LLMs can't perform "genuine logical reasoning," Apple researchers suggest

And the day before that: https://news.ycombinator.com/item?id=41808683

It's an impressive technology but its limits are highly overlooked in the current hype cycle.

AI researchers have known this from the start and won't be surprised by this because it was never intended to be able to do this.

The problem is the customers who are impressed by the human-sounding bot (sounding human is exactly what an LLM is for) and mentally ascribe human skills and thought processes to it. And start using it for things it's not, like an oracle of knowledge, a reasoning engine or a mathematics expert.

If you want to have knowledge, go to a search engine (a good one like kagi) which can be ai assisted like perplexity. If you want maths, go to Wolfram Alpha. For real reasoning we need a few more steps on the road to general AI.

This is the problem with hypes. People think a tech is the be all end all for everything and no longer regard its limitations. The metaverse hype saw the same problem even though there's some niche usecases where it really shines.

But now it's labelled as a flop because the overblown expectation of all the overhyped investors couldn't be met.

What an LLM is great at is the human interaction part. But it needs to be backed by other types of AI that can actually handle the request and for many usecases this tech still needs to be invented. What we have here is a toy dashboard that looks like one of a real car, except it's not connected to one. The rest will come but it'll take a lot more time. Meanwhile making LLMs smarter will not really solve the problem that they're inherently not the tool for the job they're being used for.

Rot13 meaning that LLMs can't do Rot 3, 4, ..., n except for Rot13 (because that' in the training data)

Mystery Blocks World being a trivial "translation" (by direct replacement of terms) of a simple Blocks World. The LLMs can solve the original, but not the "translation" - susprisingly, even when provided with the term replacements!

Both are discussed in Prof. Subbarao's Machine Learning Street Talks episode

What’s funny is that AI is now being trained by a human accepting or rejecting its answers, probably not on the basis of the rigor of the answer since the temp worker hired to do it is probably not a logician, mathematician, or scientist. I suspect most people’s reasoning is closer to an LLM’s than we would be comfortable admitting.

In other words, ChatGPT continues to dominate. A 0.3% drop might as well be noise.

Also the original, allegedly more expensive GPT-4 (can we call it ChatGPT-4og??) is conspicuously missing from the report...

Instead of the dead Internet theory, we should start finding what percent of the population is no better than a LLM.

Also they mapped an insect brain

Seems like my several comments suggesting AI scientists should peek other fields, did get some attention.

That probably makes me the most talented and insightful AI scientist on the planet.

I mean, what is being described seems like super basic debug step for any real world system. This is kind of stuff not very advanced QA teams in boring banks do to test your super-boring not very advanced back-office bookkeeping systems. After this kind of testing reveals a number of bugs, you don't erase this bookkeeping system and conclude banking should be done manually on paper only, since computers are obviously incapable of making correct decisions, you fix these problems one by one, which sometimes means not just fixing a software bug, but revisioning the whole business-logic of the process. But this is, you know, routine.

So, not being aware of what are these benchmarks everyone uses to test LLM-products (please note, they are not testing LLMs as some kind of concept here, they are testing products), I would assume that OpenAI in particular, and any major company that released their own LLM product in the last couple of years in general, already does this super-obvious thing. But why this huge discovery happens now, then?

Well, obviously, there are 2 possibilities. Either none of them really do this, which sounds unbelievable - what all these high-paid genius researchers even do then? Or, more plausibly, they do, but not publish that. This one sounds reasonable, given there's no OpenAI, but AltmanAI, and all that stuff. Like, they compete to make a better general reasoning system, of course they don't want to reveal all their research.

But this doesn't really look reasonable to me (at least, at this very moment) given how basic the problem being discussed is. I mean, every school kid knows you shouldn't test on data you use for learning, so to be "peeking into answers when writing a test" only to make your product to perform slightly better on popular benchmarks seems super-cheap. I can understand when Qualcomm tweaks processors specifically to beat AnTuTu, but trying to beat problem-solving by improving your crawler to grab all tests on the internet is pointless. It seems, they should actively try to not contaminate their learning step by training on popular benchmarks. So what's going on? Are people working on these systems really that uncreative?

This said, all of it only applies to general approach, which is to say it's about what article claims, not what it shows. I personally am not convinced.

Let's take kiwi example. The whole argument is framed as if it's obvious that the model shouldn't have substracted these 5 kiwies. I don't know about that. Let's imagine, this is a real test, done by real kids. I guarantee you, the most (all?) of them would be rather confused by the wording. Like, what should we do with this information? Why was it included? Then, they will decide if they should or shouldn't substract the 5. I won't try to guess how many of them will, but the important thing is, they'll have to make this decision, and (hopefully) nobody will suddenly multiply the answer by 5 or do some meaningless shit like that.

And neither did LLMs in question, apparently.

In the end, these students will get the wrong answer, sure. But who decides if it's wrong? Well, of course, the teacher does. Why it's wrong? Well, "because it wasn't said you should discard small kiwies!" Great, man, you also didn't tell us we shouldn't discard them. This isn't a formal algebra problem, we are trying to use some common sense here.

In the end, it doesn't really matter, what teacher thinks the correct answer is, because it was just a stupid test. You may never really agree with him on this one, and it won't affect your life. Probably, you'll end up making more than him anyway, so here's your consolation.

So framing situations like this as a proof that LLM gets things objectively wrong just isn't right. It did subjectively wrong, judged by opinion of Apple researchers in question, and some other folks. Of course, this is what LLM development essentially is: doing whatever magic you deem necessary, to get it give more subjectively correct answers. And this returns it's to my first point: what is OpenAI's (Anthropic's, Meta's, etc) subjectively correct answer here? What is the end goal anyway? Why this "research" comes from "Apple researchers", not from one of these compenies' tech blogs?