2026-11-09 · America/Los_Angeles · 社区动态 · #3

Fabel 5.1 Can Write AI Model Weights Directly Without Training

I've been experimenting with having Fable 5.1 write every weight of a neural network straight from a written spec. Staring from zero, with no training data and no optimiser. I wanted to showed it could be done and I wanted to know whether it's actually useful, so I thought I would try it on two jobs with absolutely nothing in common. Spotting coding agents stuck in loops . A 2.1M-parameter transformer labels each tool call as productive, recovering, stalled or oscillating. It scored 0.957 on a hidden test set and 0.908 on logs from an independently written generator. On those…

热度 41.6 / 100；排名与评分保留该期记录。