A. H. Zakai

LLMs can pass the Turing Test. That, not their successes on math and code, is why we can expect their general abilities to continue to make progress. The Turing Test by definition requires general ability, and Turing was very right to focus on that.

Math and code are two tasks that are easy to verify, but that we can't objectively measure LLM ability on, say, summarization of the main points in a novel, doesn't mean it isn't doing very well there.

LLMs can pass the Turing Test. That, not their successes on math and code, is why we can expect their general abilities to continue to make progress. The Turing Test by definition requires general ability, and Turing was very right to focus on that.

Good article, but I disagree. My main issue: if aliens arrived tomorrow & exhibited clearly-sentient behavior, we would believe they were conscious. We wouldn't wait to check what's inside their heads, which substrate-dependence views ("only carbon-based life can be conscious") would have us do.

It would be absurd to say "Klingons aren't conscious" in Star Trek (even before dissecting one, or before learning about the Progenitors and the shared origins)

Good article, but I disagree. My main issue: if aliens arrived tomorrow & exhibited clearly-sentient behavior, we would believe they were conscious. We wouldn't wait to check what's inside their heads, which substrate-dependence views ("only carbon-based life can be conscious") would have us do.

Revisiting Suleyman's "AI is not conscious" article from last month, I was struck by the claim that "a growing body of evidence" suggests consciousness requires life. A "growing body of evidence" - ?

(Side note: both of Suleyman's supportive links for the "growing evidence" claim are from the same author. In fact they have almost the same content. One is an academic paper and the other seems to be a more accessible, expanded version of it.)

Revisiting Suleyman's "AI is not conscious" article from last month, I was struck by the claim that "a growing body of evidence" suggests consciousness requires life. A "growing body of evidence" - ?

In the two links Suleyman provides as support, there is no empirical evidence that points one way or the other. Just philosophizing. There isn't a growing body of evidence, there is a growing body of *arguments*. Which is great! But it isn't evidence.

Revisiting Suleyman's "AI is not conscious" article from last month, I was struck by the claim that "a growing body of evidence" suggests consciousness requires life. A "growing body of evidence" - ?

Anyone who says the math of neural networks is trivial should be prepared to answer questions like "why does stochastic gradient descent not get stuck in local minimums?"

Very interesting work! I definitely *want* this to be true, and it seems plausible. If it is, it could end up a Very Big Deal, as in, very possibly the neurons in our brains are also "clean" in this sense. ...but I would wait for some peer review:

If 90%+ of neurons are ("clean") detectors, then I'm puzzled as to where computation/representation/etc. are done. The article does mention attention heads, the Transformer mechanism that may well underlie name-agnostic computation. So... maybe that is it? But is 90% of LLM work just detection..?

Very interesting work! I definitely *want* this to be true, and it seems plausible. If it is, it could end up a Very Big Deal, as in, very possibly the neurons in our brains are also "clean" in this sense. ...but I would wait for some peer review:

I mean that there can be many interpretations for a set of things, and any set of things might have *some* interpretation. But there is causal experimentation here too (steering), so this concern may be somewhat addressed.

Very interesting work! I definitely *want* this to be true, and it seems plausible. If it is, it could end up a Very Big Deal, as in, very possibly the neurons in our brains are also "clean" in this sense. ...but I would wait for some peer review:

First, as a quick summary: the article argues that the great majority (90%+) of neurons in artificial neural networks are actually easy to interpret. Like "this is a neuron for terrier dogs". And this has been staring us in the face, almost, but scientists missed it. Exciting, if verified!

I read Blindsight, since so many people are talking about it. It's full of wild, clever ideas, and doesn't have a boring moment. A fun read. ...but what the author was trying to *do* with those ideas is not really clear to me.

In Tao's defense, I think his (valid) point here is the risk of misalignment. That seems clear from his next few slides after. He isn't saying "we must fully understand medical drugs" (factually false; and maybe he is ignorant of that fact, but I suspect he's just speaking loosely).

He *is* saying "a stage 3 trial + a Lean proof is not enough". Because the AI might find a loophole to get through the medical trial, and the Lean proof might in theory depend on a Lean bug or be stated in a way that misleads as to what is proven.

The ancient Greeks were geniuses, but I don't think they knew better than us here. It is only around the Enlightenment, long after them, when the question of consciousness really becomes challenging, because of modern science.

That's why thought experiments like Leibniz's Mill came up then, which really stress the mind-body problem. Also why Descartes suggested the pineal gland was where the soul connects to the body, which for him was just a machine, a clockwork, and hence entirely separate from thought.

The ancient Greeks were geniuses, but I don't think they knew better than us here. It is only around the Enlightenment, long after them, when the question of consciousness really becomes challenging, because of modern science.