Sabitlenmiş Tweet
Why AI needs to ease up on scaling and learn how to code:
deoxyribose.github.io/No-Shortcuts-t…
@GaryMarcus @fchollet @yudapearl
English
Frans Zdyb
745 posts
Frans Zdyb
@FZdyb
Shaper of loss landscapes. No agent samples the same distribution twice, for it's not the same distribution, and she is not the same agent
Copenhagen Katılım Nisan 2014
293 Takip Edilen142 Takipçiler
@burny_tech They don't explain how hierarchical features induce low dimensional manifolds
English
What would you say are the problems with current theory of deep learning?
English
@BlancheMinerva @LucaAmb It's like saying diffusion models "transcend" the human ability to draw faces, because the faces they generate look more average. Sure, they do - is that "transcendence"?
English
@BlancheMinerva @LucaAmb The transcendence paper explains what's going on perfectly: "If these errors (rookie mistakes) are idiosyncratic, averaging across many experts would have a denoising effect, leaving the best moves with higher probability."
English
It's particularly stupid since the predictive coding hypothesis has been rather mainstream in cognitive science for long before LLMs, which effectively states that human leans by making statistical predictions
Philippe Lemoine@phl43
Once again, regardless of whether you think that ChatGPT understands anything or not, I think this argument is confused. To say that it can't possibly understand anything because it was only trained to "predict the next word" is just as idiotic as saying that humans can't understand anything because they were "trained" to survive and spread their genes. This line of argument seems to boil down to the idea that, unless something works roughly in the same way as the human brain, it can't really be intelligent, but just as the same software can run on very different types of hardware there is no reason to think that human-like intelligence couldn't be implemented in very different ways.
English
@BlancheMinerva @LucaAmb Indirect evidence would be very robust OOD generalization, eg if you always get the right product even when running in factors much longer than training examples.
English
@BlancheMinerva @LucaAmb Structural Causal Model. Direct evidence would be identifying the encoded graph and showing that different prompts correspond to evaluating queries. Eg for arithmetic, finding a circuit that is a multiplication algo and showing that different input strings result in running it.
English
@LucaAmb I agree that learning from our own actions is a small part of learning, and more comes from observing other people's actions, and even more from language. But my guess is you need the causal representations in place before these other two sources can work..
English
@LucaAmb Fair, but I'm skeptical that LLM weights learn to represent SCMs - and I think you need both actions, representations and a causal inference algorithm. LLMs may have actions, but not the other two.
English
@usablejam @AVMiceliBarone @francoisfleuret None, but I'm much better than Erdös at abstaining from amphetamine
English
English
Listen. There are two sorts of bits: The one coming from understanding, like if *I* write "2+2=4", this string is encoded in u-bits, understood-bits, good stuff.
Whereas ChatGPT writing "2+2=4" produces a string of fs-bits, aka fancy-statistics-bits. Which are crap.
Big Brain AI@realBigBrainAI
Oxford AI professor Michael Wooldridge: "ChatGPT doesn't understand anything. It's essentially doing some fancy statistics."
English
@AVMiceliBarone @francoisfleuret No, I can't even do 5 digits I think, but it doesn't matter. Understanding doesn't mean zero errors, it means deploying the correct algorithm regardless of deviation from training data distribution.
English
@FZdyb @francoisfleuret Humans can understand the concept of multiplication and still get it wrong, or just run out of working memory. Can you multiply two 10-digit numbers in your head?
English
@PersonaIData @francoisfleuret Not for arithmetic no, but it matters that it can't discover causal models in general. That's why they can't do anything humans or tools can't do, why they hallucinate, and why their capabilities are jagged - all symptoms of narrow generalization based on acausal patterns.
English
@FZdyb @francoisfleuret Does it matter if they have the ability to write their own calculator tool and use it, in machine language if it came to that?
English
@AVMiceliBarone @francoisfleuret Whereas once a human understands the concept of multiplication, algorithms, execution traces and answers all have to agree. Whatever LLMs do, it's not that. We need a different word than understanding.
English
@AVMiceliBarone @francoisfleuret You can get LLMs closer to understanding by asking it to review the algorithm in the prompt, and have it do chain of thought. But if you train it on the wrong algorithm, it might ignore it and produce correct answers; or follow it and produce wrong answers, not noticing it.
English
@paraschopra I don't see how realism makes predictions about correlates, it doesn't (and can't) posit any actual mechanism. On the other hand, the attention schema theory not only posits a mechanism, but specifically predicts that people will deny any mechanism. Seems like a slam dunk!
English
Two famous camps about the nature of consciousness is illusionism (qualia is illusion) and realism (qualia is real).
Note that both, in principle, make equivalent predictions about observables such as what will be reported by the subject ("i see red") or correlates (gamma oscillations encoding wakefulness). So, if it isn't possible to discriminate between the two, is there a point in separating them?
Pragmatism requires empirical discriminations to tell apart theories, and so far I haven't been able to discover how to tell illusionism and realism apart.
English
@aryehazan Computational functionalism + the attention schema theory are very satisfying, both independently and together. They have that "snap into place" and "nothing makes sense except in light of the theory" feel to them.
English
computational functionalism isn't terribly satisfying but it's kind of the default if you reject dualism, panpsychism, quantum spookiness
like, what else is there?
QC@QiaochuYuan
what's actually happening in the discourse is clearly stranger than this. there's a contingent of people to whom computational functionalism seems obviously correct, to whom arguments against seem like nonsense, and a contingent of people to whom computational functionalism seems obviously incorrect, to whom arguments for seem like nonsense. approximately zero meaningful communication appears to be capable of bridging this gap. this is weird! what the fuck is going on with this!
English