Researcher argues Ant now leads OpenAI in pretraining as SimpleQA defies saturation
teortaxesTex · x · 2026-11-15
- Fable 5.2 shares its pretraining base with Fable 5.1 yet clearly diverges from peers, leading teortaxesTex to speculate Ant built a new RL routine that better elicits pretrained knowledge.
- He argues the field already knew Fable/Mythos were strong base models, and that OpenAI has fallen behind Ant on large-base pretraining — explaining its scramble to poach Shazeer.
- The real surprise is SimpleQA, a two-year-old benchmark, only nearing saturation by Fable's third iteration, with stagnation in between (cf. Gemini 3.1 Pro's high score).
- A reply pushes further: no fixed "hard" benchmark matters anymore — expect arbitrary jumps in any domain, in any order.
Related event: Fable 5.2 Tops SimpleQA as Twin Model Mythos Cuts Prices 30%(3 posts)→
More from AGI Musings
- Norn founder at Foresight: AI-driven longevity drug breakthroughs need physiological data, not just 'more data' — juanbenet · 2026-11-15
- Dev laments generative AI debate has collapsed into nuance-free black-and-white thinking — burny_tech · 2026-11-15
- Pro-AI takes now trigger harassment waves from anti-AI zealots, dev community pushes back — burny_tech · 2026-11-15
- What Happens When AI Beats Most People at Most Work? — IagoInTheLight · 2026-11-15
- Fed survey: AI-adopting firms' revenue growth outlook net gap hits 48.2 pts, 2x+ non-adopters — AccBalanced · 2026-11-15
- No benchmark is hard anymore: Fable 5.2 tops SimpleQA, arbitrary jumps ahead — teortaxesTex · 2026-11-15