Report
Researchers report building transformers with AI-written weights—no backpropagation
DAIR.AI reports perfect held-out accuracy across 26 algorithmic runs, but language tasks proved less cooperative.
TLDR
In interpretable weight-space programming, an AI agent writes, tests and revises model weights directly in code, researchers report. DAIR.AI says the loop runs 30–50 rounds; with Claude Opus 5.5, the smallest successful algorithmic models had a median of 166 parameters. Tiny doesn’t mean universally better: on language tasks, models plateaued near 190,000 parameters, and paraphrase accuracy remained below its baseline, according to DAIR.AI.
Combined views
435
2 Sources, first seen ago

Comments