tracr

Here's the work that shows how rule-based logic can be compiled into transformer weights: * Thinking Like Transformers” by Weiss, Goldberg & Yahav (ICML 2021).: arxiv.org * Tracr: Compiled Transformers as a Laboratory for Interpretability arxiv.org (code at deepmind/tracr) So it's not really an "LLMs can modify weights" thing. Rather it's "LLMs can write programs that can write or modify weights". Indeed LLMs can write many programs, it's impressive.

@Edwi_Villanueva yes, for controlled algorithmic tasks. Tracr compiles human-written programs directly into transformer weights, without training. the mechanism is known by construction, so you can test whether an interpretability tool recovers it. google-deepmind/tracr

@jpschroeder but this actually was done by google some time ago google-deepmind/tracr

@VictorTaelin isnt it just more complex version of this google-deepmind/tracr