From the recent blog posted on weight-space programming: https://rikiyat.github.io/blog/weight-space-programming/
I am greatly interested in if we could specifically train (likely using reinforcement learning) LLMs to edit weights of other LLMs or itself. It sounds like OpenAI may have already tried this with Astra, given its capability mentioned in the blog.
However, I imagine Astra equipped with mathematical tools found by interpretability AI researchers through the years could greatly improve Astra’s performance. Or perhaps it could even invent its own set of tools, given its proficiency in math.
The theory in the past was this simply would not scale given the trillions of parameters LLMs have. However, I don’t buy this argument as only a small percentage of the neurons contribute to most given tasks. Thus, mathematical methods can be used to narrow down the circuits involved in a given objective to a small fraction, and from there, other methods can be used.
Perhaps this will be added as a step after post-training, or perhaps it could even lead to the AGI RSI loop all the labs are reaching for.