Reports say frontier models can generate and edit neural-network weights
Latent Space reports that several groups reproduced the Weight-Space Programming paper, finding that frontier LLMs could generate or edit weights for small neural networks without gradient training or tools. Reputations included Hugging Face’s ML Intern reproduction and a laptop run with Qwen 3.8 27B. Other reported results included Fable generating a 100,000-parameter classifier and GPT-6 Astra editing Qwen3.5-0.8B weights to raise MATH-500 accuracy from 7.9% to 48.1%. Latent Space said some claims were presented too favorably, citing weak baselines, benchmark overfitting, looping-driven gains, and a narrower interpretation in which the models act as sandboxed interpretability agents rather than natively “reading” weights.
Multiple independent replications indicate a substantive development in agentic mechanistic interpretability, but criticism of baselines, evaluation methods and the apparent mechanism limits broader conclusions about native weight understanding and reasoning gains.
This source is a digest of several news items (“[AINews] Behind Latent Space: Kill My SaaS, Jev…”). Find this story in the original.
#Weight-Space Programming#Hugging Face#ML Intern#GPT-6 Astra#Qwen3.5-0.8B#Qwen 3.8 27B#MATH-500#Fable