oh, here's an interesting #LLM -thing to play with:https://www.neuronpedia.org/llama3.3-70b-it/nlabasically there's an a...

oh, here's an interesting #LLM -thing to play with:https://www.neuronpedia.org/llama3.3-70b-it/nlabasically there's an autoencoder that reads the LLMs internal activations at a certain layer and tries to decode them. so you can kinda get the "thoughts" behind a certain token. they are a bit odd and disjointed, and it's a bit hard to say which parts are interpolations of the interpreter sometimes, but i think you get a feeling for it at some point.