Sakana AI's Fugu-Ultra agent autonomously optimizes ML training and text reconstruction

Sakana AI has developed Fugu-Ultra, an AI agent that autonomously improves machine learning training recipes. In one experiment, Fugu-Ultra iteratively edited training code and ran 123 experiments over 14 hours on a single H100 GPU, achieving a better mean bits-per-byte (BPB) score than three frontier models. In a separate test, Fugu-Ultra demonstrated superior performance in reconstructing the reading order of classical Japanese text, achieving an NED score of 0.80 compared to other models' scores below 0.24. The agent also successfully wrote a Rubik's Cube solver from scratch in Python. AI

IMPACT Demonstrates potential for AI agents to accelerate ML research and solve complex tasks, potentially outperforming individual frontier models.

RANK_REASON The item describes an AI agent's performance on specific ML research tasks and benchmarks, rather than a general product release or frontier model announcement. [lever_c_demoted from research: ic=1 ai=1.0]

Read on Hacker News — AI stories ≥50 points →

• AutoResearch
• bits-per-byte (BPB)
• Fugu-Ultra
• H100 GPU
• Karpathy et al.
• Keio Institute of Oriental Classics
• Model A
• Model B
• Sakana AI

• model release
• product
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →