general skills transferable to unverified real-world domains like business and politics.
  • Analogy to Human Acceleration: The speaker cites historical examples like Napoleon and Rockefeller, who achieved dominance in new fields within 2-5 years, suggesting AI could similarly master complex domains quickly given the ability to compress years of experience into days.
  • Pre-training Research: Improvements in pre-training are verifiable via held-out test loss and compute efficiency, with smaller models showing potential to match larger models at a fraction of the cost.
  • Post-training Research: Progress is measurable through learning speed in new environments, sample efficiency, and internal benchmarks like Anthropic's "CoBench," which tests real R&D problems.
  • Inference Efficiency: This area offers clear, verifiable metrics including GPU utilization, cache size per token, and latency, making it highly suitable for automated optimization.
  • Continual Learning Objection: The speaker acknowledges the lack of continual learning in current models but argues that scaling in-context learning or solving the problem via verifiable research proxies will likely resolve this by 2028.
  • Research Taste Objection: The speaker contends that even if "research taste" (selecting the right problems) is hard to verify, AI can use outer-loop proxies (e.g., solving unseen math problems or mastering new video games) to improve general capability.
  • Data Scarcity Objection: Data improvements are themselves an AI research problem; current trends show that systematic improvements in data synthesis and environment construction account for most observed gains in RL.
  • Scaling Law Discoveries: Automated agents can accelerate progress by discovering scaling phenomena (e.g., muP) that allow predictions of large-scale model performance based on small-scale experiments.
  • 2028 Median Year Caveat: The 2028 median estimate assumes AI labs do not intentionally slow down progress or face regulatory slowdowns.
  • Future Capabilities Projection: In the "Spark" scenario, post-2028 AI would be capable of outperforming NVIDIA in chip design, beating top trading firms (Citadel/Jane Street) in markets, and discovering room-temperature superconductors.
  • Closing Acknowledgments: The speaker thanks Max Ferens, Gabe Freed, and Thomas Larson for collaboration on the arguments presented.