Researchers from MIT say GPT-6 Astra improved the weights of Qwen3.5-0.8B directly and without any training run, using a technique called Weight-Space Programming. Alibaba’s model has 0.8 billion parameters, and the reported edits affected only 0.016% of its weights. Math performance reportedly increased by as much as sixfold and instruction-following performance tripled, although these figures represent peak rather than uniform gains. The experiment could enable rapid corrections to models deployed on resource-constrained devices, but independent teams must reproduce it and its applicability to larger models remains unproven.
Source: Read the original article
Disclaimer: this content is for information purposes only and is not financial advice. Cryptocurrencies are highly volatile: you may lose all of your capital. Always do your own research. Legal notice
