Model-weights self-improvement demos show promise but lack evidence to beat fine-tuning

ccerrato147 · x · 2026-11-04

ccerrato147 offers a sober assessment of demos where a model reads and writes another model's weights: a bit of RSI, a bit of hype.

The '20x better than RL' claim doesn't hold on reruns: Seaquest 89–53, Asterix 55–53, one Breakout rerun 58–25 — twice the trained network, not twenty times
It removes blockages, doesn't teach new skills: on a theorem prover, token-cap hits fell from 42% to 4%, but slightly fewer theorems proved
The flattery test exposed limits: with weights alone, Astra failed 4/4; it only worked once it could watch the network's attention on its own prompts — and no one verified the fix by editing the blamed head
Without a real paper with costs on harder cases, nobody can say this beats plain gradient fine-tuning or scaling backpropagation

More from Models