rikiyat.github.ioSparks of Weight-Space Programming: LLMs can natively read & write neural network weightsWe find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, well past what we’d expect from humans or conventional algorithms.
Replying to @lain@fediffusion.art
@lain just finished reading this and it's kinda nuts. Emergent capabilities my ass, how in the hell would it learn to do that? I wonder if they extensively trained the thing on the, what, binaries of the other weights? jamming it straight through the text tokenizer?