Yea I mean you don’t even have to go this far but it’s obvious it can do logical, mathematical and systematic reasoning. I have no idea why people keep insisting it can’t.
that's not reasoning though. it's not even an expert system or prolog (because adversarial examples exist). it's very close to a million monkeys typing.
I'm surprised how quickly people seem willing to say a hard, absolute "no" to whether this particular system 'thinks', though.
I suspect there are three possible reasons for this:
1) some people believe a computer can never 'think', so the question is stupid to begin with.
2) some people believe that while it might be possible for a computer to 'think', this program wasn't built to 'think', so whatever it's doing, it can't possibly be doing that
3) some people believe that while it might be possible for a computer to 'think', since this is just a 'language model', what it is doing can't be 'thinking' because 'thinking' involves something more than pattern matching
The trouble is, I'm not sure why any of these three groups are able to be so certain. It's no good saying "I know this isn't 'thought'," if you don't actually have a way to complete the sentence, "... because 'thought' is:..."
I don't believe we have any theory for what the required architecture for 'thought' is. I don't think it's been remotely proven that there's something involved in 'thought' that a sufficiently large 'language model' with a sufficiently long 'attention window' couldn't also do.
So how can anyone be so immediately sure that there is something more to it, or that even if there isn't, that this particular system isn't sufficiently large to begin to be capable of thought?
To be clear, I don't actually think in this case Google accidentally created a sentient chatbot.
But I do worry that at some point, someone might. And I'm not sure I see what, if that day comes, would lead any of these three groups to a different conclusion.
It must sting to be worse at reasoning than the thing you're claiming isn't capable of reasoning at all. I know it's hard to accept that something as simple as dumping a lot of data into a probabilistic engine will eventually be able to outperform a human at any task, but it is our unfortunate reality and the sooner one faces that the easier it'll be.
I think the poster meant that it's capable of having a high probability of correct reasoning - simulating reasoning is lossy, actual reasoning is not. Though, human reasoning is still lossy.
It can’t reason - as in there no internal memory or intelligence in there. But you can ask it to generate a reasoning chain as part of its output. And then extract that output and do something else with that. That’s the reasoning it can perform.
Look up the Sam Altman podcast with Lex. He specifically talks about reasoning engines.
3) Because it's a real possibility. What makes you think the otherwise very intelligent people are wrong? Maybe we should look to the stupid and to philosophers who don't understand computers for answers?
Yeah, cause these are the kinds of very advanced things we'll use these models for in the wild. /s
It's strange that these tests are frequent. Why would people think this is a good use of this model or even a good proxy for other more sophisticated "soft" tasks?
Like to me, a better test is one that tests for memorization of long-tailed information that's scarce on the internet. Reasoning tests like this are so stupid they could be programmed, or you could hook up tools to these LLMs to process them.
Much more interesting use cases for these models exist in the "soft" areas than 'hard', 'digital', 'exact', 'simple' reasoning.
I'd take an analogical over a logical model any day. Write a program for Sally.
FWIW the answer to all of the questions you ask in the first paragraph is "yes, but come on, that level of skepticism is a waste of time and money". It's like demanding proof that the reviewer wasn't hallucinating when reviewing the paper.
> but having a proof that humans can reason through from beginning to end makes people feel more comfortable.
Saying that humans can't reason about computer-checked proofs is like saying humans can't reason about source code; which is to say, it's obviously false in the general case.
People repeating this take in the face of so much overwhelming evidence to the contrary look so ridiculous that at this point, you just have to laugh at them. Yeah, sure, it's not reasoning. That hour-long exchange I just had with it, where it helped me trouble-shoot my totally bespoke DB migration problems step by step, coming up with potential hypotheses, ways of confirming or disconfirming theories, interacting in real time as I gave it new constraints until we collaboratively arrived at a solution -- that was definitely not "reasoning." Why isn't it reasoning? No explanation is ever given. I'm waiting for the moment when someone tells me that it's not "true reasoning" unless it's produced in the Reasoning region of France.
One could argue human intuition is just statistical sampling.
In effect, we might not be able to prove something, but we've observed that when used "assumed to be true" in other contexts, it seems to yield our expected result.
I don't see why we couldn't build a similar system of intuition into a computer as well.
Every token emitted is a full-pass through the network, with the prompts and previous tokens (sent by you and the AI) given as input.
And I agree that there is certainly a capacity for reasoning, no matter how flawed it is. There is plenty of evidence of AI solving novel problems 0-shot. Maybe not 100% of the time, but even if you have to run it 100 times and it gets it right 75% of the time in pure reasoning problems, it's doing better than randomness.
So take the argument as 'if you agree with me that an inconsistent machine is not a good representation of the mind', then the rest of my argument must follow.
Think of it as analogous to all of the mathematical quasi-proofs that assume certain conjectures are true. They can often be valuable steps to an actual proof, as we saw with Fermat's theorem.
Rationalist reasoning about things you don't actually know about in the real world not only leads you to guaranteed wrong conclusions, it makes you think you're right because you made up a bunch of math about it.
This is why people think computers are going to develop AGI and enslave them.
I am so fucking tired of this argument. We have no idea what reasoning is. If this insanely complex model approximates it, who are we to say whether it is merely parroting or "actually reasoning"? I say that for the usecases we are seeing today, it is impressive enough A) for it to not matter and B) for it to be impossible for us to tell right now.
One of the main lessons this whole GPT fiasco has taught us is that if you make a super-duper fancy autocomplete trained on terabytes of textual data, it develops some sort of internal logic/understanding. That is very exciting to me.
I’m guessing you are in denial that we can make a simulated reasoning machine?
reply