Who's Who of AI

David Ha

630 trust @hardmaru · 419,054 followers
What they're sharing

Articles & links

Sakana Fugu Technical Report https://t.co/6e6WuA8FVB Release Notes: https://t.co/7xWGpOicFN https://t.co/g2yaZvex35

Sakana AI sakana.ai
View on Bluesky · ♥ 0 ↻ 0 ↩ 0 · 5 from the directory shared this · 7d ago

RT @SakanaAILabs: Sakana Fugu Technical Report https://t.co/BRGepSmyI5 🐡 https://t.co/7d3webOkmH

Sakana Fugu Technical Report arxiv.org
AI Weekly's analysis →
  • Fugu-Ultra reportedly scores 73.7% on SWE-Bench Pro and 82.1% on Terminal Bench 2.1, beating Claude-Opus-4.8, Gemini-3.1-Pro, and GPT-5.5 baselines.
  • Fugu is itself a language model that learns to orchestrate Gemini-3.1-Pro, Claude-Opus-4.8, and GPT-5.5 as expert workers per query.
  • Training combines supervised fine-tuning with sep-CMA-ES evolutionary optimization for Fugu, and GRPO reinforcement learning for Fugu-Ultra.
Read full analysis →
View on Bluesky · ♥ 0 ↻ 0 ↩ 0 · 3d ago