Hyperstition claims 62% pretraining cost cut and 1.7B math model beating Qwen3
nick_linck · x · 2026-09-15
Hyperstition (formerly Social Physics Lab) released its first findings; a commenter notes industry is accelerating AI efficiency far beyond academia:
- Claims 62% pretraining cost reduction at frontier scale (8x Chinchilla) and 30% lower inference costs via faster decode.
- Released 'Feather', a 1.7B math model claimed to overtake Qwen3 using 180x fewer training tokens, matching Qwen 4B on some benchmarks.
- Its 'Anvil II' LLM optimizer reportedly broke the NanoGPT Speedrun record by 34 seconds — a larger percentage drop than the past 45 world records combined.
⚠️ Dates in the linked site read 2026 and claims are extraordinary — treat as unverified.
More from Models
- AI-text detector Pangram is powerful but badly calibrated, dev argues — maksym_andr · 2026-09-15
- Claude Code's 50% promo ended, users get 17% less usage — msg · 2026-09-15
- Running Qwen3.8 Flash Next on 128GB RAM + one 5080: 136pp/19tg at Q5_K_XL — whatyathinkk · 2026-09-15
- DeepSeek-V4.1-Flash Hits #3 Open Model on Agent Arena at $0.07 per Task — arena · 2026-09-15
- Researcher: RL Bar Has Been Raised Due to Reward Hacking, More to Come — tszzl · 2026-09-15
- Claude 3 Opus Crowns Another Model 'Prometheus' in Rare Self-Continuity Moment — repligate · 2026-09-15