04/01/2024
The best talk about mechanistic interpretability (the method best to hopefully understand what our new brain is up to).
https://youtu.be/2Rdp9GvcYOE?si=AYiL2-2PQoXrfZVu
Great points that are beautifully clear and others that are just as foggy as the view of our own brain.
So just like us, as we grope around our brains looking for anything to, at least, discover the seat of consciousness; if we can’t ever quite grasp how the model is performing it’s deep connections, features and weights, we can then mark the calendar as the day A was dropped leaving behind just what it is, Intelligence.
Cause if we can’t understand it or us, we’ll, that kinda says something.
Yeah, that we have clearly reached a sort of limit to knowing how something works. Unifying physics, maybe. Life starting. Time. We might realize that just because we’ve got it going on, that it should be of no surprise that we, the universe, have more to our makeup than we can grok (is that how you use it? What the f**k is grok). And if that limit of understanding that is already applied to living brains matches closely with the strange fuzz in AI, it’s then likely that it’s only a matter of time (not lths ways every talking head rattled off this kinda stuff) that rather new Turing tests will emerge with similar hypothesis’ to nail down the likelihood prompts turning into discussions.
"Looking Inside Neural Networks with Mechanistic Interpretability" by Chris Olah. Delivered at the 2023 San Francisco Alignment Workshop.Image CreditsGlenn C...
