What this shows is flaws in the test, not that ChatGPT3 has a theory of mind.
ChatGPT3 does not even have a theory of physical objects and their relations, nevermind a theory of mind.
This merely shows that an often useful synthesis of phrases statistically likely to occur in a given context and grammar-checked, will fool people some of the time, and a better statistical model will fool more people more of the time.
We can figure out from first principles that it has none of the elements of understanding or reasoning that can produce a theory of mind, any more than the Eliza program did in 1966. So, when it appears to do so, it is demonstrating a flaw in the tests or the assumptions behind the tests. Discouraging that the researchers are so eager to run in the opposite direction; if there is confusion at this level, the general populace has no hope of figuring out what is going on here.
well, I'd first need to see that it had a Theory of Feet... ;-)
More seriously, that it can actually understand and wield abstract concepts. Can it accurately and repeatedly understand that "the foot attaches to the shin bone, which attaches to the thigh bone, which attaches to the hip bone...", and that these have certain degrees of freedom, but not others, and that one foot goes in front of the other, and to easily and reliably distinguish a normal walk from a silly walk . . .
Yes, these are different levels of abstraction, especially the last one, and they need to be very accurate to even reach a young child's level of understanding, and this is just one branch of a branch of a branch in the entire fractal pattern of understanding that is necessary for a more general intelligence.
Once that is in place, and it can show evidence that it can model it's own mind, then it might be able to model someone else's mind.
While the statistical 'abstraction' and remixing seen in these "AI" systems is sometimes impressive and useful, it is frequently revealed that there is utterly no conceptual understanding beneath it. It is merely a statistical re-mixer abstracting patterns of words that occur near other words, remixing them and filtering for grammatical output.
It hasn't got a theory of anything, nevermind a theory of mind.
"statistical re-mixer" doesn't describe these systems very well. I see this complaint a lot, that supposedly DL models can only manipulate existing content without creating anything of their own. That's just false, unless your standard for originality is so high that humans can't reach it either.
These models that have hundreds of billions of "synapses", it's not very shocking to me that they can learn the abstract form of concepts. In fact, it's kind of beautiful that human concepts have this mathematical nature. It vindicates Plato, and disappoints everyone who has claimed that language and meaning is arbitrary.
But the main issue here is that for every conceivable empirical test we can perform, you'll still make the same complaint. Even after it's demonstrated better ToM abilities than you, by predicting and explaining other people's mental states better than you can, you'll say the same thing.
Maybe it's because you think that "understanding" requires not just accuracy, but having a certain kind of inner experience that a human could relate to.
Yes, I understand that it can appear to synthesize something new, and no, I'm not looking for some inner experience.
I'm looking for it to show an ability to wield not only a set of strings (with language associations), but something actually like the platonic ideals - objects, with properties and relations.
A few errors show quickly there is no such concept being weilded.
>> I saw a fine example of this failure the other day: "Mike's mom has four kids. three are named Danielle, Liam, and Kelly. What is the fourth kid's name?" ChatGPT's reply is explanation of how there isn't enough info in question to tell. Told "The answer is in the question.", ChatGPT just doubles down on the answer. (Sorry, couldn't find the original example)
>> "My sister was half my age when I was six years old. I'm now 60 years old. How old is my sister?" ChatGPT: "Your sister is now 30 years old". [0]
>> Or this one where ChatGPT entirely fails to understand order/sequence of events. [1]
Or a plethora of math problem fails found...
Similarly, the image "AI"s fail to understand relationships between objects (or parts of one object), and cannot abstract a particular person's image from a photo, showing it has no understanding of what is a body... (I can look those up if necessary).
And, of course, the answers are entirely untethered from reality - it is completely by chance whether the answer is correct or just wrong. It is run through a grammatical filter/generator at the end so it's usually grammatical, but no sort of truth filter (or ethical filter for that matter either).
I don't expect some abstract experience, I expect it to be able to break down it's work into fundamental abstract concepts and then construct an answer, and this it cannot do, or it would not be making these kinds of errors.
> A few errors show quickly there is no such concept being weilded
I would have given similar examples to show that ChatGPT makes the same kinds of mistakes that humans do. The first one is good, because ChatGPT can solve it easily when you present it as a riddle rather than being a genuine question. Humans use context and framing in the same way; I'm sure you've heard of the Wason selection task:
https://en.wikipedia.org/wiki/Wason_selection_task
When posed as a logic problem, few people can solve it. But when framed in social terms, it becomes apparently simple. This shows how humans aren't using fundamental abstract concepts here, but rather heuristics and contextual information.
The second example you give is even better. It's designed to trick the reader into thinking of the number 30 by putting the phrase "half my age" before the number 60. It's using context as obfuscation. In this case, showing ChatGPT an analogous problem with different wording lets it see how to solve the first problem. You might even say it's able to notice the fundamental abstract concepts that both problems share.
The third problem is also a good example, but for the wrong reason: I can't solve it either. If you had spoken it to me slowly five times in a row, I doubt I could have given the right answer. If you gave me a pencil and paper, I could work through the steps one by one in a mechanical way... but solving it mentally? Impossible for me.
> It is run through a grammatical filter/generator at the end so it's usually grammatical, but no sort of truth filter (or ethical filter for that matter either).
I kind of thought it did get censored by a sort of "ethical filter" (very poorly, obviously), and also I wasn't aware of it needing grammatical assistance. Do you remember where you heard this?
But comparing 1 human to 1 GPT is mistaken to begin with. It's like comparing 1 human with 1 Wernicke's area or 1 angular gyrus. If you had 100 different ChatGPTs, each optimized for a different task and able to communicate with each other, then you'd have something more similar to the human brain.
>>trick the reader into thinking of the number 30 by putting the phrase "half my age" before the number 60
Yet it is exactly the process of conceptualizing "half" and applying it to "at six years old" instead of "of 60" that is the key to solving it.
These things aren't abstracting out any concepts, they only operate at the level of "being fooled by" semantics. The fact that humans sometimes fail this way gives us little more than [sure a human not really thinking about the problem may offer a bad solution based only on the superficial semantic]. ChatGPT reliably gives us the error based on the superficial semantics.
>>If you had 100 different ChatGPTs, each optimized for a different task and able to communicate with each other, then you'd have something more similar to the human brain.
YES, that is the route we need to go to get towards actual intelligent processing. Taking 100 of these tuned for different areas, and abstracting out the various entities and relationships.
Kind of like the visual cortex model that extracts out edges, motion, etc., and then higher areas in the visual cortex, combined with other areas of the brain allow us to sort out faces, bodies, objects passing behind each other, the fact that Alice entered the room before Bob, and that this is because Bob was polite...
They also mut know when they are making errors, and NONE of these systems comes even close — they happily spout their bullshirt as confidently as any fact.
I gave a deposition in a legal case where the deposing attnys used an "AI" transcription system. Where a human would ask if anything was unclear, and always at the next break get proper spellings of all names, addresses, etc., this thing just went merrily along inserting whatever seemed most likely in the slot. Entire meanings of sentences were reversed (e.g., "you have a problem" edited to "I have a problem"), names were substituted (e.g., the common "Jack Kennedy" replaced "John Kemeny").
There's the Stable Diffusion error with a bikini-clad girl sitting on a boat, where we see her head and torso facing us, as well as her butt cheeks, with thighs & knees facing away. It looks great for about 1.5 sec. until you see the error that NO human would make (except as a joke).
The mere fact that some humans can sometimes make superficial errors which resemble the superficial errors these "AI" things frequently and consistently make does not mean that because humans often have a deeper mode, these "AI"s must also have a deeper understanding.
It means either nothing, i.e., insufficient data to decide, or that these are indeed different, because there is zero evidence of deeper understanding in a ChatGPT or Stable Diffusion.
You might like some of the work being done under the label "Factored Cognition". It's an approach that treats LLMs as building blocks instead of being complete AIs. Instead of asking the LM to solve a problem directly in one pass, you ask it to divide the problem between several different virtual copies of itself, which then themselves subdivide further, and so on until each subtask is small enough that the LM can solve it directly. For this to work the original problem needs to be acyclic and fairly tree-like, i.e., not something that requires having a sudden "Eureka!" moment to solve.
But I've only seen this done with a single model. Sometimes it gets prompted to act like a different agent in different contexts, or given API access to external tools, but it's still just one set of weights.
Hmm, that sounds like a nod in the right direction, but a rapid initial skim maybe indicates that it's more parallelizing the problem than abstracting it. I've got to read more about it - thanks!
While Minsky & Papert's book on Perceptrons was enormously destructive, I think there is something to their general concept of Society Of Mind, that multiple sub-calculating 'agents' collude to actually produce real cognition.
We aren't doing conscious reasoning about the edges detected in the first couple layers of our visual cortex (which we can't really even access, 'tho I think Picasso maybe could). We're doing reasoning about the concepts of the people or objects or abstract concepts or whatever many layers up. The first layers are highly parallel - different parts of the retina connecting to different parts of the visual cortex, and then starting to abstract out edges, zones, motion, etc. and then synthesize objects, people, etc.
I think we need to take a GPT and a Stable Diffusion and some yet-to-be-built 3D spatial machine learning/reasoning engine, and start combining them, then adding more layer(s) synthesizing about that, and maybe that'll get closer to reasoning...
ChatGPT3 does not even have a theory of physical objects and their relations, nevermind a theory of mind.
This merely shows that an often useful synthesis of phrases statistically likely to occur in a given context and grammar-checked, will fool people some of the time, and a better statistical model will fool more people more of the time.
We can figure out from first principles that it has none of the elements of understanding or reasoning that can produce a theory of mind, any more than the Eliza program did in 1966. So, when it appears to do so, it is demonstrating a flaw in the tests or the assumptions behind the tests. Discouraging that the researchers are so eager to run in the opposite direction; if there is confusion at this level, the general populace has no hope of figuring out what is going on here.