AI megathread

[Laird]

Am I right to interpret that "like" as an endorsement of that operational definition of intelligence-without-consciousness, @Sci?

I think this discussion on whether AI can ever become consciousnes is interesting even though of course the argument presented meets fierce resistance:

Alexander Lerchner, a Senior Staff Scientist at Google DeepMind, has published a paper arguing that no amount of scaling will make an AI model conscious. Not in ten years, not in a hundred.

…


Here is the argument. Computation is not a thing that exists in physics on its own. It is a description we lay over physical processes, mapping continuous voltages and transistor states onto discrete symbols we call meaningful. The mapping requires what Lerchner calls a "mapmaker," an experiencing agent who decides that this pattern of voltages counts as the number 7 and that one counts as the letter A. Without that agent, the computer is just physics. The symbols are not in the hardware. They are in us

https://aihola.com/article/deepmind-abst...sciousness

Am I right to interpret that "like" as an endorsement of that operational definition of intelligence-without-consciousness, @Sci?

I think it depends on what you mean by intelligence still.

Is a more basic proof assistant program intelligent? Some basic game AI for, say, chess?

I dislike the term "emergence" because to me it suggests some kind of "something from nothing" event.

OTOH, if we see the potential for intelligence is in all matter that can be purposed toward the actions of a Turing Machine...where is the intelligence? In the program? The Universe entire?

[Laird]

I think it depends on what you mean by intelligence still.

The definition I provided stipulates what I mean (operationally). I agree that "natural" intelligence is exercised consciously or at least by a mind. We're talking about "artificial" intelligence though. It also seems likely that the basis of the intelligence of these LLMs is different to that of the intelligence of us humans, perhaps quite radically different.

What more meaning, and what more qualifications, do you want?

Is a more basic proof assistant program intelligent?

Not in my view; it's too narrow and rigid in an algorithmic sense.

Some basic game AI for, say, chess?

Ditto, although that's closer to meeting the bar.

I dislike the term "emergence" because to me it suggests some kind of "something from nothing" event.

I don't mean it in the sense that a new ontological entity is generated. I mean it in the sense in which new behaviour and functionality are generated, i.e., when not expected nor predicted given the rules which generated them. In other words, I mean it in the sense of the emergence of complex patterns given the simple rules of the Game of Life.

OTOH, if we see the potential for intelligence is in all matter that can be purposed toward the actions of a Turing Machine...where is the intelligence? In the program?

Yes.

[Laird]

Another like-but-don't-reply. It seems that you're going to make me work for this, @Sci! Well, OK then, here's that explicit question again, updated:

Now do you endorse that operational definition of intelligence-without-consciousness?

Another like-but-don't-reply. It seems that you're going to make me work for this, @Sci! Well, OK then, here's that explicit question again, updated:

Now do you endorse that operational definition of intelligence-without-consciousness?

Are the patterns that end up produced by the Game of Life genuinely unpredictable? AFAIK the program is not probabilistic but "deterministic" in the computational sense of following a set of algorithmic steps in order.

I guess it's unclear to me what you think is the intelligence in an LLM that is beyond game AI but also beyond theorem provers that existed in the past?

I'm fine with saying the best chess AI is intelligent, if all that means is following some set of algorithmic steps in response to human decisions in such a way as to viably compete with humans in the program's domain space.

But you seem to be saying the LLM is something beyond this, and that is where I am confused.

[Laird]

Are the patterns that end up produced by the Game of Life genuinely unpredictable?

No, they're predictable in principle, because, as you say:

AFAIK the program is not probabilistic but "deterministic" in the computational sense of following a set of algorithmic steps in order.

I meant unpredictable in the sense of (i.e., to augment and extend the meaning of) "unexpected", the other adjective I used.

I guess it's unclear to me what you think is the intelligence in an LLM that is beyond game AI but also beyond theorem provers that existed in the past?

It's simply a matter of degree and scope. At a certain point (not exact, of course), the degree and scope of intelligence becomes astonishing, especially when it is for most purposes - given the same interface - indistinguishable from human intelligence, and in some respects even already superior to at least the average human. Sometimes it can even perform at expert human level, e.g., the solving of the maths problem that restarted this exchange when I posted about it. That's why I'm puzzled that you're not astonished.

I'm fine with saying the best chess AI is intelligent, if all that means is following some set of algorithmic steps in response to human decisions in such a way as to viably compete with humans in the program's domain space.

But you seem to be saying the LLM is something beyond this, and that is where I am confused.

The intelligence of LLMs is different because:

1. It's more human-like: it can communicate in our languages; it can to some extent "think" like we do (see chain-of-thought prompting); when augmented it can "see" and "hear" like we do, especially so as to transcribe text from audio and images; etc.
   
2. It's therefore more generally-applicable (with respect to "scope" above) and a likely/potential candidate for AGI, at least with future refined/augmented algorithmic techniques.
   
3. It's emergent in the above sense of unexpected and unpredictable: chess algorithms are designed and coded to do exactly what they do; LLMs were designed to predict the next word, not to reason, solve maths and logic problems, and code. Those abilities "came along for the ride" without it being known in advance that they would, and weren't built in explicitly like playing chess is built into chess AI.
   

Model Collapse Ends AI Hype by Theos Theory

Abstract: Do Large Language Models (LLMs) think and reason? Are they perpetual information machines, producing endless coherent and correct text from finite training data? We explore how LLMs work and whether they produce rational thought and endless information. We show how theoretical considerations and experimental results from philosophy, statistics, information theory, and machine learning argue against the thesis that LLMs are rational, information-generating entities.

While the output from LLMs is very impressive it’s important (for the sake of ones sanity) to keep in mind that they are just “functions”. The context is it’s arguments and the parameters/weights of the neutal net the definition of the function.

Chain-of-thought looks like genuine reason but in reality it’s a trick where the llm iteratively generates more context on top of the existing user provided context (it enriches the problem you could say).

It’s very far from human intelligence if this is the benchmark for agi. With that said I believe it’s only a matter of time before there will be world models that can recursively self-improve from continuous input. It’s certainly not possible to disprove we can build such thing.

While the output from LLMs is very impressive it’s important (for the sake of ones sanity) to keep in mind that they are just “functions”. The context is it’s arguments and the parameters/weights of the neutal net the definition of the function.

Chain-of-thought looks like genuine reason but in reality it’s a trick where the llm iteratively generates more context on top of the existing user provided context (it enriches the problem you could say).

It’s very far from human intelligence if this is the benchmark for agi. With that said I believe it’s only a matter of time before there will be world models that can recursively self-improve from continuous input. It’s certainly not possible to disprove we can build such thing.

This is something we seem to agree on. I am happy we have common ground in some regard.

However... I am not sure that a world model is even remotely possible with how current "AI" fundamentally functions.

To compare, we conscious beings do not start from models ~ we start from raw experience, and then conceptualize abstract objects by which we identity similar elements in experience. That is, we form the concept of a chair, and then we build upon that when we encounter other objects that fit our concept of what a chair is. So... what might be a chair to you might not be a chair to me, irrespective of whether we agree on a generalized definition.