Dwarkesh Patel44 min
I thought superintelligence was a decade away. Not anymore.
Interview, Lecture, Other
- Forecasts suggest a significant compression of progress timelines, with AI researchers potentially equalling the best human experts by 2028, a median estimate that carries a 25% chance of delay beyond 2032 if hard-to-automate components persist or government intervention occurs.
- Two post-automation scenarios are weighted at 50% probability each: a "Spark" scenario compressing three to six years of progress into a single year, and a "software-only singularity" emerging within months without additional compute increases.
- Rapid capability gains are projected within 12 months of automating research, potentially jumping from a GPT-3.5 equivalent to a GPT-6 Astra equivalent, with post-Spark models reaching top human performance in real-world domains within months of deployment.
- Compute availability is expected to scale to support 10,000 agents running daily GPT-3 sized training runs by the end of 2027, building on a historical global AI compute growth rate of roughly 3x annually over the last half-decade.
- Automated researchers are predicted to leverage synthesized data, virtual worlds, and parallel experimentation to solve verifiable problems, including discovering new scaling phenomena like muP, designing next-generation AI accelerators better than NVIDIA, and solving Millennium Prize problems.
- Operational efficiency improvements are anticipated through the ability to predict hyperparameter optimality from smaller models, minimizing compute usage for training smaller models, and shifting from long pre-training runs to rapid innovation cycles.
- Risk factors include the possibility of hitting the 2028 milestone later due to non-automatable aspects of research, as well as the "software-only singularity" timeline requiring hardware scaling which could take years to achieve post-Spark.
- Current adoption trends indicate surging internal usage, evidenced by OpenAI users increasing coding agent spend from $1,000 to over $7,000 daily in summer months and Claude managing 26% of Anthropic's R&D work compared to under 1% in February.
- Future models are expected to train on vast historical corpora including all newspapers until 1962, every scientific discovery and election, and unfathomable virtual worlds to achieve transfer learning and sample efficiency superior to biological evolution.
- Verification mechanisms for progress include held-out test set losses, inference efficiency metrics like GPU utilization, and the ability to predict large-scale improvements from small-scale ablations, making AI research progress highly testable and scalable.