ARC's Greg Kamradt: AGI gains must come from self-built tools, not one-shot ability
GregKamradt · x · 2026-09-09
ARC Prize founder Greg Kamradt argues the 'general' in AGI means doing things a system wasn't trained for. If weights are frozen, capability gains must come from external memory and tools the model builds for itself.
He cares less about one-shot performance and more about what a model can accomplish through building its own tools given enough time and inference. He adds this is a short-term win for the 'harness crowd' and creates friction for benchmarkers.
Related event: Debating AGI: Untrained Capability vs Ability to Learn Anything(5 posts)→
More from AGI Musings
- Should we hand the lightcone to a never-tested concept? CEV under fire — xuenay · 2026-09-09
- Kaj Sotala argues alignment should favor wise-advisor designs over untested CEV — xuenay · 2026-09-09
- OpenAI claims 10,000 agents solved Navier–Stokes Millennium Prize problem in 88 hours — scaling01 · 2026-09-09
- Sequoia mines 15 years of X hype data, mapping top topics to each year's biggest startup — Scobleizer · 2026-09-09
- Evitable founder David Krueger on what a superintelligence takeover moment would look like — DavidSKrueger · 2026-09-09
- A text's AI share is either 0% or 100%, argues economist — paulnovosad · 2026-09-09