AI bros benefit more from people NOT understanding AI than otherwise. If suddenly we can peer into these and see how much of a waste they are, it might bring the whole market down a peg.
OK, so the article suggests he basically ‘re-discovered’ those edge, line and corner-detecting neurons in the optical cortex? (But as an analogue encoded in LLMs). If I remember anything from my college biology classes…
Big Deal™. It does diddly squat in addressing the ‘confidently incorrect’ hallucination / Bullshit-generator problem fundamental to LLMs.
It does diddly squat in addressing the ‘confidently incorrect’ hallucination / Bullshit-generator problem fundamental to LLMs
Eh, I can see some potential relevance. If an AI starts by dividing objects into “pointy things” and “round things” where a human would divide them into “animals” and “artifacts”, the AI would be more likely to mistake a cat for a car where a human would mistake a cat for a dog. Neither mistake is objectively worse than the other; but to a human, the former is a bizarre hallucination while the latter is an excusable misperception.
Yeesh, if the “AI safety folks” are celebrating this, they’re easily impressed.
The big revelation seems to be that neurons that appear to recognize more than one thing (e.g. cars and cats) are actually just recognizing a common feature between those two things (in this case, a certain “V” shape).
And… yeah? Obviously? That’s long been our understanding of how neurons work. It’s nice to have confirmation, but I don’t see how this makes LLMs (or any other large convolutional network) any more “transparent”.
