NaiveAI's first model is built and trained largely by AI itself, cutting decode latency 72%
机器之心 · wechat · 2026-09-28
Startup NaiveAI has AI models handle most of its R&D, with humans setting goals, constraints, and key decisions. Its first model Naive-N0.5-Flash (309B total/15.5B active params, native 1M-token context, MIT license) targets coding and AI research.
It scores 63.2 on PaperBench, 73.7% on MLE-bench-30, and pushed the NanoGPT SpeedRun record from 77.5s to 73.8s. Companion cases: the NaiveRT inference runtime cut speculative decoding from 12.3ms to 3.4ms (-72.4%) over 151 experiments in six days; the AutoWM world model reached the top tier of WorldArena after 400 hours of experiments.
Technically, it replaces global attention with DeepSeek Sparse Attention in a SWA-DSA hybrid, using mixed sequence parallelism to train 1T tokens in 4 days on 512 GPUs. The release mirrors a broader shift: Anthropic says Claude now leads 26% of its internal AI R&D, and OpenAI has hit its automated research intern goal.
Related event: NaiveAI Open-Sources 309B MoE Model Built Largely by AI Itself(3 posts)→
More from AGI Musings
- Dev argues LLMs can't yet do out-of-distribution science like novel algebraic geometry — burny_tech · 2026-11-12
- Blogger: Some people would rather see AI fail than benefit from companies they dislike — VraserX · 2026-11-12
- Team proposes lending idle agent quotas to science, Folding@home style — my-aga · 2026-11-12
- Gary Marcus Meets Pope Leo XIV, Hails Him as Clearest AI Thinker Among World Leaders — sapinker · 2026-11-12
- Isomorphic Labs President on Using AI to Turn Drug Discovery into Precision Engineering — No Priors · 2026-11-12
- Roblox CEO Baszucki: AI-made games hit top 20, 80% of top 100 AI-touched — Stratechery · 2026-11-12