Anthropic Solved Interpretability Again? (Walkthrough)

[@drhxa]

That was an excellent walkthrough, thank you. I've learned a lot. Would love to see more walkthroughs of the prior/related work

[@TheInsideView]

Thanks! My walkthrough of the previous Anthropic paper (prior work): kzbin.info/www/bejne/fnLblWt6pL-UjZY For other interpretability papers I'd recommend checking out Neel Nanda's series of walkthroughs (he's actually leading a mechanistic interpretability team at DeepMind): kzbin.info/aero/PL7m7hLIqA0hpsJYYhlt1WbHHgdfRLM2eY&si=tLqxLua5XZEdbyCy

[@christopherwoodall3464]

Great overview. Really enjoyed the fact that you showed previous work that was built upon.

[@TheInsideView]

Thanks! To be honest I only briefly mentioned their previous work and don't think I actually went through previous work in the literature (was just doing a walkthrough of their blogpost, still doing daily uploads), but I'll definitely consider this preference to discuss previous work for future videos

[@Alice_Fumo]

This seems actually useful and has real-world applications. It seems this allows for actually adjusting the personality of the model, so one could make it more adverse to writing code with bugs, more flirty, more honest or whatever. The big AI labs could adjust small details without needing to retrain the AI. Also, I guess this could be done with open source models to figure out their "deny response" features and set them to very low values. It can be done with retraining, but that also just changes the model. Not needing such brute-force-y methods is neat.

[@TheInsideView]

Yeah exactly, that enables to steer them in the way that you'd prefer. If you haven't tried it yet I'd recommend checking out Golden Bridge Claude (which I talk about in the video) available on claude.ai for a limited time, which basically gives a concrete example of what having a custom steered LLM would be like.

[@Alice_Fumo]

@@TheInsideView I asked it to go one prompt without mentioning the bridge and tell me a bedtime story and it got extremely internally conflicted, retrying several times and wondering why it had such difficulty with this. It's extremely interesting to witness. Thanks for notifying me that they were hosting that model, I didn't know.

[@ThomasMeliWellness]

Crystal clear. Thank you for sharing this. Subscribed!

[@TheInsideView]

Thanks! Tomorrow's video will be another walkthrough so hopefully worth the sub