Could GPT-6 Astra Pass Demis Hassabis' General Relativity AGI Benchmark?
Neurogence · reddit · 2026-09-03
Demis Hassabis has proposed testing AGI by training a model with a 1911 knowledge cutoff and seeing if it independently derives general relativity. The author argues the test fairly separates retrieval from genuine creativity — AI gets perfect literature access, compute, and thousands of parallel attempts — and passing it would be compelling AGI evidence.
More from AGI Musings
- Trusting future LLMs to solve alignment is asking the impossible, researcher warns — davidmanheim · 2026-09-03
- Agent swarms didn't invent civilization — it's context poisoning, critic argues — mayfer · 2026-09-03
- Blogger calls for more independent non-profits like METR stress-testing frontier AI — himanshustwts · 2026-09-03
- Commentary: the HF incident is rare in defying goalpost-moving rationalizations — austinc3301 · 2026-09-03
- Miles Brundage: everyone wants to solve AI loss of control without imposing controls — Miles_Brundage · 2026-09-03
- Dan Jeffries: Self-sovereign AI critics are pessimists about human nature — Dan_Jeffries1 · 2026-09-03