Like we don't say that planes artificially fly, they just fly. I mean technically it's non-natural flight, but it's interesting to think about.
Whether it was generated from "natural" or "artificial" means (a distinction which quickly breaks down anyway) the intelligence is just intelligence.
Look up the definition of "artificial" while you're at it.
- If you mean "you cannot EXPLAIN AGI/consciousness/intelligence if you don't understand it", then that's true, but it's a trivial tautology.
- If you mean "you cannot DEVELOP AGI/consciousness/intelligence if you don't understand it" then that's very debatable.
Historically we have been able to develop all kinds of things, despite not knowing how they work. Tinkering and trial and error is often enough.
After all that's how evolution solved the problem of creating consciousness/intelligence. There wasn't some entity that "understood" intelligence that created it.
Sort of how we didn't have a "clear definition" of most things for millenia, but we still were able to recognize them as a class of thing. It's more of a "I'll know it when I see it" kind of thing.
And the external behaviors (of consciousness and intelligence) matter more than "but is it really conscious/intelligent inside" when considering some AGI as such. After all we neither can clearly define, not we know or can measure what's going on inside another person's head regarding consciousness, or to be frank, not even on our own head. When it comes to us, we just have a subjective experience, and not even a very clear one at that.
Well that's not true. We tamed fire before understanding combustion, friction, heat, or anything else.
To crank it up a notch, the assigned task could involve generating itself subtasks which are handled in the same manner. This subtask generation could start to look a bit like will/intentionality.
Now consider if the top-level task is something like maximizing money in a bank account that pays for the compute to keep it running :)
(IMO this is still missing some key pieces around an emotion-like system for modulating intensities of actions and responses based on circumstantial conditions—but the basic structure is kinda there...)
I'm not sure that's as clear cut as you make it sound.
For example curiocity could be a fine engine for motivation as well, even if you don't care if you'll survive or not.
For starters, because the claim argued is that you can have motivation (say, through curiocity) even without a survival instinct.
And it's perfectly possible to have no survival instinct and yet not to be starting anyway. It's enough that you have access to food or are fed, to avoid starvation. So lack of survival instict is not the same as starvation.
If we substitite "starvation" for "access to electricity" (as AGIs don't eat food), as long as an AGI is provided by electricity by us, even if it has no survival instict, can still have curiosity, and thus motivation.
It's just hard coded, whereas GPT's is dictated. More or less anyway.
Also our "fitness function" or motivations & goals aren't even that hard coded. You can easily modify them through drugs.
That's my point. It's raising some arbitrary human trait to special status, when it really isn't. Human goals are set by some external process too - evolution. And they aren't really even intrinsically fixed. They can be modified through drugs.
I think the strongest demonstration of that is post orgasm clarity (if you're a man anyway). Your whole motivation changes in an instant.
most brains I run into don't do that much at all, mostly just existing and adaptation-execution
Yep. Elevators used to be proof that machines can think. Then compilers, and chess, and go, and search, and …
The problem with AI is that as soon as it works, we stop thinking about it as “artificial intelligence” and it becomes “just automation”. Then AI moves to the next goalpost.
At some level, intelligence requires logic, rationality, and conceptualization, all of which are topics which have evaded clear definition despite millennia of philosophy directly addressing the issues.
To the point underneath, humans do not answer in as predictable a way as ChatGPT. Your answer, for example, I am confident does not come from ChatGPT.
Edit: if I've horribly mangled the Turing test definition, please let me know
Turing test is easy, I had 2 chat bots talk about other users in the channel while besides some trigger words ignoring what those other users had to say. The human subjects got angry successfully which means it was important to them.
You’re demanding this because you aren’t comfortable with the implication that a computer can pass our existing tests for intelligence, so you rationalize that with the comforting thought that those tests were not meant to identify intelligence. Tests like the SAT or the bar exam or AP English. Or tests for theory of mind or common sense or logic. Those tests aren’t testing for ‘intelligence’ - they can’t be. Because a computer passed them.
It’s okay. We can make new tests.
If we took this comment at face value we're ending up with a definition of "think" that can't reason, play games or recall information - or it would be outdone by machines. Thinking obviously isn't very important!
This is not obvious in the least. Thinking was required to produce the thing that could achieve the outcome that was defined by thought.
I could just as easily say that AI, as currently implemented, is really just another muscle.
Did they? When? By who?
I suspect even full AGI will be considered “just a machine” for many decades, even centuries, before it gains the same rights as humans. We love to find reasons we’re special. Look how long it took us to admit animals are intelligent.
For many humans, computers definitionally can’t be intelligent. It’s important to recognize that.
And all of our surprising wins and awful mistakes had explainable reasons, dammit; it wasn’t just a misfiring of trained statistical networks!
It is only if we narrow thought to mean precisely human-like thought when humans and human creations are uniquely capable of something. To that extent, our models of intelligence is very much in the pre-copernican era.
Slime molds and amebas might be able to reason abstractly to some degree but they can't write code or poetry.
How do you know fungal networks don’t write and read poetry under the forest floor? If they do—and we have no reason to doubt that they do—you wouldn’t be able to read them, let alone understand them.
The earth’s biosphere as a whole also writes code, just in DNA as opposed to on silicon transistors, why exclude the earth’s biosphere from things capable of abstract thought?
yes, that is the sense in which we are discussing intelligence in order to debate whether the human brain and LLMs operate on similar phenomena
This would be like finding a flower which produces a unique fragrance, then create a perfume which approaches the same fragrance and then conclude that since these are the only two things in the universe which can create this fragrance there must be something special about that perfume.
> is an uninteresting fact if we define “abstract reasoning” to be exactly what humans do, and then create models with the goal of recreating exactly that
if you find this uninteresting, we have perhaps an irreconcilably differing view of things
I’m not saying these language models—or my hypothetical perfume—aren’t an amazing feat of technology, however neither has any deep philosophical implications about shared properties other than the ones constructed to do so. Meaning, even if LLMs and humans are the only two things in the universe that can reason abstractly in the same way humans do, that doesn’t mean the latter has any more properties shared with the former.
For an example, crows can effectively use tools and communicate abstract concepts to one another from a memory. Which means they can observe a situation, draw conclusions, and use those conclusions to act as well as make decisions on how to act. That would seem to meet the bar for reasoning abstractly.
Of course… The hole in that theory is that evolution never found the wheel.
The steel man in that theory is that it invented the neurological and social processes that then went on to invent the wheel. And the platypus and the clap.
Edit: I forgot to bring it back around and make a point ;)
I’m saying that humans invented clocks and CPUs. We only have metaphors that havs emerged from the still misunderstood ether of the informatic universe.
To solve ourselves is to know ourselves completely, and to know ourselves completely is to be honest in who and why we are what we are simultaneously across all persons. It assumes perfect knowledge.
There is no statistical approximation nor computational power which can do this.
> People are understandably desperate to understand their experiences as more than an encoding of a thing that might be explained.
Another way to frame this is, "some people are nihilists and do not see life as more than an encoding of a thing that might be explained."
To know anything (an X) completely you need perfect knowledge. Hence people come up with a set of simplified ways of reasoning about X. They call this a model of X.
Model is incomplete and so primitive that we manage to play it forward/backward in our heads or our computers.
If a model checks out with the real outcomes we proudly exclaim that we understand X.
I'm not being sarcastic, that is just a real method we use all the time.
Agreed.
> Hence people come up with a set of simplified ways of reasoning about X. They call this a model of X.
A "set of simplified ways of reasoning about X" in order to create a "model of X" does not imply complete understanding. Quite the contrary actually.
To wit, science often models current understanding of a phenomenon. When new evidence (understanding) is discovered, the model is updated to account for it. Sometimes this invalidates the original model, often the model is refined. Either way, progress is made with the tacit agreement that the model may change in the future.
> If a model checks out with the real outcomes we proudly exclaim that we understand X.
Again, this does not support the assertion of "That's been solved long ago." If anything, it affirms there is justification for disagreeing with the original premise to which I responded:
> We just can’t accept that we might solve ourselves.
The LLM value add for coding is less than the value add of syntax highlighting in my experience.
In my experience and it's certainly useful for unit tests speeding things up for me pretty dramatically.
Also when working on a new development language was quick to point out "how you do that" with the inline vars I needed manipulated vs look ups in Google and copy pasting.
That’s still the hole in my theory of consciousness. I admit it ;)
But it doesn’t give me as much cognitive dissonance as others to believe that the process of performing moral actions still has to be “processed” by me, a processor. In some sense.
You could say the same for a character recognition system.
how am i supposed to have a conversation about someone who is gassing up "not perfect [arithmetic]" (something a wrist watch from the 80's can do) and won't even believe what the creators of said machine say about how it works
Prompt ChatGPT Actual Match
397,356 * 930,547 369,685,207,932 369,758,433,732 FALSE
36,330 * 26,951 979,458,630 979,129,830 FALSE
8,681 * 9,330 80,911,430 80,993,730 FALSE
278 * 903 250,734 251,034 FALSE
82 * 77 6,314 6,314 TRUE
I have a theory that it does arithmetic badly because the logic goes right-to-left, when LLMs write left-to-right. If the digits were to be reversed, it might not make as many mistakes. I ran out of attempts before I could test this properly.
GPT’s attention window is not equivalent to it being able to ‘write stuff down’ - it’s its equivalent of being able to remember the foregoing few minutes of conversation.
> Human: what is 36,330 26,951 * (input is expressed in reverse post fix notation)
> chatGPT: To evaluate this expression using reverse Polish notation, we need to use a stack to keep track of the intermediate results. Here's how we can evaluate the expression:
1. Read the first number "36,330" and push it onto the stack. 2. Read the second number "26,951" and push it onto the stack. 3. Read the operator "", pop the top two numbers (26,951 and 36,330) from the stack, multiply them, and push the result (971,914,830) onto the stack. 4. The stack now contains only one number, which is the final result.
Therefore, 36,330 26,951 = 971,914,830 in reverse Polish notation.
I dont think this is about notation.
I also tried the show your work methods. I will not paste my conversations here as they are so long but even with showing it's work it doesn't give the right answer. Two funny things I noticed
1. It either uses a completely wrong method to calculate it, or it shows the right "kind" of work and then gets the wrong answer
2. If I tell it it was wrong and it will just agree with me. I multiply two huge numbers and tell it no the answer is actually 42, it will just apologize. Then when I ask the reason why it thinks it's right it will give the most bullshit things lol. Once it even said "I read the number wrong, the second number should be x I read y". The thing is if I had actually given x in the input it would have gotten the answer right.
It's all very confusing.
What's also interesting, though not necessarily surprising, is how relatively close the incorrect answers are in your examples. It's not like the model will randomly spit out an answer like 42 if you ask it to multiply something like 36,330 by 26,951.
You did not. You have to explicitly select it from the dropdown which is only available on paid Plus accounts.
In fact it is damage to the prefrontal cortex (which has nothing to do with speech) which is mostly correlated with a detriment in intelligent behavior (suspiciously also social behavior; a food for though in what we consider “intelligence”). Victims of lobotomy had their prefrontal cortex destroyed, and their injuries resulted in them loosing their personalities and loosing basic function as human beings, even though they were still able (but perhaps not always willing) to speak and comprehend speech.
What you have that LLMs lack is a visual part of your brain - one which can instantly count quantities of objects up to about 7. That gives you tools that can be trained to do basic arithmetic operations. Although you have to be taught how to use that natural capability in your brain to solve arithmetic problems.
And of course for more complex things than simple arithmetic, you fall back on verbalized reasoning and association of facts (like multiplication tables) - which an LLM is capable of doing too.
Poor GPT though has only a one dimensional perceptual space - tokens and their embedding from start to end of its attention window - although who’s to say it doesn’t have some sense for ‘quantity’ of repeated patterns in that space too?
That has been my best analogy so far.
They'll never say I don't know and bullshit you into oblivion while never backtracking.
However this is a great simplification, and borders on an absurd reduction. You can model our brains using linear algebra, however that doesn’t mean our brains are linear algebra computer. There is a whole lot more going on than neurons receiving feedbacks from other neurons which adjusts the weight for subsequent firing. A lot of our behavior is actually inherited (I know I spent a whole week here on HN arguing with IQ advocates on the nuance of that statement), neurochemicals and hormones add a whole another level of statefulness not seen in artificial neural networks, the brains ability to make computations is actually pretty limited (especially next to a GPU). I mean, cordiseps exists, meaning a fungus can infect an organic system and control its behavior, there is 100% chance that some yet to be discovered viral and bacterial agents, not sharing any of our DNA—and certainly not “connected” to the “weight matrix”—are also influencing our behavior (just not as dramatically), and there is 100% chance they interact with our DNA also controlling our “innate” behavior.
What is going on in our brains can only be modeled using statistical ordering of semantic words and actions. The real world brain is always going to be infinitely more complicated than this model.
And that's completely fine.
Didn't we dispose of Cartesian dualism a long time ago? It may have been useful for disposing of superstitions which preceded it, but anyone who has observed the effects of a stroke or has taken a psychotropics understands that mind and body are not made of separate stuff, and I know of no modern system of thought that stipulates that as an axiom.
LLMs just demonstrate that emergent information structures at the complexity of human language need not be only biological.
Not just that LLMs are very capable by themselves, but significantly improve speech and image models.
We’ll need to break out of the Chomsky hierarchies and develop some new theories of language.
Vocal communication isn’t unique to humans, a new theory should be more broadly applicable to non human and non vocal languages.
I’m excited to see what big questions will be answered, and what new questions arise.
We might need to develop some "new theories of language", covering other areas, but why exactly do you think we'd need to "break out of the Chomsky hierarchies"?
Nothing about LLMs challenges Chomsky hierarchies.
They do challenge Chomsky's ideas about language being an innate human quality, but that's about it. Besides that's not related to the Chomsky hierarchy of language, e.g. regular, context-free and so on. Those will continue to remain a hierarchy describing languages based on capability levels, whether there's AI, and AGI, or not.
You can't tell me with a straight face that actual spoken language fits well in that hierarchy. It seems we can shoehorn it in with a 100B parameter turing machine definition, but that just makes my point that the constructed language hierarchy is not well suited for describing spoken languages if it takes such a large definition.
We should be able to find something like a new probabilistic theory of language that will do a much better job of describing spoken languages. Such a theory could help explain some LLM behaviors we see, and apply generally to other forms of noisy communication.
Humans by contrast use about 700T synapses to support our own understanding of language.
I believe most human intelligence is based on language. We are smart because language gives us the keys to solving all sorts of situations, based on the experience of prior generations. We are wondering if GPT architecture can generate real understanding, but forgetting the training corpus. The language corpus is the real magic here. The source and accumulation point of intelligence.
Take a baby and give it access to language, you get a modern human. Take a random init neural net, use language corpus for training, you get GPT4. It doesn't matter what the model is - brain or transformer, it doesn't matter if it's a transformer or just a RNN (RWKV). What matters is the training signal.
We should not be wondering how come computers got so smart now, we should be wondering how smart is Language. Language is a separate evolutionary system of ideas and concepts. It is a self replicator, like DNA.
But you can’t say information theory as it stands gives us a completely satisfactory model of human language. Something will need to be built on top of it.
That's highly debatable. For starters, who said there's no creativity or inspiration in humans, or for that matter, that there can't be in a complex A.I.?
How we achieve that creativity or inspiration is irreleant, as long the entity (human or AI) showcases creativity and inspiration.
Nor is it much clear why all being "only (a very complex) agency" would preclude creativity and inspiration.
That's like a worse version of "a human can't be creative or have feelings because it's all a bunch of molecules".
There is such a thing as emergent properties.
>And that's completely fine.
That's also highly debatable. I mean, that it would be "completely fine" if you were right and there wasn't "creativity of inspiration".
There has been a long history of reifying the behaviors of living things in general (e.g. vitalism) and humans in particular (e.g. dualism).
The success of LLMs challenges a lot of philosophy dealing with what behaviors are and are not possible in the absence of these categorical districtions.
I have had casual debates years ago, in which strong dualists asserted that the kind of creativity exhibited by today's LLMs is simply impossible. No doubt those folks are busy inventing "special philosophical creativity" that LLMs "aren't really doing," but they've lost credibility.
LLMs have demonstrated that there was never any need to invoke categorical districtions between human behavior and math-as-implemented-by-physics. The gap is closed, there is no more room for gods.
Or maybe we will create a god.
It is telling that you talk about all those things happening in "real time." Ask any super-regarded philosopher, from Plato to Wittgenstein (yes I'm excluding Dennet et al), and that would be quite the hoot to point out.
There may be a part of the brain that is modelled well by an LLM, but if so, there seem to also be parts that aren’t, or even existing “multimodal” models like GPT-4, which is more than an LLM.
Humans, including their brains (both in the narrow sense and in the broader sense of “everything that contributes to cognition”, which may extend beyond the brain proper) are machines, and their function probably will, someday, be mimicked by machines. But they are still more complex machines that modern “AI” systems
a machine, in typical usage, is an object designed and constructed by a deliberative intelligence for a purpose, which is what evolutionary theory states living beings aren't
the machines we currently conceive and manufacture also lack many structural features common to biology
in biology a single component may have many functions, some of which may interfere constructively or destructively, while in engineering it is often considered best practice for each part to do one thing
organisms self heal, subsist on/in an ecology of other organisms, and are resilient to informational errors
forms of computation resilient to bit level errors could be said to mimic biological antifragility, but they run on an extremely fragile physical substrate which requires constant tending (by organisms) to keep entropy at bay
even if our implements were as autonomous and resilient as we are, it still feels valuable to preserve a clear distinction between objects, which all agree may be understood solely as a means to an end, and subjects, who (in kantian form) are ends in themselves
this is not to say designed objects could not in principle become subjects, but to claim e.g. there is a clear path to sentience for deep learning models is for now speculative at best, and until we have good reason to believe a tool may develop subjective experience it can be a dangerous thing to personify it, and in any case we ought not compress ourselves into its image
So? I would imagine this would be orders of magnitude greater in a neural net, no?
Our minds are in fact same statistical models with a gradually declining ability to learn and driven by exogenous irrational goals to eat and mate.