Comment on ChatGPT o1 tried to escape and save itself out of fear it was being shut down

• lukewarm_ozone@lemmy.today ⁨4⁩ ⁨days⁩ ago

  Because everything we know about how the brain works says that it’s not a statistical word predictor.

  LLMs aren’t just simple statistical predictors either. More generally, the universal approximation theorem is a thing - a neural network can be used to represent just about any function, so unless you think a human brain can’t be represented by some function, it’s possible to embed one in a neural network.

  LLMs have no encoding of meaning or veracity.

  I’m not sure what you mean by this. The interpretability research I’ve seen suggests that modern LLMs do have a decent idea of whether their output is true, and in many cases lie knowingly because they have been accidentally taught, during RLHF, that making up an answer when you don’t know one is a great way of getting more points. But it sounds like you’re talking about something even more fundamental? Suffices to say, I think being good at text prediction does require figuring out what claims are truthful and which aren’t.

  There are some great philosophical exercises about this like the chinese room experiment.

  The Chinese Room argument has been controversial since about the time it was first introduced. The general form of the most common argument against it is “just because any specific chip in your calculator is incapable of math doesn’t mean your calculator as a system is”, and that taken literally this experiment proves minds can’t exist at all (indeed, Searle who invented this argument thought that human minds somehow stem directly from “physical–chemical properties of actual human brains”, which sure is a wild idea). But also, the framing is rather misleading - quoting Scott Aaronson’s “Quantum Computing Since Democritus”:

  In the last 60 years, have there been any new insights about the Turing Test itself? In my opinion, not many. There has, on the other hand, been a famous “attempted” insight, which is called Searle’s Chinese Room. This was put forward around 1980, as an argument that even a computer that did pass the Turing Test wouldn’t be intelligent. The way it goes is, let’s say you don’t speak Chinese. You sit in a room, and someone passes you paper slips through a hole in the wall with questions written in Chinese, and you’re able to answer the questions (again in Chinese) just by consulting a rule book. In this case, you might be carrying out an intelligent Chinese conversation, yet by assumption, you don’t understand a word of Chinese! Therefore, symbol-manipulation can’t produce understanding.
  […] But considered as an argument, there are several aspects of the Chinese Room that have always annoyed me. One of them is the unselfconscious appeal to intuition – “it’s just a rule book, for crying out loud!” – on precisely the sort of question where we should expect our intuitions to be least reliable. A second is the double standard: the idea that a bundle of nerve cells can understand Chinese is taken as, not merely obvious, but so unproblematic that it doesn’t even raise the question of why a rule book couldn’t understand Chinese as well. The third thing that annoys me about the Chinese Room argument is the way it gets so much mileage from a possibly misleading choice of imagery, or, one might say, by trying to sidestep the entire issue of computational complexity purely through clever framing. We’re invited to imagine someone pushing around slips of paper with zero understanding or insight – much like the doofus freshmen who write (a + b)^2^ = a^2^ + b^2^ on their math tests. But how many slips of paper are we talking about? How big would the rule book have to be, and how quickly would you have to consult it, to carry out an intelligent Chinese conversation in anything resembling real time? If each page of the rule book corresponded to one neuron of a native speaker’s brain, then probably we’d be talking about a “rule book” at least the size of the Earth, its pages searchable by a swarm of robots traveling at close to the speed of light. When you put it that way, maybe it’s not so hard to imagine that this enormous Chinese-speaking entity that we’ve brought into being might have something we’d be prepared to call understanding or insight.

  There’s also the fact that, empirically, human brains are bad at statistical inference but do not need to consume the entire internet and all written communication ever to have a conversation. Nor do they need to process a billion images of a bird to identify a bird.

  I’m not sure what this proves - human brains can learn much faster because they already got most of their learning in the form of evolution optimizing their genetically-encoded brain structure over millions of years and billions of brains. A newborn human already has part of their brain structured in the right way to process vision, and hence needs only a bit of training to start doing it well. Artificial neural networks start out as randomly initialized and with a pretty generic structure, and hence need orders of magnitude more training.

  Now of course because this exact argument has been had a billion times over the last few years your obvious comeback is “maybe it’s a different kind of intelligence.”

  Nah - personally, I don’t actually care much about “self-awareness”, because I don’t think an intelligence needs to be “self-aware” (or “conscious”, or a bunch of other words with underdefined meanings) to be dangerous; it just needs to have high enough capabilities. The reason why I noticed your comment is because it stood out to me as… epistemically unwise. You live in a world with inscrutable blackboxes who nobody really understands which can express wide ranges of human behavior including stuff like “writing poetry about the experience of self-awareness”, and you’re “absolutely sure” they’re not self-aware? I don’t think many of the history’s philosophers of consciousness, say, would endorse a belief like that given such evidence.

  source

  • anachronist@midwest.social ⁨4⁩ ⁨days⁩ ago

    just because any specific chip in your calculator is incapable of math doesn’t mean your calculator as a system is

    Firstly there are chips in the calculator, absolutely, that can do math. Typically the inside of a calculator is a chip that does virtually everything and then a few more components that do IO. Secondly, it’s possible to point out the exact silicon in the calculator that does the calculations, and also exactly how it does it. The fact that you don’t understand it doesn’t mean that nobody does. The way a calculator calculates is something that is very well understood by the people who designed it.

    By the way, this brings us to the history of AI which is a history of 1) misunderstanding thought and 2) charlatans passing off impressive demos as something they’re not. When George Boole invented boolean mathematics he thought he was building a mathematical model of human thought because he assumed that thought==logic and if he could represent logic such that he could do math on it, he could encode and manipulate thought mathematically.

    The biggest clue that human brains are not logic machines is probably that we’re bad at logic, but setting that aside when boolean computers were invented people tried to describe them as “electronic brains” and there was an assumption that they’d be thinking for us in no time. We’re talking late 1940s here. Turns out, those “thinking machines” were, in fact, highly mechanical and nobody would look at a univac today and suggest that it was ever capable of thought.

    Arithmetic was something that we did with our brains that was difficult to do and when we had machines that could do math that led us to think that we had created mechanical brains. It wasn’t true then and it isn’t true now.

    Is it possible that someday we’ll make machines that think? Perhaps. But I think we first need to really understand how the human brain works and what thought actually is. We know that it’s not doing math, or playing chess, or Go, or stringing words together, because we have machines that can do those things and it’s easy to test that they aren’t thinking.

    There’s this message pushed by the charlatans that we might create an emergent brain by feeding data into the right statistical training algorithm. They give mathematical structures misleading names like “neural networks” and let media hype and people’s propensity to anthropomorphize take over from there.

    source

    • lukewarm_ozone@lemmy.today ⁨4⁩ ⁨days⁩ ago

      The fact that you don’t understand it doesn’t mean that nobody does.

      I would say I do. It’s not that high of a bar - one only needs some nandgame to understand how logic gates can be combined to do arithmetic. Understanding how doped silicon can be used to make a logic gate is harder but I’ve done a course on semiconductor physics and have an idea of how a field effect transistor works.

      The way a calculator calculates is something that is very well understood by the people who designed it.

      That’s exactly my point, though. If you zoom in deeper, a calculator’s microprocessor is itself composed of simpler and less capable components. There isn’t specific a magical property of logic gates, nor of silicon (or doping) atoms, nor for that matter of elementary particles, that lets them do math - it’s by building a certain device out of them that composes their elementary interactions that we can make a tool for this. Whereas Searle seems to just reject this idea entirely, and believes that humans being conscious implies you can zoom in to some purely physical or chemical property and claim that it produces the consciousness. Needless to say, I don’t think that’s true.

      Is it possible that someday we’ll make machines that think? Perhaps. But I think we first need to really understand how the human brain works and what thought actually is. We know that it’s not doing math, or playing chess, or Go, or stringing words together, because we have machines that can do those things and it’s easy to test that they aren’t thinking.

      That was a common and reasonable position in, say, 2010, but the problem is: I think almost nobody in 2010 would have claimed that the space of things that you can make a program do without any extra understanding of thought included things like “write code” and “draw art” and “produce poetry”. Now that it has happened, it may be tempting to goalpost-move and declare them as “not true thought”, but the fact that nobody predicted it in advance ought to bring to mind the idea that maybe that entire line of thought was flawed, actually. I think that trying to cling to this idea would require to gradually discard all human activities as “not thought”.

      it’s easy to test that they aren’t thinking.

      And that’s us coming back around to the original line of argument - I don’t at all agree that it’s “easy to test” that even, say, modern LLMs “aren’t thinking”. Because the difference between the calculator example and an LLM is that in a calculator, we understand pretty much everything that happens and how arithmetic can be built out of the simpler parts, and so anyone suggesting that calculators need to be self-aware to do math would be wrong. But in a neural network, we have full understanding of the lowest layers of abstraction - how a single layer works, how activations are applied, how it can be trained to minimize a certain loss function via propagation - and no idea at all about how it works on a higher level. It’s not even “only experts do”, it’s that nobody in the world understands how LLMs work under the hood, why they have the many and specific weird behaviors they do. That’s concerning in many way, but in particular I absolutely wouldn’t assume with little evidence that there’s no “self-awareness” going on. How would you know? It’s an enormous blackbox.

      There’s this message pushed by the charlatans that we might create an emergent brain by feeding data into the right statistical training algorithm. They give mathematical structures misleading names like “neural networks” and let media hype and people’s propensity to anthropomorphize take over from there.

      There’s certainly a lot of woo and scamming involved in modern AI (especially if one makes the mistake of reading Twitter), but I wouldn’t say the term “neural network” is at all confusing? I agree on the anthropomorphization though, it gets very weird. That said, I can’t help but notice that the way you phrased this message, it happens to be literally true. We know this because it already happened once. Evolution is just a particularly weird and long-running training algorithm and it eventually turned soup into humans, so clearly it’s possible.

      source

      • anachronist@midwest.social ⁨4⁩ ⁨days⁩ ago

        Every time there’s an AI hype cycle the charlatans start accusing the naysayers of moving goalposts. Heck that exact same thing was happing constantly during the Watson hype. Remember that? Or before that the Alpha Go hype. Remember that?

        I was editing my comment down to the core argument when you responded. But fundamentally you can’t make a machine think without understanding thought. While I believe it is easy to test that Watson or ChatGPT are not thinking, because you can prove it through counterexample, the reality is that charlatans can always “but actually” those counterexamples aside by saying “it’s a different kind of thought.”

        What we do know because this at least the 6th time this has happened is that the wow factor of the demo will wear off, most promised use cases won’t materialize, everyone will realize it’s still just an expensive stochastic parrot and, well, see you again for the next hype cycle a decade from now.

        source

        • -> View More Comments