Wasn’t it designed to do just that? My dog has demonstrated theory of mind, dogs were bred to do that. Theory of Mind is no biggie.
No, they weren’t designed to do that at all. No one even thought they were capable of it.
And I’m not sure your dog has theory of mind. Theory of mind in animals is still speculative. Chimps may have it, and maybe dogs, but we’re not sure.
Moreover, it’s pretty hard to imagine that a simple ‘lookup table’ or ‘stochastic word generator’ could exhibit theory of mind. And of course the LLMs didn’t until they got to a humongous size - then it just emerged.
They were designed to analyze text and derive human and behavior from that. It’s clearly using Theory of Mind to talk to humans. I don’t say it understands humans in the way humans do, but it’s pretty easy to logically guess at a human’s state of mind based on their words and comparable words from others. Humans love to tell you their state of mind, they write about it all the time for some reason.
It is not at all clear that it has an actual theory of mind; it is responding in a way that emulates how the textual base that forms its training set indicates how people respond, but emulation is not cognition. When ChatGPT or another language tool comes up with a unique insight that isn’t just some convolution of text that it has studied, wake me up.
Stranger
Are you sure humans do anything different? Programming a computer to emulate Theory of Mind in the simple language analysis world of current AI doesn’t sound like a stretch. It’s not like all humans are that good comprehending the state of mind of others, surely you’ve worked with engineers before and have observed that, but is it a necessity to define intelligence?
Even if a machine intelligence system were actually “thinking”, it would be very different from what a human brain does because how it is structured and functions is fundamentally different. But even the most sophisticated chatbot like ChatGPT isn’t actually formulating any kind of theory of mind, and it certainly isn’t in any way “programmed to emulate a Theory of Mind”; it is given an algorithmic framework to analyze textual data and respond in ways that are consistent with what it finds in that data, using Bayesian statistics to refine its responses until it roughly approximates an actual human’s response. The ability to formulate these responses certainly is an emergent phenomenon, but that doesn’t mean that there is a deeper sapience within it, much less an ability to actually conceive of an independent sentience or sapience. ChatGPT is ‘clever’ in the sense that it uses an enormous amount of computational power at its disposal to produce somewhat human-like textual responses, but there is zero evidence that it is actually producing anything novel, and in many cases clearly not grasping semantic errors.
It is one thing to not understand the inner thoughts of another person, and if that is your personal standard for having a theory of mind then you could legitimately argue that no one does. But the vast majority of people—yes, even engineers—understand that other people have their own internal thoughts and ideas. A chatbot is just responding as its very complex data model indicates that it should respond, and its ‘understanding’ of anything is built from that model. When such a system actually presents a novel idea or produces some work that isn’t clearly derived from something it has been trained on, then I’ll show some enthusiasm for the notion that it is actually ‘thinking’ in some manner that we might consider to be ‘intelligent’ instead of just ‘clever’.
Stranger
But that’s not what happened. No one has ‘programmed’ these LLMs to do this. They were programmed to ingest a whole lot of material, and then went through reinforcement learning and unsupervised learning from a large dataset. During this process, the AI constantly updates its neural net parameters to improve scores on a fitness test, using back propagation and other methods. There are no algorithms for poetry, theory of mind, arithmetic, translation, or anything else. Just a giant neural net. Theory of mind just emerged after a certain number of parameters and training FLOPS, as did everything else these models do. No one coded it, planned it or even expected it.
I understand what you are trying to say, but software has long advanced beyond such a simple idea of what coding is. It was coded to be able to produce output as if it had Theory of Mind. It’s just software, not self-realizing yet.
@Stranger_On_A_Train , I don’t contend these early AIs do what humans seem to do process-wise. I’ve seen no indication they attempt to emulate a person’s mind to understand their state of mind. But we will have to contend with many machine behaviors that at some point will be observably indistinguishable from our own but arrived at in a complete different manner.
I don’t hold out hope for the “intelligence” of ChatbotGPT. I was demonstrating it earlier and asked it a straightforward question. “Who won the most recent Super Bowl?” Chatbot insisted that the most recent Super Bowl was LVI and that it was played on 2/13/2022 at Raymond James stadium and that the Tampa Bay Buccaneers emerged victorious over the Cincinnati Bengals. I asked it multiple times and in various ways that it should check and make sure, but it kept giving me the same incorrect response. For some reason it was mixing up the results from Super Bowl LV and LVI, and didn’t even realize that LVII had already been played and is the most recent Super Bowl. Needless to say the person I was demonstrating it to was not impressed.
For one, it has “limited knowledge of events past 2021”. It says right on the front page. Its training data ends around there. Secondly, it is not very good right now at those trivia-type questions. Simple web look-ups are not its strong point.
Who doesn’t?
The difference between ChatGPT and more generally all large language model AIs is an important one to make, especially in light of the uses to which someone might put these AIs.
The theory of mind bit is interesting, thanks for linking to it. My thinking on this is leaning towards this explanation found in the discussion of the paper you linked to:
It is possible that GPT-3.5 solved ToM tasks without engaging ToM, but by discovering and leveraging some unknown language patterns. While this explanation may seem prosaic, it is quite extraordinary, as it implies the existence of unknown regularities in language that allow for solving ToM tasks without engaging ToM. Such regularities are not apparent to us (and, presumably, were not apparent to scholars that developed these tasks). If this interpretation is correct, we would need to re-examine the validity of the widely used ToM tasks and the conclusions of the decades of ToM research: If AI can solve such tasks without engaging ToM, how can we be sure that humans cannot do so, too?
I think Kosinski sells this explanation a little short - individual humans are incapable of examining or even reading data sets as large as the training data given to ChatGPT; therefore if theory-of-mind-like ideas are found in patterns of text, it would be impossible for a human without AI assistance to find them anyway. It should be unsurprising that if such patterns existed, we would be unable to find them without the help of an AI. What do you think?
I have been using an Ai program lately that has literally blown my mind with what it can do. I have struggled my entire life with being fragmented and I spend too much time going down rabbit holes. The AI program allows me to do this without missing a beat. It remembers everything and can put it into any order I ask it to. It gives me a lot of prompts that fit right into the work. I got more work done in two days than I have in 5 years. It is all my work, I don’t feel at all like I am coping out some how. It is bigger than just a writing tool, it allows the writer to become bigger in a lot of ways. If we don’t take the lazy approach it really helps to produce quality work. But I can also see just the sheer amount of AI produced material watering everything down that is produced. This will be interesting.
Isaac Asimov, one of my heroes (obviously) said somewhere that the most important moment for a scientist is the one that as the result of an experiment says, “That’s odd.”
Think of the Michaelson-Morley experiment that failed to find the expected velocity of the earth with respect to the ether. Or Rutherford’s experiment that seemed to show that the atom is nearly entirely empty space. At any such point a scientist stops, thinks and eventually comes up with a conjecture. And then designs experiments to try to refute. The more they fail to refute it, the more plausible the conjecture becomes. This, to me, is the true scientific method and nothing I have seen of AI suggests it is remotely near being able to carry it out.
FWIW
AI has reached a point that it can generate reasonable hypotheses and develop approaches to test them. It can scan very disparate literature sources, representing abstracts as vectors, and find surprising commonalities that suggest novel hypotheses.
Awareness is not required.
Interesting. I wouldn’t have believed it. I wonder what AI would do with the Riemann hypothesis or the P = NP question (the biggest unsolved problems in math and computer science, resp.).
This is so incredibly true. When doing my PhD work, there was a moment where my AI was getting the right answer for the wrong reason. It was very much an “That’s odd (and frustrating)” moment. But when I figured out why, it was a huge step forward for the research.