What do you believe intelligence to be?
Simplifying things to this degree can show how you get from A to B.
We don't know what 70s AI and ChatGPT will evolve into. That's why everyone keeps debating and prognosticating about it, because nobody actually knows. But whatever people mean by AGI, we don’t know if it will or can evolve from the AI platforms that have come before.
We do know that the thing that evolved into ChatGPT is entirely different from primitive cellular life, so there's no reason to believe that ChatGPT will keep evolving into some kind of humanistic, self-aware intelligence.
I agree, and I think Goodman & Tenennaum [2] is a great place to see other things that may pop up in the road to AGI. LLMs are great, but they do too much at once. I think moving towards AGI requires some form of symbolic reasoning, possibly combined with LLMs, which may play the role of intuition and kitchen-sink memory.
You need something like 3 petabytes just to model the neurons. There are 100 trillion atoms in a neuron. And he said subatomic, so you need some significant factor of that, plus you have to save some state.
So 300 trillion petabytes+ might be a memory the size of the moon. I think relativity does matter at that scale. Also, you will probably have to mine silicon asteroids and invent an AI factory to make the memory, and power it with what, fusion?
So yeah, just model the brain at the subatomic level, bro, it's easy bro, and you're just a naysayer for saying it can't be done, bro.
But, please. If you have a solution to this, don't post it on HN, file a patent. Not just for the model, but for the revolutionary new dense memory, the location of silicon asteroids, the planetary memory factory, the power source for it, and so on.
We only seem “irrational” when we are talking about a narrow view of “rationality”, i.e., as defined by the cold hard logic of machines. We do not question why we have this definition of rationality. Our “irrationality” simply seems so because we have not bothered to understand the larger complexity of our evolutionary programming. It’s the same as not bothering to understand how a car works, and then claiming that the car works on magic. If one understands how it works, then it is no longer magic. In the case of humans, we may never fully understand how we work, but we can work towards a compassionate understanding of the same.
/rant
You can, as a human, map the 'wrong' response to external stimuli resulting in, say, your death (or the deaths of your progeny, for that matter)
If something seems irrational, that's because there are rational things layered on top of one another in a way that creates a misapprehension in a partially-informed observer.
Why? You assume nature must be comprehensible to us because you define it that way? Nature's not under any obligation to us.
> If something seems irrational, that's because there are rational things layered on top of one another in a way that creates a misapprehension in a partially-informed observer.
Possible irrational things: Collapse of the wave function, consciousness, why anything exists. There's also non-computable functions, paradoxes like the liar paradox, and the inability to refute various skeptical possibilities like living in a simulation.
Combine that with the fact that all models are wrong, some are just more useful. We can't model the entire universe with perfect accuracy without the model being the universe, which means there are always things left out. Our rational explanations are approximations.
We can't look at quantum mechanics and call it irrational. The way the universe works is by definition the right / correct / only way. If logic points one way and reality points another, the logic isn't taking everything into account.
>Rationality is something that makes sense for us
But I say:
Rationality is something that makes sense.
So nature cannot be irrational. For instance: we had the very logical Newtonian mechanics. But Mercury's orbit did not make sense. Later on Einstein saw another layer of logic that put both Newton and Mercury in more context, and now both make sense again.
When we see things now that make no sense, it indicates faults in our mental models, not in the things.
In the context of a person behaving or thinking "irrationally": their brain is just atoms and chemicals. It has to work right. If their behaviour and thoughts do not line up to reality, that's a feature of said brain.
Perhaps the irrationality is due to the brain giving too much weight to certain factors, or perhaps they are in fact behaving perfectly logically given what they know and you don't. Perhaps its workings can be improved in software, or in the next hardware generation.
But there's no spooky mechanism that can't in principle be modelled and implemented in a different architecture, ie, transistors.
But even if there are random things, are they thus irrational? Or just unpredictable? Is today's lottery number wrong in some way?
IMO, the two most important properties of human cognition are that it evolved through natural selection, and that it is embodied, meaning that it is part of a feedback loop between perception and action. I don't see why either of these things requires having a biological brain or body. It's true that current methods aren't close to achieving these things, but there's nothing in principle stopping us from creating evolved, embodied intelligences either in simulations or in robots.
Gradient descent is literally highschool calculus-level math.
In this case sounds like better algorithm + a lot of self-generated data (but no prior knowledge other than rules).
You've phrased the skeptic-not-contrarian case perfectly and succinctly.
I think chatgpt did show proof of what was considered at least until very recently « intelligence ». aka: understand enough about context and concepts to provide relevant answers to very complex and open questions.
For example, "squeeze a tube to cut water flow" and "put pressure on a deep wound to stop blood loss" can only be related, if not already in the training data, if there is understanding and intelligence.
The ability to do that is intelligence.
Otherwise, it's just a search and optimization problem.
Secondly, this poses the problem of 1) finding examples like the one you gave that it can't understand (regarding your specific example, see [1]), 2) ruling out that there was something "too close" drawing the equivalence in the training data, 3) ruling out that the failure to draw the equivalence is something structural (it can't have real understanding) rather than qualitative (it has real understanding, but it just isn't smart enough to understand the specific given problem)
So I'm back to my original question of how we would know if these are structurally different things in the first place.
[1] vidarh: How does squeezing a tube to cut water flow give you a hint as to what to do about a deep wound?
ChatGPT (GPT4): Squeezing a tube to cut off water flow demonstrates the basic principle of applying pressure to restrict or stop the flow of a fluid. Similarly, applying pressure to a deep wound can help control bleeding, which is a critical first aid measure when dealing with serious injuries.
[followed by a long list of what to do if encountering a deep wound]
It just seems to understand. This is useful, and deeply impressive, but it's not the same thing.
ChatGPT clearly needs reinforcement to know its limits the way humans are constantly reinforced from childhood that making confident claims about things we don't actually know well is often going to have negative consequences, but until we've seen the result of doing that, I don't think we have any foundation to say anything about whether ChatGPT's current willingness to make up answers means it's level of understanding is in any way fundamentally different from that of humans or not.
Those people are reasoning, however rightly or wrongly, based on that wrong information. LLMs are not.
Claiming we can tell that there's a distinction that merits saying people are reasoning and LLMs are not, is "hallucination" to me. It's making a claim there is insufficient evidence to make a reasoned statement about.
E.g. ChatGPT can (badly) perform maths or symbolically evaluate code that is guaranteed not to be in its input in a way that would be very hard to explain as "just stringing words together".
Here is an example where I loosely describe a language that is a mix of an INTERCAL-inspired "come from" construct and Ruby, and then give it an example program and asks it what it will output:
> We're going to write a program in a "new" language that is a mix of Ruby and INTERCAL. We're going to take the "come from" statement and use it to allow "hijacking" the return of a function. Furthermore, we're going to do it conditionally. "come from <method> if <condition>" will execute the following block if <method> was executed and <condition> is true. In <condition>, "result" can be used to refer to the result of executing <method>.
>
> Can you deduce what this program will print, and why?
>
> def foo(str)
> str
> end
>
> come from foo if result == "hi" do
> print "foo returned hi"
> end
>
> puts "1:"
> puts foo("some string")
> puts "2:"
> foo("hi")
Here's ChatGPT(GPT4)'s output:
> Sure, let's analyze this program step by step. The new language is a mix of Ruby and INTERCAL, using the "come from" statement to conditionally hijack the return of a function.
>
> Here's the program:
> 1. The `foo` method is defined, taking a single argument `str` and returning it.
>
> 2. The "come from" statement is introduced with a condition to hijack the return of the `foo` method if the `result` is equal to "hi".
>
> 3. The program outputs "1:".
>
> 4. The program calls `foo("some string")` and outputs its result, which is "some string". The "come from" statement is not executed because the result is not equal to "hi".
>
> 5. The program outputs "2:".
>
> 6. The program calls `foo("hi")`. Since the result is "hi", the "come from" statement is executed, and "foo returned hi" is printed. Note that the return value of the `foo` method is not printed here, as there's no `puts` statement wrapping the call to `foo("hi")`.
>
> So, the output of the program will be:
>
> ```
> 1:
> some string
> 2:
> foo returned hi
> ```
This isn't a very complex problem, but it's the kind of problem that requires stepping outside how you're used to think to the point where I'd like to e.g. pose it as an interview question, but I'd worry about using it as I know from experience of conversations with people about INTERCAL's "come from" that a surprising number of people struggle with reasoning about it.
I asked a followup question:
> "come from" alters the control flow, and prevents a method call from returning to its call site. Suggest a couple of implementation and specification issues specific to dealing with the control flow and its implications.
I then followed up with one more:
> Are there other options to unwinding the stack? E.g. when the block passed to "come from" exits, what are our options for where code execution will continue?
Because of looking through this I recognised that the block I'd added to the "come from" syntax made it somewhat ambiguous and might hint at different semantics, I asked this:
> Does the fact I passed a block to the "come from" statement give any hints as to which of these options I might have intended?
> Yes, the fact that you passed a block to the "come from" statement does provide some hints about your intended behavior. It suggests that you might want the "come from" statement to behave like a method interceptor or an "around" aspect, executing the block of code when the specified condition is met and then returning control back to the original call site. [snipped it pointing out that matches one of the options it had described before and copying it's previous explanation]
If this is just "stringing words together", then I know a lot of developers who are bad at just stringing words together.
- In my most recent tests, it will tell you when the data you've provided doesn't match the task (instead of inventing an answer out of thin air).
- It will also add (unprompted) comments/notes before or after the result to disclaim a plausible reason why certain choices have been made or the answer isn't complete.
You have to take into account that not everyone wants the model to not hallucinate. There is a lot of competing pressure:
- Some people would like the model to say "As an AI model trained by OpenAI, I am not qualified to provide an answer because this data is not part of my training set" or something similar, because they want it to only talk when it's sure of the data. (I personally think this use case - using LLMs as search engines/databases of truth - is deeply flawed and not what LLMs are for ; but for a large enough GPT-n, it would work perfectly fine. There is a model size where the model would indeed contain the entire uncompressed Internet, after all)
- Some people want the model to never give such a denial, and always provide an answer in the require format, even if it requires the model to "bullshit" or improvise a bit. An example is, if that as a business user I provide the model with a blog article and ask for metadata in a JSON structure, I want the model to NEVER return "As an AI model..." and ALWAYS return valid JSON, even if the metadata is somewhat shaky or faulty. Most apps are more tolerant to BS than they are to empty/invalid responses. That's the whole reason behind all of those "don't reply out of character/say you are an AI" prompts you see floating around (which, in my experience, are completely useless and do not affect the result one bit)
So the reinforcement is constantly going in those opposite directions.
With respect to the competing draws here, I'm not sure they necessarily compete that much. E.g. being able to ask it to speculate but explain justifications or being able to provide a best guess or being able to ask it to just be as creative as it wants, would be sufficient. Alternatively one that knows how to embed indication of what it knows to be true vs. what is speculation or pure fiction. Of course we can "just" also train different models with different levels of / types of reinforced limits. Having both a "anything goes" model and a constrained version that knows what it does know might both be useful for different things.
If only that were universally true!
there is a qualitative difference: humans may be wrong about facts because they think they are true, ChatGPT is wrong because it does not know what anything means. You cannot fix that, because it's just the way LLMs work.
For example, if asked about a URL for something, a human may remember it wrongly, but will in general say "I don't know, let me check", while ChatGPT will just spew something.
Does stockfish "understand" chess, or does it just "seem to understand" chess ?
For all practical purposes and intent, this doesn't make much difference.
But in AGI, G is the important bit, and neither Stockfish nor ChatGPT have demonstrated general understanding of the world.
E.g. consider this exchange with chatgpt
> hey assistant, can you provide me with a link to learn about LLMs?
> Certainly! LLM stands for Master of Laws, which is a postgraduate degree in law. Here is a link to a website that provides information about LLM programs, including what they entail and how to apply: > https://www.llmstudy.com/what-is-an-llm/what-is-an-llm-.aspx > I hope you find this helpful! Let me know if you have any further questions.
> that is not a real link, can you provide another?
> I apologize for that mistake. Here is the correct link: > https://www.llmstudy.com/what-is-an-llm/what-is-an-llm-.aspx > Please let me know if you have any further questions or if there is anything else I can help you with.
ChatGPT makes up a URL, twice. Does it understand URLs? Does it understand that the URL is wrong? I don't think it does, and I don't think it requires consciousness to do this interaction in a sane way.
Still this does not mean it's useless, it's super useful!
> Essentially all we've done is pushed the basic model proposed by linear regression to it's absolutely limits
No, we haven't pushed linear regression to its limits. If it was only linear regression, it wouldn't work. Neural networks need a non-linearity to model complex things.
The beauty is that given an infinite series of nonlinearities, one can model any mathematical function. In practice we found it takes much less than "infinite", a handful will already get you a long way.
A stack of this basic building block, as you describe it, is really all it takes - we know that mathematically already. The interesting question is: how complex are the functions we need to model? So if we create a neural network of a certain size, is that size large enough to model the problem space?
> However we are only near AGI if this is in fact how intelligence works (or can work) and I don't believe we've seen any evidence of this. And there are some very big assumptions baked into this approach.
I think ChatGPT is good evidence of this. What evidence do we have that this isn't how intelligence works?
What evidence do we have that this is how intelligence works?
Lol, this is precisely what chatgpt doesn't do well at all! It fails to understand simple concepts, logical inference, simple math, etc.
It doesn’t work programmatically—that’s why it fails at logic. But it can reason inductively very very well. Do you have an example besides logic/math where it doesn’t understand simple concepts?
For example it cannot identify musical chords because despite (I presume) ample training material including explanations of how exactly this works, it cannot reasonable represent this as an abstract rigorous rule, as humans do. So I ask what is C E G and it tells me C major correctly, as it presumably appears many times throughout the training set, yet I ask F Ab Db and it does not tell me Db major, because it did not understand the rules at all.
In a sense though, this is just a logic problem.
The reason that chatGPT can write quantum computer programs to in any domain (despite the lack of existing programs!) is because it can deal with the concepts of quantum computing and the concepts in a domain (eg, predicting housing prices) and align them.
Not everything is logic and math; very little of human reasoning is. And this despite the fact that, as a Pythagorean at heart, I believe the world is made of math.
Can you be more specific? I literally don't know what you mean. What can you say about quantum mechanics that is not mathematical or logical in nature? Barring metaphysical issues of interpretation, which I assume is not what you mean.
1. Superposition: In quantum mechanics, particles can exist in multiple states simultaneously, until they are measured. This is called superposition. It's like a coin spinning in the air, being both heads and tails at the same time, until it lands and shows one face.
2. Wave-particle duality: Particles like electrons, photons, and others exhibit both wave-like and particle-like properties. This means they can sometimes behave as particles, and at other times, as waves. This dual nature has been experimentally demonstrated through phenomena like the double-slit experiment.
3. Quantum entanglement: When two particles become entangled, their properties become correlated, regardless of the distance between them. If you measure one of the entangled particles, you'll immediately know the state of the other, even if they are light-years apart. This phenomenon is often referred to as "spooky action at a distance."
4. Heisenberg's uncertainty principle: This principle states that we cannot simultaneously know the exact position and momentum of a particle. The more precisely we know one of these properties, the less precisely we can know the other. This inherent uncertainty is a fundamental aspect of quantum mechanics.
5. Quantum tunneling: In quantum mechanics, particles can "tunnel" through barriers that would be insurmountable in classical physics. This is because the particle's wave function, which describes its probable location, can extend beyond the barrier, allowing the particle to appear on the other side.
6. Quantum superposition of states: Quantum systems can exist in multiple states at once, and when you measure a property of the system, it "collapses" into one of the possible states. This is a fundamental difference between quantum and classical mechanics, where systems have definite properties even before measurement.
These concepts can be discussed and reasoned about without delving into the complex mathematical equations that govern quantum mechanics. While a mathematical understanding is necessary for rigorous study and application of the theory, non-mathematical discussions can still provide valuable insights into the strange and fascinating world of quantum mechanics.*
QED
Claiming these concepts are not mathalematical is like saying addition is not mathematics because you can explain it with words or diagrams to a child!
I hate to break it to you, but humans aren't thinking logically, or exclusively logically. In fact I would say that humans are not using logic most of the time, and we go by intuition most of our lives (intuition is shorthand for experience, pattern matching and extrapolation). There is a reason of why we teach formal logic in certain schools....
It even told me about the role of each in a chord progression and how even though they share the same notes they resolve differently
Humans clearly don't think logically anyhow. Thats why we need things like abacus to help us concretely store things, in our head everything is relative in importance to other things in the moment
So it gave you a wrong answer and when you spelled out the correct answer it said "OK" x) Was that it? Or am I missing something.
In music theory a set of the same notes can be one of several chords, the correct answer as to which one it is depends on the "key" of the song and context which wasnt provided, so the AI decided to define the root note of the chord as the bottom one which is a good and pretty standard assumption.
It's kind of like asking the AI about two equivalent mathematical functions expressed in different notation, and it saying a bunch of correct facts about the functions like their derivative, x intercept and stuff and how it can be used, but needing a prod to explicitly say the fact that they are interchangable. It's the kind of trivial barely-qualifies-as-an "oversight" I would expect actual human people who fully understand the material to make.
All the time. It often fails to understand simple concepts. It doesn't really seem to understand anything.
For example, try to get it to write some code for a program in a moderately obscure programming language. It's terrible: it will confidently produce stuff, but make errors all over the place.
It's unable to understand that it doesn't know the language, and it doesn't know how to ask the right questions to improve. It doesn't have a good model of what it's trying to do, or what you're trying to do. If you point out problems it'll happily try again and repeat the same errors over and over again.
What it does is intuit an answer based on the data it's already seen. It's amazingly good at identifying, matching, and combining abstractions that it's already been trained on. This is often good enough for simple tasks, because it has been trained on so much of the world's output that it can frequently map a request to learned concepts, but it's basically a glorified Markov model when it comes to genuinely new or obscure stuff.
It's a big step forward, but I think the current approach has a ceiling.
I don’t think its ability to program in an obscure program is really a great test. That’s a matter of syntax more than semantics, no?
Novel conceptual blends are where it excels. Yes, it needs to understand the concepts involved to blend them —but humans need that too.
It doesn't seem to have any "meta" understanding. It's subconscious thought only.
If I asked a human to program in a language they didn't understand, they'd say they couldn't, or they'd ask for further instructions, or some reference to the documentation, or they'd suggest asking someone else to do it, or they'd eventually figure out how to write in the language by experimenting on small programs and gradually writing more complex ones.
GPT4 and friends "just" take an input that seems like it could plausibly answer the request. If it gets it wrong then it just has another go using the same generative technique as before with whatever extra direction the human decides to give it. It doesn't think about the problem.
("just" doing a lot of work in the above sentence: what it does is seriously impressive! But it still seems to be well behind humans in capability.)
I think artifacts can easily reflect the understanding of the designer (Socrates claims an etymology of Technology from Echo-Nous [1])
But for an artifact to understand — this is entirely dependent on how you operationalize and measure it. Same as with people—we don’t expect people to understand things unless we assess them.
And, obviously we need to assess the understanding of machines. It is vitally important to have an assessment of how well it performs on different evals of understanding in different domains.
But I have a really interesting supposition about AI understanding that involves it’s ability to access the Platonic world of mathematical forms.
I recently read a popular 2016 article on the philosophy of scientific progress. They define scientific progress as increased understanding — and call it the “noetic account.” [2] Thats a bit of theoretical support for the idea that human understanding consists of our ability to conceptualize the world in terms of the Platonic forms.
Plato ftw!
[1] see his dialogue Cratylus
[2] Dellsén, F. (2016). Scientific progress: Knowledge versus understanding. Studies in History and Philosophy of Science Part A, 56, 72-83.
Is that really any different than asking me to attempt to program in a moderately obscure programming language without a runtime to test my code on? I wouldn't be able to figure out what I don't know without a feedback loop incorporating data.
>If you point out problems it'll happily try again and repeat the same errors over and over again.
And quite often if you incorporate the correct documentation, it will stop repeating the errors and give a correct answer.
It's not a continuous learning model either. It has small token windows where it begins forgetting things. So yea, it has limits far below most humans, but far beyond any we've seen in the past.
To the extent that humans have encoded the concepts into words, and that text is in the training set, to that degree ChatGPT can work with the words in a way that is at least somewhat true to the concepts encoded in them. But it doesn't actually understand any of the concepts - just words and their relationships.
But I suspect your notion of understanding is not measurable, is it? For you, chatGPT lacks something essential such that it is incapable of understanding, no matter the test. Or do you have a way to measure this without appeal to consciousness or essentialism?
Does understanding require consciousness? Maybe yes, for the kind of understanding I'm thinking of, but I'm not certain of that.
How do you measure understanding? You step a bit outside the training set, and see if whoever (or whatever) being tested can apply what it has learned in that somewhat novel situation. That's hard when ChatGPT has been trained on the entire internet. But to the degree we can test it, ChatGPT often falls down horribly. (It even falls down on things that should be within its training set.) So we conclude that it doesn't actually understand.
How come humans are so efficient when these computers are using enormous amounts of energy? When will we have a computer that only requires a few slices of pizza worth of energy to get to the next step?
I don't think this is the case at all. Language is how we encode and communicate ideas/concepts/practicalities; with sufficient data, the links are extractable just from the text.
What GPT does is not "text" - although it centers that as the interface - but "symbols". The billions of parameters express different syntaxes and how they relate to each other. That's why GPT can translate languages and explain things using different words or as different personas.
So when we ask it to solve a spatial problem, we aren't getting a result based on muscle memory and visual estimation like "oh, it's about 1/3rd of the way down the number line". Instead, GPT has devised some internal syntax that frames a spatial problem in symbolic terms. It doesn't use words as we know them to achieve the solution, but has grasped some deeper underlying symbolic pattern in how we talk about a subject like a physics problem.
And this often works! But it also accounts for why its mathematical reasoning is limited in seemingly elementary ways and it quickly deviates into an illogical solution, because it is drawing on an alien means of "intuiting" answers.
We can definitely call it intelligent in some ways, but not in the same ways we are.
Ask a random person stuff like whats 465*42/12 and you'll get ridiculous answers, and LLM can use tools now.
I'd argue that it's actually superior to humans in that regard
Does chatGPT need a way to verify reality for itself to become truly intelligent ?
1) My confidence was around 80%, without any sense of absolute certainty. 2) I had no idea why I knew that, and where did I originally learned about what a house centipede looks like.
There's a good argument to be made you learned about potentially dangerous insects as a child, from books aimed at children. A reasonable person would make this argument. It's also likely that you retained that information because your would-be peers who aren't around, aren't around because their ascendants inserted their genetic material in to the humanity's descent's to a lesser extent than your extant peers.
You might not be able to immediately recall where and when you learned something, and that shouldn't be equated with genuinely having no idea.
I'd settle for a poor fascimillie, and argue strongly against human-like.
Human-like would be, let's say, a Cylon skin-job. ChatGPT is just a fraking toaster, at best.
But I agree with GP's eager-anthropomorphization complaint. When algorithms produce verifiably wrong output we call them errors.
Hallucinations are a mistake in perception.
The key difference appears a combination of two factors: We appear to be more likely to have learnt which subjects we're not very knowledgeable about through extensive feedback, be it through school or conversations where we're told we're wrong, and which also would appear to teach us to be more cautious in general. We also appear to have a (somewhat; far from perfect) better ability to separate memory from thoughts about our knowledge. We certainly can go off on wild tangents and make stuff up about any subject, but we get reinforced from very young that there's a time and a place for making stuff up vs. drawing on memory.
Both goes back simply to extensive (many years of) reinforcement telling us it has negative effects to make stuff up and/or believe things that aren't true, and yet we still do both of those, just not usually as blatantly as current LLMs without being aware.
So I'd expect one missing component is to add a training step that subjects the model to batteries of tests of the limits of their knowledge and incorporating the results of that in the training.
Agreed. I'd like to add another point to the discussion. It seems to me, as if LLMs are held to a higher standard regarding telling the truth than humans are. In my opinion the reason for this is that computers have been traditionally used to solve deterministic tasks and that people are not used to them making wrong claims.
Imagine subjecting a random human to the same battery of conversations and judging the truthfulness of their answers.
Now, imagine doing the same to a child too young to have had many years of reinforcement of the social consequences of not clearly distinguishing fantasy from perceived truth.
I do think a human adult would (still) be likely to be overall better at distinguishing truth from fiction when replying, but I'm not at all confident that a human child would.
I think LLMs will need more reinforcement from probing the limits of their knowledge to make it easier to rely on their responses, but I also think one of the reasons people hold LLMs to the standard they do is also that they "sound" knowledgeable. If ChatGPT spoke like a 7 year old, nobody would take issue with it making a lot of stuff up. But since ChatGPT is more eloquent than most adults, it's easy to expect it to behave like a human adult. LLMs have gaps that are confusing to us because the signs we tend to go by to judge someones intelligence are not reliable with LLMs.
Paradoxically, it seems as if the people who are pushing this the hardest are the same people who flat out deny even the slightest flicker of what could be considered intelligence.
Saying “most humans make dumb statements” is not a defense to build a machine that makes dumb statements imo
If we made a ChatGPT that has some “senses” it can use to verify its perceptions, and it does as well with this as humans generally do, I’m sure we’ll get interesting results but I’m not sure we’ll be any closer to solving the problem.
In other words, the human has a concept of truth or facts and simply had a memory lapse. This thing has no concept of truth at all.
People don't need the gigantic amount of input data that ChatGPT needs to learn. However I'm not sure what exactly "this" is you and GP are referring to, and it may be possible to improve existing ideas so that it works with less input data.
The training data isn't the input. It's a part of the algorithm.
Sure they do. Humans rely on tons of audio and video before they can even read (or walk).
>Let us consider the GPT-3 model with 𝑃 =175 billion parameters as an example. This model was trained on 𝑇 = 300 billion tokens. On 𝑛 = 1024 A100 GPUs using batch-size 1536, we achieve 𝑋 = 140 teraFLOP/s per GPU. As a result, the time required to train this model is 34 days.
https://arxiv.org/pdf/2104.04473.pdf
I'm not sure expressing brain capacity in FLOPs makes much sense, but I'm sure if it can be expressed in FLOPs, the amount of FLOPs going to learning for a normal human is less than that.
For reading, the same applies. Our brains are equipped with many of the foundational aspects required for reading, and we only _learn_ a part what is necessary for the skill of reading.
Unlike computer models, brains are no tabula rasa. So we don't need the same input as computer models to learn.
You are not storing it all, but you are finding patterns and correlations in it from the day you were born. This all forms a base where after a decade+ you can learn things fast from books, but that’s comparable to an LLM’s in context learning. It’s fast, but it depends on a deep base of prior knowledge.
And even if I did, most of it would not be of the same category as what LLMs are learning; information about how a fabric feels or the sound of an ambulance doesn't teach me anything about programming or other things GPT can do. So when comparing the inputs, it makes no sense to count all the information that our senses are getting, to all the input for an LLM.
>This all forms a base where after a decade+ you can learn things fast from books, but that’s comparable to an LLM’s in context learning
Not really comparable
I think you discount general knowledge acquired non-verbally too easily. The sound of an ambulance and how it moves through space and how it correlates with visual information represents an astronomical amount of information that can be generalized from handsomely. All these data streams have to be connected, they correlate and intertwine in fantastically complex ways.
I think the sound of an ambulance does teach you things that eventually help you “program”. You have witnessed similar (non-)events thousands if not millions of times. Each time it was accompanied with shitloads of sensory data from all modalities, both external and internal.
The base of general patterns you start out from once you are ready for language is staggering.
Again not saying LLMs work like that, because they do not. All I mean to do is put their information requirements in perspective. We ingest a lot more than a bunch of books.
The common perception has been that children aren't exposed to enough data to arrive at their grammatical language skills, implying there's some proto language built in. Comparative analysis of languages has looked for what aspects are truly universal but there's actually not a lot of concrete true universals to ascribe to our genetic innate language.
But if it is genetic that doesn't really mean it's fundamentally different than ChatGPT, it just took a different and drastically longer training period and then transfer learning when children learn their mother tongue.
It doesn't necessarily mean it's fundamentally different, but it doesn't mean it is comparable either. Geoff Hinton doesn't think the brain does backpropagation. Training a neural net uses backpropagation. So if Hinton is correct then saying "it just took a longer training period" while brains doesn't learn like our current neural nets is glossing over a lot of things.
Animals have bodies and have to survive in an environment. They're not just sitting there waiting on a prompt to generate text. They do things in the real world. Language is later invention by one particular social animal which serves our needs to communicate, which is different than waiting on prompts.
There are certainly some missing elements, such as ongoing training and memory, but it's not at all clear that there's any major conceptual differences. There could be, we just don't really know.
If you conclude, after watching this, that this is all intelligence is - namely cascading optimization and data structure problems chained in a row, then you and a fair number of people here wont ever find common ground. That is not to say that you're wrong, it just seems as if there is a missing component here that we haven't yet discovered.
They won't find common ground with who?
I think it is easy to distracted by the specific mechanisms by which these models work, but most of the technological detail is because we want something to happen on systems at the scale of what we can actually build. We simply can't build a human brain scale neural net yet. We build what we can and besides maybe with all this research we will figure something significant about what intelligence actually is.
The notion "This can't be all thought is" is as old as the idea of AI. I think it informed Turing when he proposed the Imitation Game. The insight is that people would be resistant to the idea of a bunch of simple things stuck together becoming thinking until they were faced with something that behaves sufficiently indistinguishable from themselves that to doubt that they were thinking would be akin to doubting everyone you meet.
In the end some people won't even accept an AI that does everything a human does as actually thinking, but then again some people are actually solipsistic.
Older still:
>It must be confessed, moreover, that perception, and that which depends on it, are inexplicable by mechanical causes, that is, by figures and motions, And, supposing that there were a mechanism so constructed as to think, feel and have perception, we might enter it as into a mill. And this granted, we should only find on visiting it, pieces which push one against another, but never anything by which to explain a perception. This must be sought, therefore, in the simple substance, and not in the composite or in the machine.
- Leibniz, 1714
For extraordinary claims ('intelligence'), the burden of proof is on those making the claim, not on others to prove the negative.
The sine function is nonlinear.
You give sprinkle in a tiny nonlinearity (e.g. x^2 instead of x) and suddenly you can get infinite complexity by weighted composition - which is also the reason why we're so helpless with nonlinear functions and reach for linear approximations immediately (cf. gradient descent).
Except you can "compose" affine functions using Horner's schema to get any polynomial... No need to sprinkle tiny non linearities. It's a buffet. Grab as much as you need.
y = b + x1 + x1^2 + x1^3 + ... + x1 * x2 + (x1 * x2)^2 + ... + x2 + x2^2 + ...
By that point you're making a Taylor approximation of the latent function through linear space, which is also a universal approximator.
So the commenter above is wrong -- neural networks are indeed just glorified linear regression from this point of view.
The main difference is that this kitchen sink regression is computationally inefficient which neural nets are extremely efficient computationally.*
I'm not an expert, but the motivation seems more like this:
- Linear regression and SVM sometimes work. But they apply to very few problems.
- We can fit those models using gradient descent. Alternatives to gradient descent do exist, but they become less useful as the above models get varied and generalised.
- Empirically, if we compose with some simple non-linearities, we get very good results on otherwise seemingly intractable problems like OCR. See Kernel SVM and Krieging.
- Initially, one might choose this non-linearity from a known list. And then fit the model using specialised optimisation algorithms. But gradient descent still works fine.
- To improve results, the choice of non-linearity must itself be optimised. Call the non-linearity F. We break F into three parts: F' o L o F'', where L is linear, and F' and F'' are "simpler" non-linearities. We recursively factorise the F' and F'' in a similar way. Eventually, we get a deep feedforward neural network. We cannot use fancy algorithms to fit such a model anymore.
- Somehow, gradient descent, despite being a very generic optimisation algorithm, works much better than expected at successfully fitting the above model. We have derived Deep Learning.
Deep learning has been a series of *engineering* successes stacking over each other rather than theory being applied rigorously.
It's hard to scale training on the "dumb" approximators like a kitchen sink regression, and balancing overfitting becomes a nightmare.
(Edit: this whole thread feels like a setup for this punchline)
ChatGPT points to us possibly going in the right direction. It is so good that people need to write long articles about why it is not as good as a human. Contrast to older efforts, which were pretty clearly not so great. I find this pretty compelling, and GOTCHA examples that show that ChatGPT isn't as good as humans in everything miss the point.
Birds fly and drones don't look like birds - but they fly. If the goal is flying, it's okay that we achieve it through something that doesn't exactly mimic what inspired it. Do we need a "full human" for all we do? Our billions of machines, many performing work previously done by humans, show that we don't.
If we can largely replicate intelligence, it's not super important whether "this is how human intelligence works".
Edit: the important bit is that the subject already has some qualities in common with the proposed classification which is a significantly different proposition than a purely hypothetical orbital teapot
The fact that ChatGPT doesn’t behave like an intelligence. (Though it converses like one with radical deficiencies in certain areas, which highlights the narrowness of the Turing Test, which itself is a big step.)
OTOH, it gets less bad at this when you wrap it in a loop that provides recall and other capacities with carefully constructed prompting as to how (“thought” process-wise, not just interaction mechanics) to integrate those capacities, so there’s maybe a decent argument that it models an important component of general intelligence.
1) Given an infinite series of nonlinearities, one can model any mathematical function to an arbitrary degree of precision. This has been proven and is accepted by any working mathematician.
2) The human brain / intelligence can be simulated by a sufficiently complex mathematical function. It is somewhat accepted that simulating the human brain is sufficient for intelligence. Disagreements usually boil down to: It won't have a soul, you can't simulate biological / quantum processes on digital computers, or something abut Qualia I don't really understand,
3) How big / complex is the function that we need to get to AGI / simulate human or above intelligence? TBD, it seems like experts disagree but many working in the field have been surprised by the capabilities that GPT has and we are likely closer to AGI than anyone thought 2 years ago.
Our ability to reason, from an evoultion perspective came about from needing to communicate with others, i.e to give reasons for things, it's this same system we use to be 'logical' or think we are being inteligent from, but it's effectively just a system to make excuses for things, hence why we might choose something in a store for reasons of falling for some form of marketing, but give a completely different logical sounding reason for doing so, hence conspiracy theorists, when confronted with evidence against that theory, don't change there theory, but change there excuse.
I believe this could be more profound and have bigger implications than we have realised.
This is no different to what a lot of AI models are doing right now when they are wrong and people are calling that out saying there is no intelligence behind that, as it's spouts some nonsense reason for being right, but the problem is, humans do this all the time and I think we do it way more than we realise, the trouble is we are so caught up in it, we cannot see it.
What we have with the human brain is something that wasn't designed logically but via evolution which gives the illusion of intelligence or arrives at mostly the same result, but in completely different ways than expected, therefore I think the path to an AGI, might not be quite what some are saying or expects it to be, because the human brain certainly doesn't work like that.
If that's the case, there doesn't need to be anything new; fundamentally, linear regressions may be sufficient.
Lol, let me ask ChatGPT what it thinks about that. :)
> I don't believe we've seen any evidence of this
What kind of evidence would you like to have? Do you want a mathematical proof or what? There is evidence that we are making forward progress in solving problems which were previously in the domain of human cognition. There is evidence that yesterday's "impossible" problem become "possible" while the "hard" problems become "easy" or even "trivial". (Just look at this xkcd[1]. When it was published in 2014 telling whether or not a picture contained a bird was indeed a "5 year and a team of researchers" project. Today it is what, an afternoon? A tutorial you pick up to learn a new ML framework?)
There is also evidence that our solutions are tending toward more generalised ones. Previously you would need a "sentiment detection" network, and a separate "subject disambiguation" network to parse the meaning out of a text. Today you can achieve the same or better with an LLM trained to follow instructions.
Obviously these are not "hard evidence" that this path will lead to AGI. But it is certainly not unreasonable to think that it might.
Schmidhuber, is it you?
But seriously, it's partially true most techniques we use today could be found initially envisaged in 90th ANN-related papers. I think Geoffrey Hinton summarized most of them well in his famous "Neural Networks for Machine Learning" lectures. Essentially, what is available today is compute resources unimaginable in 90th, so scaling up from a shallow 2-layered Multi Layer Perceptron to something like 96-layered deep architecture is possible only recently. We also found some of the tricks work better than others in practice when we scale up (like *ELU non-linearity, layer-norm, residual connections). What stays the same however is the general approach: training and validation sets, cross entropy loss, softmax, learnable parameters based on data-in/data-out training pairs, and differentiation chain rule. IMO this requires some innovative revision, especially generalization is still very weak in all architectures today.
Honestly I've always surprised at the scepticism AI researchers have had of about the limits of training large neural nets with gradient descent. Where I've had my doubts is in the architecture of existing models and still think this is their primary limiting factor (more so than compute and network size).
I think the question that remains now is whether existing models are actually capable of producing a general intelligence that's well rounded and reliable enough to be competitive with human general intelligence. Personally, I think LLMs like GPT-4 are generally intelligent in most ways and should be considered AGI already (at least in a weak sense of the word), but they clearly have notable gaps in their intelligence such as their ability to be consistent, long-term memory, and ability to discern reality from delusion.
I don't think scaling existing models could possibly address these limitations – they seem to be emergent properties of an imperfect architecture. So I suspect we're still a few breakthroughs away from a general intelligence as well rounded as human general intelligence. That said, I suspect existing models (perhaps with a few minor tweaks) are generally intelligent enough that larger models alone are likely still able to replace the majority of human intellectual labour.
I guess what I'm touching on here is the need to more nuanced about what we mean by "AGI" at this point. I think it's quite likely (probable even) that in a few years we'll have an AI that's generally intelligent and capable enough that it can replace a large percentage of existing knowledge work – and also generally intelligent enough to be dangerous. But I suspect despite this it will still have really clear limitations in its abilities when contrasted with human general intelligence.
For me AGI is achieved when AutoGPT is at a point when it's able to improve its own algorithm (specifically improve on the GPT architecture).
Flash attention came out a bit less than a year ago (27 May 2022), which was a great scaling improvement for getting rid of O(n^2) memory requirement in the length of attention.
I guess the next one we need will be one of the solutions that change the FLOPS for computing attention from n^2 to n*log(n) (there was already a paper that achieved it using resizable gauss filters where the size and the base convolution filter is learned separately).
A few of these kind of ,,breakthroughs''/algorithmic improvements are the only ones needed to get to AGI (self improving machine) in my opinion.
There are various types of retorts: i) the brain is also doing gradient descent; ii) what the brain does does not matter (if you can fake intelligence you have intellegence) iii) the pace is now so fast that has not happened in decades will happen in the next five years etc.
None of them is remotely convincing. I would really start getting worried if indeed some of the AI fanboys had a convincing argument. But we are where we are.
Somehow this whole episode will be just another turn of the screw in adopting algorithms in societal information processing. This in itself is fascinating and dangerous enough.
There is a case to be made that with sufficient ingenuity at some point people will expand the algorithmic toolkit into more flexible and powerful dimensions. It may integrate formal logic in some shape or form, or yet to be conceived mathematical constructs.
But the type of mental leap that I think would be required just to breakout of the statistical fitting straight-jacket cannot be invented to satisfy the timing of some market craze.
If you look at the broad outline of the development of mathematics there are many areas where we have hit a complexity wall and generations of talent have not advanced an iota.
Even if we condition on some future breakthrough, the next level of mathematical / algorithmic dexterity we might reach will follow its own intrinsic logic, which will probably be very interesting but may or may not have anything to do with human intelligence.
Not taking a stance on whether "intelligence" can emerge out of gradient descent but it's certainly not biological.
Just make humans dumber and more predictable. Problem solved. (This isn't that hard since humans will adapt to the stupidity level of the "AI" they use daily on their own without special social engineering efforts.)
Where's the data to show that we've hit the limits?
It's obviously not linear regression. You mean gradient descent?
AGI ≠ a lot of AI. They are fundamentally different things.
The first computer was designed in 1837, long before a computer was ever built. We know how fusion reactions work, now we’re tweaking the engineering to harness it in a reactor.
We don’t know how human intelligence works. We don’t have designs or even a philosophy for AGI. Yet, the prevailing view is that our greatest invention will just suddenly “emerge.”
No other field I’m aware of so strongly purports it will reach its ultimate breakthrough without having a clue of the fundamentals of that breakthrough.
It’s like nuclear scientists saying “if we just do a lot of fission, we think fusion will just happen.”
Going by the differences between gpt3.5 and gpt4 is really interesting. It is better able to reason in basically any problem I throw at it. Personally I think that a hypothetical system that is able to generate a sufficiently good response for ANY text input is AGI.
There aren't really any "gotcha" cases with this technology that I'm aware of where it just can't ever respond appropriately. Most clear failings of existing systems involve ever more contrived logic puzzles, which each successive generation is able to solve, and eventually at some point the required logic puzzle will be so dense few humans can solve it.
This isn't a case a case of "studying for the test" of popular internet examples either. I encourage you to try and invent your own gotchas for earlier versions then try them on newer models. Change the wording and order of logic puzzles, or encase them within scenarios to ensure its not responding to the format of the prompt
There are absolubtely cases of people overhypting it, or it overfitting to training data (see the debacle about it passing whatever bar exam, university test etc.). But despite the hype there is an underlying level of intelligence that is building and I use it to solve problems pretty much every day. I think of it atm as like a 4 year old that has inexplicably read every book ever written
In many ways, AGI and AI are opposites. We don’t actually “train” humans. We present them information. And a human can choose to ingest that information or - just as importantly - choose not to.
A 3yo learns English not through being force fed, but through their own curiosity and motivation to learn.
GPT and the like are proving to be amazing tools. The compute model of probabilistic algorithms (“plain” AI) is going to transform industries.
But the more sophisticated these tools get, the further from AGI they will become.
When we create AGI, it will begin with a blank slate, not the whole of human knowledge. It will study some topics deeply and be uninterested in others. It might draw its own conclusions instead of just taking the conclusions presented to it. Or it may decide not to study at all, preferring to instead become a hermit.
David Deutsch is the most compelling on this topic: https://open.spotify.com/episode/5nGOkhcJ7VszXm2WVgxKyC?si=8...
I think some people explitly dont want AGI to simulate emotion or self-defined goals because that would be little more than an artistic curiousity. Why intentionally make a tool less useful? Less willing to do the things you want it to do? Perhaps you hold some belief that simulating such things is necessary for some goal you have in mind, but personally I don't think that's true, for literally any goal
LLMs are not able to create new knowledge, only organize existing knowledge. Critically, we don't use induction to create new knowledge - but induction is all these LLMs can do. (e.g. To explain the origin of the universe, there is nothing we can induce, as the big bang is not observable.)
I don't see how training these AI models more will cause this property to emerge. But knowledge creation is the thing that we want when we say we want AGI.
The reason I describe AGI as such is because we only have one example of what an AGI is, namely humans. A popular idea is that machine AGI will look very different from human AGI (e.g. no consciousness or intrinsic motivation or qualia).
But this is a bold claim. It contends that there are multiple kinds of general intelligences vs some universal kind. Other universal concepts don't look like this. Consider the universality of computation: a computer is either Turing-complete or it isn't. There is no other kind of general purpose computer.
Once we created a Turing-complete computer, it could perform all conceivable computation. This is the power of universality. The first Turing-complete computer could technically play Call of Duty (though no one would have had the patience to play it).
It's not like there is some mode of ChatGPT that will produce AGI-level responses if given months or years of time to compute it.
Fundamentally, its ability to do whatever it is that AGIs do (AGI's version of Turing computation) is missing.
Similarly if I gave it a phenomena I wanted to know about it could propose experiements to run to understand that thing.
taking independant action in the world is indeed beyond the limits of chatgpt on it's own, but its not hard to build systems relying heavily on chatgpt to do that, a simple loop calling it with information from sensors and a prompt to experiment with outputs is enough to get some kind of agent. It can process any information, and given more power it can draw ever-higher inferences from that information. This is analogous to a turing machine, something that can eventually get anywhere, but it will take an arbitrarily long time to do it.
I think of it kinda like the CPU in a Von Neumann architecture. A cpu on it's own doesn't do much of anything unless you tell it to. We have that part solved, we just need to settle on a good structure for repeatedly querying it.
It's obvious it's an AI because it's too smart. It's too intelligent. Dumb it down, remove the AI branding, and it would fool the majority of the world.
We did agree on a definition intelligence! For 50 years, the Turing test was the unquestioned threshold beyond which machines would be considered "intelligent". The simple fact of the matter is that we've reached that point, but folks remain unimpressed, and so have set about moving the goalposts.
I am of the opinion that, when the dust settles, the point in time which will be selected as "the moment we achieved Artificial Intelligence", or even rudimentary AGI, will not be in the future, but in the past. This will be true because, once we take a step back, we'll remember that our definition of intelligence should not be so strict that it it excludes a significant percentage of the human population.
Consider this--is there any reasonable definition of intelligence which excludes Chat GPT 4.0, but which includes all human beings who fall in the bottom ~5% of IQ?
First of all it is now more then 70 years. Secondly the question of the paper isn't whether a computer is intelligent but whether a computer can win the imitation game, so naturally it doesn't contain any definition of intelligence within it.
The game IS Turing's definition of intelligence! From the paper:
> I PROPOSE to consider the question, ‘Can machines think?’ This should begin with definitions of the meaning of the terms ‘machine’ and ‘think’. The definitions might be framed so as to reflect so far as possible the normal use of the words, but this attitude is dangerous. If the meaning of the words ‘machine’ and ‘think’ are to be found by examining how they are commonly used it is difficult to escape the conclusion that the meaning and the answer to the question, ‘Can machines think?’ is to be sought in a statistical survey such as a Gallup poll. But this is absurd. Instead of attempting such a definition I shall replace the question by another which is closely related to it and is expressed in relatively unambiguous words. The new form of the problem can be described in terms of a game...
If we’re talking about something like agency or free will, that’s trickier.
The problem in this space is people use terms that nobody agrees on.
At the end of the day all ML is using gradient descent to do some sort of non-linear projection of the data on to a latent space, then doing some relatively simple math in this latent space to perform some task.
Personally I think the limits of this technique are far better than I would have thought 10 years ago.
However we are only near AGI if this is in fact how intelligence works (or can work) and I don't believe we've seen any evidence of this. And there are some very big assumptions baked into this approach.
Essentially all we've done is pushed the basic model proposed by linear regression to it's absolutely limits, but, as impressive as the results are, I'm not entirely convinced this will get is over the limit to AGI.