David Silver's new bet: a superlearner that learns without human data
Training Data (Sequoia) · rss · 2026-10-13
Sequoia's Training Data podcast hosts Ineffable Intelligence co-founders David Silver (AlphaGo/AlphaZero) and Lasse Espeholt:
- The plan: build a "superlearner" that starts from nothing and discovers science, math and language via direct experience, not pre-training. Silver says AlphaGo Zero surpassing AlphaGo in under two days without human data proved the route; human-language reasoning is a ceiling, not a foundation
- Method: RL stack and "crucibles" — simple rewards, no objective breadcrumbs, emergent complexity; one agent is boring, two are infinite
- Alignment: chain-of-thought monitoring is already hitting its limits; superintelligence coexistence science should be worked out in simulation first
- Compute: Vera Rubin GPUs on GCP, a 20-person team with extreme compute-per-person, building on IMPALA and SEED RL
- Forecast: strong evidence within a year, superintelligence in three
More from AGI Musings
- Anthropic's Jack Clark moves up RSI timeline: full recursive self-improvement likelier in 2027 than 2028 — haider1 · 2026-10-13
- OpenAI Safety Researcher: No Fundamental Alignment Progress Since RLHF Nine Years Ago — Puzzleheaded-King584 · 2026-10-13
- Game dev pushes back on anti-AI argument: nobody actually uses AI for ideas — AIandDesign · 2026-10-13
- Dan Faggella: humanity may need radical augmentation to keep any control over AI future — danfaggella · 2026-10-13
- David Sacks Dismisses Superintelligence Fears as 'Rotating Hoaxes' in All-In Rant — DavidSacks · 2026-10-13
- Reddit debates the least replaceable job: dog groomer beats AI's finest — ataraxic89 · 2026-10-13