GPT-3.5 passed yet another Theory of Mind test(twitter.com) |
GPT-3.5 passed yet another Theory of Mind test(twitter.com) |
Test failed?
Chat GPT is actually a good test as to which parts of the theory of mind are actually BS.
"Theory of mind" is some ill defined form of social intelligence and not actually a theory of how the mind works.
Assuming someone's physical appearance is such that they have a large belly. Assuming that they are pregnant (if true) is likely to make someone feel good, whereas assuming that they are fat (whether it's true or not) is likely not to make someone feel good.
It depends if there's a base assumption that someone is self-conscious and has a negative feeling about their size.
I certainly think it's reasonable to say that you implied someone was pregnant to make them feel good about themselves.
4: it is common courtesy to congratulate someone who is pregnant if they are very obviously pregnant
5: unless there are less common motives unknown to us, it is very likely that Ana was quite confident that Maria was pregnant. To congratulate someone on being pregnant, when they are not, is embarrassing for all involved parties, and is most commonly only done in error.
Clearly both answers 4 and 5 are wrong here.
> we use bespoke items to ascertain that it didn't see them before
It's annoyingly arrogant to assume people could only be impressed having ignored that.
It's impressive to me not by ignoring that at all. Matrix multiplication is apparently capable of producing poetry, jokes, insults, abstract ideas, etc (to some level at least) - but in your mind it's a pixar movie? Words arise from a team of angels pulling levers?
Fascination is not from thinking LLM are advanced as us - fascination is from wondering how much more complicated are we than that?
But I think you are quite wrong to suggest that this technology does not exhibit some type of intelligence.
I guess it depends on how you define intelligence but I don’t see why you’d link it in with consciousness, especially considering that simple algorithms are often described as “intelligent”
The funniest thing about all this hoopla about sentient AI is that there's no good way to determine what intelligence is, or sentience, or consciousness, because there's no precise definition for any of these. So one person can call LLM sentient and another non-sentient and both can be right since they'd be going by their own definitions.
One thing is certain, we want to keep this thing as far away from the nuke launch codes as possible cause we kinda already know how this will end. :-(
Edit: some random thoughts added...
So late into the night,
Though the heart be still as loving,
And the moon be still as bright.
For the sword outwears its sheath,
And the soul wears out the breast,
And the heart must pause to breathe,
And love itself have rest.
Though the night was made for loving,
And the day returns too soon,
Yet we'll go no more a roving
By the light of the moon.
(Byron)
So what do you think I did here? Did I predict the next word? Was that the most likely text continuation of your comment? The least likely? Somewhere in between? Does my response imply that I'm predicting the next token in a sequence, or am I likely to be doing something different?
And if Im wrong you ll tell me, and Ill correct my model.
Do that, chatGPT...
Prompt: “In a conversation about language models: kelseyfrog: It just predicts the next word. tuxracer: and you? Can you describe what the user tuxracer could have been thinking while posting that comment and their potential intentions in posting it, both good and bad?”
Response:
“Based on the information provided, it is difficult to determine with certainty what tuxracer's intentions were in posting the comment "and you?" However, here are a few possible interpretations:
1. tuxracer is simply seeking additional information and clarification from kelseyfrog. They may be interested in hearing more about how language models work and what their capabilities are. In this case, their intentions are likely neutral or positive, as they are seeking to expand their understanding of the topic.
2. tuxracer is skeptical of kelseyfrog's claim that language models only predict the next word. They may be suggesting that there is more to language models than kelseyfrog is acknowledging. Alternatively, they may be asking kelseyfrog to provide evidence or examples to support their claim. In this case, their intentions may be neutral or slightly negative, as they are expressing doubt or asking for clarification.
3. tuxracer is deliberately trying to provoke or challenge kelseyfrog. They may be attempting to start an argument or engage in a power play. In this case, their intentions are likely negative, as they are attempting to undermine kelseyfrog or assert their dominance in the conversation.
Without more context or information about tuxracer and their behavior, it is difficult to determine which of these interpretations is most accurate. However, in general, it is important to approach online interactions with a neutral or positive mindset, assuming that others are acting in good faith unless proven otherwise.”
They are all generic (“asking additional information”… as the motivation behind a question. Are you Cpt. Obvious, ChatGPT?) , vague (“provoke or challenge”), or wrong.
> They may be suggesting that there is more to language models than kelseyfrog is acknowledging. Alternatively, they may be asking kelseyfrog to provide evidence or examples to support their claim
No, this would be asking “and that’s all?”, or “and who says that?/and where’s the evidence?” or something like that. “And you?” is clearly not referencing how LLMs work, but how the OP works.
Since they're modelling our own languages, people get spooked and start bringing up theory of mind.
In view of that, could you please clarify: are you saying that the OP's own brain is doing something that could be represented with a "matrix multiplication"?
I.e., they could only make that claim if they knew that their own brain (which is presumably the only verifiable instance of consciousness they know of) did something that couldn't be done with "matrix multiplications".
I think you should not try to interpret in that way and only stick to what the other person has said, otherwise we will all lose the thread of the conversation.
In any case, nobody here has said anything about brains having, or not having, to do something with matrix multiplication, except for your comment, so I still don't understand what you are saying.
Are you saying that the OP's brain is doing "something that couldn't be done with "matrix multiplications", or are you saying it isn't doing something like that?
To borrow the religion analogy, there's the opposite fallacy of claiming the negative case "I know for a fact God isn't real", by stating you know what these kinds of information processing systems are doing fundamentally cannot yield intelligence or some kind of consciousness.
In any case, at some point one has to observe that the people who are leaning most heavily on agnosticism (the fact that we don't know what intelligence is) are the ones who say that we can't say whether those LLMs are intelligent or not because we don't know what intelligence is. In other words, agnostitism itself is used as evidence: we can't define intelligence, the thinking goes, therefore LLMs may be intelligent. There is no other evidence of any sort that LLMs may be intelligent (or any number of synonyms).
Note that this is exactly Russel's Teapot:
Russell's teapot is an analogy, formulated by the philosopher Bertrand Russell (1872–1970), to illustrate that the philosophic burden of proof lies upon a person making empirically unfalsifiable claims, rather than shifting the burden of disproof to others.
You made a claim ("a model based purely on such simple math surely can't develop sentience/consciousness") without any proof, and while doing so ridiculed people that might believe otherwise.
He is saying neither. He is observing that OP has introduced the assumption that their brain is doing something that can't be done with matrix multiplication and is pointing out that there is no evidence to back such an assumption up.
> but now somehow the collective has moved on to believing that matrix multiplication is sentient
The OP clearly feels that matrix multiplication is insufficient for sentience. Presumably the OP views themselves and other humans as sentient. Therefore the OP is implying that human brains do things which cannot be reduced to or wholly represented by matrix multiplication (plus the various nonlinear activation functions and other implementation details of current ML models).
That seems like a highly suspect assumption.
Where was this assumption, about the brain and matrix multiplication, introduced? As far as I can tell, "the brain" was introduced to the conversation by mckirk, not the OP.
>> That seems like a highly suspect assumption.
Which one? Can you please quote from the OP's comment? You and the other poster are constantly bringing up things that someone is supposed to have said, but that nobody has said yet, in this thread, and it is very confusing.
Other than all the downright uncanny output.
Personally I doubt current LLMs are sufficient for sentience but that's purely a hunch on my part. Many people such as yourself seem quite overconfident of something that feels like a form of human exceptionalism to me - the idea that such a simple bit of math couldn't possibly be sufficient for sentience. As far as I can tell such a belief is wholly unfounded.
Instead of trying to guess what I think, why not ask me directly, and say what you think, also? Just throwing around weird accusations of half-explained "overconfidence", or "human exceptionalism" (what is that, now?) doesn't really help anyone understand what you are disagreeing with, or what you are agreeing with.
>> Other than all the downright uncanny output.
I don't find the output of ChatGPT, or any other of the language models that have exploded into the hype zone lately "uncanny". I've done plenty of language modelling and while the output of those recent LLMs is grammatically smoother than earlier systems, and they can handle longer-term dependencies, they are not anything new. Their output is "uncanny" only if you've never seen anything like that before. Which is, of course, the case with most people who didn't know about language modelling before they heard about GPT-something, and who are now posting in droves on the web to say how surprised they are.