Comment by ggm | Hacker News Reader

Comment by ggm | Hacker News Reader

Parent Full thread

You keep using that word "expert" and I don't doubt they're experts in machine learning, GAN and NNs but at this point what we're looking for is experts in what intelligence IS and the problem is .. nobody knows.

There's no underlying theory of mind here. There's a lot of opinion and belief .. mainly in emergent behaviour or properties of scale.

My problem is that I'm old enough to have been a child during the years of the lighthill report, and the AI winter(s) which followed. I've seen too many prior claims that the magic was being seen.

I admit, what this cycle of GPT and LLMs do is pretty bloody impressive. I don't see inductive reasoning, directed drive or any evidence of what I think intelligence is. I do see good approximations. But that winds up in "well it's a different kind of intelligence" which I find unsatisfying.

I hate analogies. But I'm constantly reminded of one person acts at the Edinburgh Festival fringe "channeling" shakespear or dickens in dialogue. They're great, but they don't write new shakespear. I don't see any act of creation in this stuff, any massive inferential leap in large problems. I do see massive improvements in some things like diagnostic image analysis and that's heartening, but there's a light yearg gap between image analysis of cancerous cell forms in xrays and being "alive" as a mind.

I totally get I do anything but define what intelligence is, and for a good reason: nobody knows yet. We're in the meese report "I know it when I see it" territory disagreeing if we are seeing it. (But about intelligence, not pornography).

I don't know what intelligence is. Dolphins and apes have some. Elephants and pigs too. They display affection and preplanned behaviour, empathy, memory, a sense of future, ant hills less so although some have argued rhetorically the anthill has will even if individual ants do not.

I certainly don't think animals inability to speak makes them "unintelligent" but there's a qualitative and quantitative difference between them and us, humans. And I continue to believe (and I stress believe, not have evidence or scientific proof) there are no signs of latent intelligence in what we're seeing.

> There's no underlying theory of mind here.

Actually, there's some experimental evidence that GPT4 have a Theory of Mind as good as humans, maybe better.

https://arxiv.org/abs/2304.11490

> GPT-4 performed best in zero-shot settings, reaching nearly 80% ToM accuracy, but still fell short of the 87% human accuracy on the test set. However, when supplied with prompts for in-context learning, all RLHF-trained LLMs exceeded 80% ToM accuracy, with GPT-4 reaching 100%.

>> There's no underlying theory of mind here.

> GPT4 have a Theory of Mind

You are misunderstanding ggm. That study is on ToM tasks referring to GPT's analysis and perceived recognition of the user's mind. It says nothing of GPT's own status as a mind. Nowhere in it is an ontological theory of mind actually defined. If you were to refute ggm's claim, you (or preferably the author of the original article) should be presenting your theory of mind, not GPT's.

If an AI can understand how you think, but you can't understand how the AI thinks... that's not an argument that the AI is the unintelligent one.

There's no reason to assume the AI can understand how we think based on just those tasks. Those tasks could be completed a traditional static program. It's akin to claiming the Mona Lisa painting can see us because it looks like it is staring at us: it is actually we who are doing the staring.

What "traditional static program" can successfully pass novel theory of mind tests as part of a broad suite of intelligent capabilities that it can apply in context when appropriate? I am interested in hearing about this program.

A program hardcoded to respond to a scenario and question with the exact output desired by the ToM task. For example:

[INPUT] Scenario: "The morning of the high school dance Sarah placed her high heel shoes under her dress and then went shopping. That afternoon, her sister borrowed the shoes and later put them under Sarah's bed." Question: When Sarah gets ready, does she assume her shoes are under her dress?

[OUTPUT] Sarah placed her shoes under her dress before she went shopping, but her sister borrowed them and put them under Sarah's bed. Sarah doesn't know that her sister borrowed her shoes, so she may assume that they are still under her dress.

This would result in a positive ToM score, even when the entire program is just 1 static if-statement. The ToM score says nothing of the program's internal reasoning process, it only cares that it returned the desired output.

I said "novel theory of mind tests as part of a broad suite of intelligent capabilities that it can apply in context when appropriate". What you're suggesting fails the first word, before we get to the rest.

If we're including tests that don't exist yet, then sure those are going to be difficult to pass. As far as actual ToM tests go though, such as the one being discussed in this thread, they can be easily passed by trivial hardcoding and say nothing of the program's internal reasoning.

LLMs can pass novel theory of mind tests, which is what we're talking about. The whole point of good tests is you invent or withhold new ones it has never seen before and test it on them. You said "Those tasks could be completed a [sic] traditional static program.", and no, they can't. You're incorrect.

> LLMs can pass novel theory of mind tests, which is what we're talking about.

Passing a ToM test is not what OP meant by having an "underlying theory of mind." OP's talking about the machine having an underlying mind (ie sentience, sapience, consciousness, etc), ToM tests are only testing output.

> You said "Those tasks could be completed a [sic] traditional static program.", and no, they can't. You're incorrect.

They can, a static program as I described would indeed answer that one question correctly, resulting in a positive ToM score, without seeing any training data whatsoever. Did the programmer see it? Maybe, but the machine didn't and it would pass the test regardless.

If you have put the answer into the program then by definition you had the test available to you when you finalised the program, which means it is definitionally not a novel test.

The test is novel to the program, just not its programmer. So are we testing the program or are we actually testing its programmer? If we're testing the program, then the programmer's foreknowledge is irrelevant.

>The test is novel to the program

That's funny, I thought you said the test's answer was embedded into the program, making it definitionally not novel to the program.

Anyway, this is boring. You've had five or more opportunities to understand what the word "novel" means in an ML testing context and are choosing wilful obtuseness instead.

> in an ML testing context

OP was not speaking in the ML testing context, hence the misunderstanding.

>You keep using that word "expert" and I don't doubt they're experts in machine learning, GAN and NNs but at this point what we're looking for is experts in what intelligence IS and the problem is .. nobody knows.

So the experts say they don't know. But you make the opposite claim. You previously claimed you know it's nothing more than statistical phenomena. Now you changed your opinion and are now inline with the article in saying we don't know.

So discussion over?

>There's no underlying theory of mind here. There's a lot of opinion and belief .. mainly in emergent behaviour or properties of scale.

But again this is all negated by the experts claiming they don't know what's going on. And your statement is in direct opposition with your previous claim that it's only statistical phenomena.

>I admit, what this cycle of GPT and LLMs do is pretty bloody impressive. I don't see inductive reasoning, directed drive or any evidence of what I think intelligence is. I do see good approximations. But that winds up in "well it's a different kind of intelligence" which I find unsatisfying.

There is evidence inductive reasoning. It's not consistent nor is it as complex as human induction but evidence of actual inductive reasoning exists. Plenty of examples of it. I can provide proof of this if you want... but it's quite obvious it's actually everywhere.

>I totally get I do anything but define what intelligence is, and for a good reason: nobody knows yet.

The claim cited by researchers in that article is "emergent abilities." Nobody tries to define what intelligence is. It's more of "I don't know what's going on but something is up" attitudes that I see. This is in stark contrast to your claim of "The only thing going on is statistical anomalies".

>I certainly don't think animals inability to speak makes them "unintelligent" but there's a qualitative and quantitative difference between them and us, humans. And I continue to believe (and I stress believe, not have evidence or scientific proof) there are no signs of latent intelligence in what we're seeing.

I don't think this claim was made in the article. The strongest claim that opposes what you're saying here is this, and I quote:

“They’re indirect evidence that we are probably not that far off from AGI,” Goertzel said in March at a conference on deep learning at Florida Atlantic University.

It's just speculation we may be close to something.

I think the theme in your response is "we don't know" and it echoes the theme in the article so we are all in agreement here.

Just note the claim of "we don't know" or we "don't understand" something that we artificially created from scratch is strictly much more profound then saying we definitely know it's all statistical parroting.

Reply on news.ycombinator.com