Forethought model: ~60% odds software intelligence explosion fits >3 years into <1
ben_j_todd · x · 2026-10-05
Responding to a discussion, Ben Todd cites Forethought's paper by Tom Davidson and Tom Houlden on how quick and big a software intelligence explosion would be once AI fully automates AI R&D:
- 60% probability of compressing >3 years of AI progress into <1 year via software improvements alone;
- 20% probability of compressing >10 years into <1 year;
- Continued compute growth during this period would add further progress.
Todd's own take: "2-5 years of progress in 1 year" is the more realistic estimate, more conservative than the paper's scenarios.
More from AGI Musings
- Professor's grim math: slowing AI costs 100,000 lives per week — DoctorJosh · 2026-10-05
- Countervailing Technologies: how distillation-style tools can diffuse AI power concentration — DavideCrapis · 2026-10-05
- NeuroAI Researcher: You Don't Need to Emulate the Whole Brain to Get Intelligence — aran_nayebi · 2026-10-05
- Consciousness debate sparks: 'neurons are just math' argument mocked — inductionheads · 2026-10-05
- Gary Marcus Mocks Lawmaker's Argument That LLM Randomness Rules Out Consciousness — GaryMarcus · 2026-10-05
- Don't bet against the airplane: researchers push back on AI skeptics — aran_nayebi · 2026-10-05