Voyage AI founder Tengyu Ma: frontier models at Ultra settings are still fundamentally dumb
lateinteraction · x · 2026-09-20
Tengyu Ma (lateinteraction), founder of Voyage AI and Stanford professor, describes the daily agony of extracting merely OK work from frontier models even at "Ultra" settings — requiring levels of detailed context and elaborate feedback he has never given a human, despite feedback being his literal job. His verdict: as of at least September 2026, public frontier models, like those of 2020, "know a whole lot but are fundamentally quite dumb and unadaptive" — a sharp counterpoint to AI-replacing-work narratives.
Related event: Researcher Slams Frontier Models as Still Fragile After 8-Hour Sessions(4 posts)→
More from AGI Musings
- If early autocomplete-era LLMs could code, do we even need post-training? — menhguin · 2026-09-20
- Pharma exec admits he barely uses AI day to day despite the hype — TheMoonMidas · 2026-09-20
- Jerry Kaplan argues AI existential risk is sci-fi, not science — zetalyrae · 2026-09-20
- The Jev killer isn't another model — it's a 1M-token context window — cocktailpeanut · 2026-09-20
- Bill Gurley: My P-second-COVID is way higher than my p-doom on AI — kevinnbass · 2026-09-20
- AI art is theft: three years of ignored artist grievances fueled the anti-AI wave — wordgrammer · 2026-09-20