Jeff Bezos on AI (1998) [video]
The common sentiment around AI in the 90s and early 2000s was that it didn't work; it had its hype, it had its heyday, but it seemed like a dead-end for the most part. The Perceptron was merely a linear function approximator. And the Multi-layer Perceptron was a little more capable, but the many orders of magnitude it would have to scale up in order to be convincing just wasn't feasible back then (it finally was in the 2010s).
Simple statistical models that aren't "AI" so much as just generic ML were and are quite useful: like recommendation and newsfeed engines ("the Algorithm" as we call it so often today). Love 'em or hate 'em, they can be quite good at predicting interest/engagement.
The resurgence in deep learning in the 2010s has shown us new magic tricks, but they're still just that: parlor tricks. At least they're more convincing tricks than what we had 40 years ago!
That's what ultimately depresses me about AI. It's still just a parlor trick. We haven't actually taught computers to think, to reason, to be innovative. Deep learning is definitely having its day, but I suppose this too will pass unless we can unlock certain ways to make AI reliable and responsible. Or... to just start understanding it in general.
> That's what ultimately depresses me about AI. It's still just a parlor trick. We haven't actually taught computers to think, to reason, to be innovative.
And what do you feel when we make these parlor tricks more capable than us at the majority of tasks?
And what do you feel when we understand it well enough to realize we're the same type of parlor tricks?
To me it seems like you're most interested in a magic 'aha' moment and will miss or not be prepared for how the road in front of us likely unfolds.
And what do you feel when we understand it well enough to realize we’re the same type of parlor tricks?
That’s called positivism and it has a lot of philosophical issues. I wouldn’t be so quick to assume that sensory appearance is equivalent to reality.
Sensory appearance not being equivalent to reality does not have any relevance to the question of AI and humans ultimately being the same kind of information-processing system. Just handwaving "that's X philosophical position and it has problems" does not strike me as a good argument either unless you manage to explain how these problems pertain to the question at hand.
Unless AI becomes indistinguishable from human beings on a cellular level, yes, it’s entirely relevant and is the single most relevant thing. A lot of people seem to think that if an AI can simulate the appearance of a human being, that makes them equivalent to one. It might introduce some problems WRT to determining if an entity is human or not, but this doesn’t somehow prove they humans are just a “parlor trick.”
This is a positivistic argument and as I pointed out, positivism has a lot of issues. The best counter argument IMO being that it’s needlessly reductive. This is all covered pretty clearly in the link.
> Unless AI becomes indistinguishable from human beings on a cellular level, yes, it’s entirely relevant and is the single most relevant thing.
I disagree.
Thought experiment: design a circuit which has as many inputs and outputs as a biological neurone, such that it always maps inputs to outputs in the same way (including the observation that this isn't a static map but one which changes over time), then connect them as neurons are in one of us.
While clearly nothing like an natural brain on a cellular level, I believe this is a sufficient similarity to be "the same parlour tricks".
The question then is: how close does the design actually need to be, while not losing anything of importance?
Perceptrons were only ever a toy model, so they may well be insufficient; but on the other hand, for a sense of scale, GPT-3 is about the complexity of a rodent brain rather than a human brain — and that suggests that humans could learn to be simultaneous experts in many dozens of fields and languages with a mere tenth of a percentage point of our brains if only we lived long enough to read the entire internet.
Which matters most — neurons, connective structure, learning environment, or something else — is, I think, still an open question. But even between all the differences, AI collectively are general purpose enough to at least suspect these things have got a lot of similarities where it matters.
>A lot of people seem to think that if an AI can simulate the appearance of a human being, that makes them equivalent to one.
That is not what BoiledCabbage was saying. He was saying: "And what do you feel when we understand it well enough to realize we're the same type of parlor tricks?"
>This is a positivistic argument and as I pointed out, positivism has a lot of issues. The best counter argument IMO being that it’s needlessly reductive. This is all covered pretty clearly in the link.
You're not really making any specific claim about what is wrong with BoiledCabbage's speculation, and why this specific thing is wrong. "That's wrong because positivism, it's all in this 5k word wikipedia article!" just doesn't prove anything.
If you haven’t done the reading, I can’t explain it to you in a HN comment. I’m not trying to be snarky about it, but I genuinely don’t know what else to tell you. This is a pretty foundational ideal in the philosophy of science.
What’s wrong with the speculation is that it’s a positivistic argument that is needlessly reductive. It’s reductive because it assumes that appearing human-like is equivalent to being human.
The fact that we can understand how “AI” works as a parlor trick yet appears human-like in no way implies that human beings are nothing more than the same parlor trick processes. To argue that it does is to make a positivistic argument that doesn't take in account a whole host of other things. As noted in the Criticism section of the article (which is hardly 5,000 words) there are many issues with this approach.
>It’s reductive because it assumes that appearing human-like is equivalent to being human.
I don't read that assumption into BoiledCabbage's statement at all: "[..] when we understand it well enough to realize we're the same type of parlor tricks?" This clearly implies a (hypothetical) deeper understanding of processes in the brain and their specific qualities, rather than (as you seem to be implying) a mere comparison of the outputs.
Edit: anyway, the criticism section opens like this:
>Historically, positivism has been criticized for its reductionism, i.e., for contending that all "processes are reducible to physiological, physical or chemical events," "social processes are reducible to relationships between and actions of individuals," and that "biological organisms are reducible to physical systems."
This (at least the 1st and 3rd quoted item, while I think the 2nd one is just out of scope) is exemplary of the kind of things that are obviously true for anyone but a subset of philosophers clinging to magical and unprovable beliefs about the human mind. I asked you to elaborate your argument precisely because if it all boils down to simply rejecting physicalism (in philosophy of mind terms) there's nothing new to argue about. The recurring discussion about "AI can never be like humans" is only interesting when the participants do a little bit more than just staking out their own position in idealism vs dualism vs physicalism terms and regurgitating all the known debates between these camps.
I don't read this statement as hypothetical at all.
And what do you feel when we understand it well enough to realize we're the same type of parlor tricks?
It seems pretty clearly when and not if. Not sure what you're reading there.
"When" is in the future so I don't get your issue. Unless your point is that we can in principle never hope to understand certain things about it.
When implies that the thing is going to happen. If implies that it might happen.
https://dictionary.cambridge.org/grammar/british-grammar/if-...
The comment is written in such a way that it assumes this is the nature of reality and at some point, we will learn that the human mind is no different from the parlor tricks of LLMs. This is what I was criticizing.
>The comment is written in such a way that it assumes this is the nature of reality and at some point, we will learn that the human mind is no different from the parlor tricks of LLMs.
Again your whole point seems to boil down to simply denying physicalism (unless you have something more subtle in mind about the limits of physical science). It's a valid position to take but not very fruitful for further conversation.
No, I'm not talking about physicalism at all. I'm talking about reductionism and emergence. Even if it were shown that human thought processes operate at lower levels in ways similar to machines, that does not imply that human beings are equivalent to machines. That is a reductive argument.
Now I don't understand at all anymore why you're disagreeing so strongly. I don't think anyone has proven that it's inherent to emergent properties that they can't be understood or explained in detail. One just has to explain the mechanisms of emergence in addition to explaining the component parts reductively (and if that is your objection to reductionist arguments I would be inclined to agree with you). But such an explanation of how intelligence can emerge from simple component parts is exactly what could potentially be provided by a better understanding of AI systems.
That statement also has no basis in neuroscience.
Computers are already better than humans at a wide variety of tasks. Text generation just happens to now be one of those tasks. But if you look at the prompt -> output -> prompt feedback loop, it's clear that the human submitting the prompts is still doing all the thinking. We're not yet at the point where the AI can prompt itself and improve its output in a logical manner.
> We're not yet at the point where the AI can prompt itself and improve its output in a logical manner.
Self-play is widely used to train game AI, and is the "A" in "GAN"; is there any point doing it on an LLM? Especially on the ones being sold as services where people get upset if they change over time?
You really should take a look at Code Interpretor:
> And what do you feel when we make these parlor tricks more capable than us at the majority of tasks?
This seems like the logical fallacy of "begging the question" since it is far from apparent to me that they are "more capable than us at the majority of tasks."
It's certainly difficult to enumerate all the things we humans actually do.
There's a lot of stuff we consider to be "common sense", sometimes those things are used to criticise AI and sometimes they're used to criticise other humans for not knowing them, but that is a category that we don't even think about until we notice the absence.
For the things not considered common sense, like playing chess (beats all humans) or speaking/reading foreign languages (more than I can name to a higher standard than my second language), to creating art (even if it regularly makes the common sense mistake of getting the number of fingers and limbs wrong it's still better and not just faster than most humans), to arithmetic (a Raspberry Pi Zero can do it faster than all humans combined), to symbolic maths, to flying planes…
A dev conference I was at recently had someone demonstrate how they hooked up their whatsapp voice calls to speech recognition, speech synthesis trained on their own voice, and an LLM, and the criticism of the people who got the AI replies was not "you're using an AI" (he had to actively demonstrate his use of AI to conversation partners who didn't believe him) but "you can't have listed to my message, you replied too quickly to have even played it all back."
It is impossible to enumerate all the things that we humans do. However, we can enumerate all the things that we create can do. Every system we create has its limitation due to the limitations that we create in them. All systems we create cannot exceed those limitations.
We make machines that are stronger, faster, and can have much finer motor control than we have as individual abilities. No machine we have created has the dexterity that we have.
Every computational system can be analysed in fine detail to determine the limits that we have built into them. It may take an enormous amount of time and effort to do so, but we can do it. No computational system that we have built is able to exceed the limited programming we place in it.
There is an enormous amount of hype that goes on about the current generation (and future generations) of these systems, but all of them are in the abilities that we have programmed into them. They are in all essentials completely stupid (in the worst possible way - non-sentient, non-intelligent).
Every logic error that we have made in building these systems is hidden in that code. One day, those errors will come back and bite us, but there is nothing intelligent or sentient in these systems. It is our errors, for which we are responsible, that will cause those problems.
We can use them as adjuncts to our sentience and intelligence - but all they are are tools, never anything more.
However, if we cede control to these systems, we are ceding control to something that is no better than fire (a good servant - a horrendous master). After forty years, I have seen far too often, hype by humans convince other humans to cede control to the systems that humans have made and the result has been various levels of chaos.
If anything, what we need to be careful of is how humans use these systems against other humans. This is the perennial problem that we face as we build new technology.
Mostly I agree with you, but
> However, we can enumerate all the things that we create can do.
Not really, no. Even before AI, "Turing Complete" makes things extremely hard to enumerate; see Busy Beaver numbers for how small a system can be and still outside our ability to fully comprehend — needing to use up-arrow notation because exponentials aren't big enough is always good for a laugh.
With your example of "Turing Complete", we know what cannot be done and in this way, we have enumerated the things that can be done, if you like. You appreciate the humour required for the up-arrow notation - a very human quality.
You example of the Busy Beaver numbers, which was a recent interesting read, is a good example of what I was trying to point out. We have a definition and even if we cannot enumerate each number, we discuss and think about these in a rational way. At the moment, I am quite interested in Computer Algebra Systems (of which there are a variety) and I find it interesting just how limited these systems are and just how difficult it is to program into them the capabilities that humans use to solves various problems. The various discussions have been quite enlightening.
Mathematics is an interesting subject and I think shows up the intractability of ever getting that highly feared singularity.
All artificial computing systems are limited in ways we are not. Your "Turing Machine" example is one such case. The Halting Problem being a class example.
I think that far too often, we fail to recognise that what we create is not that great. We often stand in awe of the things we make without comprehending that these things are a very poor reflection of what is around us and what we ourselves are.
Every time some hype comes about these artificial stupidity systems, I look at my youngest granddaughter and see in her, capabilities that far exceed anything that we have created. Even my old buck of a goat demonstrates capabilities far, far in excess of anything we have created in all of our computational systems.
As I have said elsewhere here, we have to be careful that we do not cede control of our lives to systems that we think are more than they really are - systems that are limited, fragile and prone to failure.
> All artificial computing systems are limited in ways we are not. Your "Turing Machine" example is one such case. The Halting Problem being a class example.
You appear to be asserting that humans can tell if a loop will end, when that loop is defined so that if it does it doesn't and if it doesn't it does.
> Even my old buck of a goat demonstrates capabilities far, far in excess of anything we have created in all of our computational systems.
How so?
Not saying this is necessarily false — GPT-3 is about as complex as the brain of a rodent, so it wouldn't exactly be surprising even though the LLM only does text and completely different AI do other things — but still, what exactly do goats do that's "far in excess"?
> You appear to be asserting that humans can tell if a loop will end, when that loop is defined so that if it does it doesn't and if it doesn't it does
We can determine by looking at certain problems (The Halting Problem is one such example) what the outcome will be without actually having to execute that code. The Halting Problem is one of the simpler problems that cannot be solved by computational means, which includes all artificial computational system.
You ask [How so?] to my comment about my goat. I would suggest that to understand this you need to go and observe what happens in the environment with such beasts, whether it be a cow in a local paddock or a pet dog or cat. Take time to observe the interactions that occur and think about how little [training] is involved here.
Watch children around you, take some serious time and observe them in their interactions and I think that when you think about how we program our various artificial stupidity systems that we are still at the caveman stage in our computational systems. We have barely discovered fire so to speak.
As for [GPT-3 is about as complex as the brain of a rodent], I don't think GPT-3 has even reached a single bacterium cell state of intelligence.
I would like you to try the following: Using your index finger on your left hand, touch the tip of your nose.
Now think about this: How did you do that very simple task? When did you learn and how did you learn to do that simple task?
If you think about it carefully, the task that I describes is incredibly complex.
Now what would be required to get an artificial stupidity system (AS) to do the same task? What programming do we need to do to achieve this task? What programming was done to you to achieve that same task?
When you start asking questions like this, it becomes very clear that all of our computational systems (including all of the AS systems) are incredibly simple and not at all comparable to what we find within ourselves.
We can build very useful tools that we can use to good purpose. But no tool is ever more than a tool for us.
I suppose what concerns me about our current state of affairs is that we are far too impressed by our caveman antics. There is not a single industrial system built by mankind that comes close to the integrated control and manufacturing systems found in a single living cell. None of our communication systems come close to what is found in the various control/communication systems found in even the simplest of chordate organisms.
It was very obvious, 40 odd years ago during my engineering undergraduate days, just how fragile much of our technological base was then. It is far more fragile today and yet we appear to be enamored by our [current technological prowess] which is actually far more fragile than it was 40 years ago.
> We can determine by looking at certain problems (The Halting Problem is one such example) what the outcome will be without actually having to execute that code.
No, we definitely can't do that in general, and that lack of generality is the halting problem.
> Take time to observe the interactions that occur and think about how little [training] is involved here.
Apart from "all of evolutionary history", (though everyone agrees AI are slow learners when counting how many examples they need if that's your point?), there's continuous feedback from pleasure and pain and probably a lot more emotions that don't necessarily map onto any human qualia.
> I think that when you think about how we program our various artificial stupidity systems that we are still at the caveman stage in our computational systems. We have barely discovered fire so to speak.
I'd use a different analogy; dinosaur perhaps.
> As for [GPT-3 is about as complex as the brain of a rodent], I don't think GPT-3 has even reached a single bacterium cell state of intelligence.
I think my Roomba-clone does that: touch an obstacle, back off, rotate a few degrees, go forward again.
Now that, I'm fairly sure was programmed rather than learned (in the robot; still learned in the bacteria via evolution).
> I would like you to try the following: Using your index finger on your left hand, touch the tip of your nose.
> Now think about this: How did you do that very simple task? When did you learn and how did you learn to do that simple task?
How: a network of neurons, if I remember right about 40 deep, integrating mostly proprioceptor input as it's continuously updated when my muscles move.
When: unclear, either as an infant before and autobiographical memories, or genetic (which is arguably "not me").
I'm not sure this matters, either way though, as we do have robots which are navigating entirely by proprioception, and which again learned by training rather than being programmed.
> What programming do we need to do to achieve this task?
basically:
""" from FooLibrary import AiModel
model = AiModel()
model.learn(input, expected_output) """
With a lot of optional parameters in the constructor for different hyperparameters like "learning rate" and neurons/layers…
> We can build very useful tools that we can use to good purpose. But no tool is ever more than a tool for us.
When tools stops being mere tools — regards of this is mere perception, or when peasants and slaves revolt, or when (Australia) pest-control animals themselves become pests — we generally have big problems.
I agree the world is more fragile; I don't know if AI will help or not.
AI systems are vastly better than humans at a wide variety of tasks. Better at handwriting recognition, better at scheduling, better at playing games, better at speech recognition and transcription, etc.
I am skeptical on many of those. Speech recognition is not even close to human level. Whisper, and whatever Google uses will make a lot of mistakes on audio files that are trivial to any native speaker.
In actual tests it is beyond human level. Humans actually mishear about 1 in 20 words during transcription tests; whisper does better.
But we don’t solely rely on how well we hear since we have knowledge that allows us to correct for poor hearing based on what is being said rather than forging ahead with a nonsense transcription. Machine transcription is definitely faster and cheaper but the end product isn’t “better,” and anyone who has read it can attest to that.
> But we don’t solely rely on how well we hear since we have knowledge that allows us to correct for poor hearing based on what is being said rather than forging ahead with a nonsense transcription.
Good voice transcription AI already do that too; that's why they work best if they know which language they're operating in, as that means they can use the language to create a model of the most likely words.
I think the most recent WWDC from Apple even has a video about adding custom vocabulary for their speech engine to pick up on that covered some details in this exact topic, though I can't search right now.
Undoubtedly so but I have yet to see one that doesn't make mistakes a human would be unlikely to. It is not an easy capability to reproduce and wouldn't have been my first choice if I wanted to talk about things it can do better than people.
> Undoubtedly so but I have yet to see one that doesn't make mistakes a human would be unlikely to.
Absolutely, AI is very rarely human in its failure modes, and often has novel and exciting failure modes instead.
But, on average… or so the marketing claims… it makes fewer mistakes.
For a while, it was possible to improve upon super-human chess AI by pairing them with a human; the combination was called a centaur. Eventually the AI got too good even for that as they stopped making the sorts of mistakes humans could spot, but in the meantime, even though they were superhuman, they had failure modes that we could help out with.
Assuming the intended audience is also humans, "exciting" errors seem worse in this instance, so I find it hard to credit these marketing claims.
Seem worse, sure.
If it is or not, depends on the specific use of the transcription.
Consider "I went to Lenny's" being transcribed wrong by a human as "I went to Denny's" or by an AI as "Ivan to Lenny's".
Both are wrong, but if you get a human to check, we can be oblivious to the human mistake for the same reason it was made in the first place plus the effect where seeing text alters our perception of what we hear; the AI error being inhuman means we can spot it when the human error is imperceptible.
Well, those "actual tests" clearly don't reflect reality. This is obvious if you actually use whisper.
The question to ask is why?
The answer is that we have programmed these systems to do what we require. They cannot exceed but they fail becasue of errors that we have placed in these systems.
All of the tasks that you have mentioned have been programmed that way. It has taken human ingenuity to work out how to do this programming. The end result is a machine (non-sentient, non-intelligent) that is doing what we require.
If you look at game playing, a system was created to play Go and won and yet that same system fails to win against humans under many circumstances. The literature is there, yet not publicised for all the world to see. A result of keeping the hype in play.
If you look at speech recognition, these systems still fail when we humans work against them and yet, we humans still recognise what the machines fail at.
Just keep in mind that a tractor can move a greater amount of material than a human can, but it is still only a tool. A plane can travel faster and fly higher that a human can, but it is still only a tool.
We use these systems to augment our abilities and yet they are all limited in so many ways that we are not.
The upshot is that we can do amazing things with the things we create, but none of those things exist without us and all those things fail without us.
> All of the tasks that you have mentioned have been programmed that way. It has taken human ingenuity to work out how to do this programming.
The successful Go AI were programmed to learn; we still can't program a decent Go AI with rules humans come up with.
> The literature is there
Do you have a link? Two Minute Papers just had a video about an AI systematic finding ways to confound other AI, but I thought we'd passed the point where the best Go AI could be so manipulated by humans…
Your example of the Go AI being programmed to learn is not all that accurate for what has been achieved here. I didn't keep the link for the discussion on the confounding of the Go AI system. What the discussion covered though was that there were simple Go configurations that the GO AI failed abysmally on when playing a human - it didn't learn here.
I have spent forty years dealing with all sorts of computer systems - designing, building, maintaining, repairing, redesigning and rebuilding. One thing I have learnt over that time is that none of the systems ever built has been error free in terms of the logic entailed within them. All to often, I have seen systems that were used to make decisions with and those using them assuming that the outputs were correct or reasonable. Yet on investigation, the logic entailed in them was completely rubbish.
We make assumptions and often we do not carefully check that those assumptions are actually real. I don't trust anything I write until I have gone over it with a fine tooth comb and then I will try to document all my assumptions and this usually shows up various logic errors or conditions that I didn't think about. I don't see this happening much out in the real world.
> Your example of the Go AI being programmed to learn is not all that accurate for what has been achieved here.
What do you mean?
AlphaZero was trained entirely on self-play, and is a generic reinforcement learning algorithm. All it starts with are the rules (Chess, Go, Shogi) and a few million games later it beats — so far as I can see from a quick Google — all the humans, and most matches against AlphaGo Zero which learned the same way and which in turn beat AlphaGo Lee in every match, and that (unlike the aforementioned) was trained on examples of human matches in addition to self-play… but still learning from those examples as there's no known useful[0] set of rules that even says if a Go game is over let alone which moves are good.
There are AI which can find and exploit its weaknesses, but I've not seen anyone else suggest humans can defeat it.
> I didn't keep the link for the discussion on the confounding of the Go AI system. What the discussion covered though was that there were simple Go configurations that the GO AI failed abysmally on when playing a human - it didn't learn here.
Do you remember the name of the AI?
A bit of rummaging got me KataGo, but the humans had to use another AI to discover the weaknesses of KataGo rather than figuring it out for themselves.
And yes, KataGo absolutely does learn. The fact you can trivially stop the learning process is a feature not a bug for AI, precisely because it means any safety testing of the sort you're calling for is actually possible (albeit rather different than formal logic).
[0] pathological cases are easy — "board empty == not finished" — but not helpful.
> What do you mean?
Whose intelligence programmed this system?
> Do you remember the name of the AI?
If I recall correctly - Go AI.
They used a simple regular pattern and the system failed to beat the human. It didn't [learn] from this.
All such systems use a set of rules (whether specific or pattern based or mathematically based - there is some form of logic involved, even when using probabilistic functions), you and I can make choices based on illogical decisions - irrational decisions if you like. No computational system is capable of irrational decisions, the decisions may be surprising but of you look at the code then that option was always there somewhere, It cannot take a path that does not exist.
We can create a completely new path not previously available.
> If I recall correctly - Go AI.
I see.
Well, that's too generic to even be searchable.
> They used a simple regular pattern and the system failed to beat the human. It didn't [learn] from this.
Anything written like that would struggle against an amateur.
The machine learning based Go AI don't do that, and do beat humans.
> All such systems use a set of rules (whether specific or pattern based or mathematically based - there is some form of logic involved, even when using probabilistic functions), you and I can make choices based on illogical decisions - irrational decisions if you like. No computational system is capable of irrational decisions, the decisions may be surprising but of you look at the code then that option was always there somewhere, It cannot take a path that does not exist. We can create a completely new path not previously available.
Whatever standard I use for logical or illogical decisions, wherever I put that line, humans and AI seem to be on the same side.
We have electrical impulses flowing though messy networks, crossing tiny chemical barriers where they can be influenced by neurotransmitters; to me, that's not different enough from information flowing through an artificial neural network with weights and biases that have been automatically modified through feedback after winning and losing millions of games to say that the machine "isn't learning" — or that humans and machines aren't on the same side of "logical", at the fundamental lowest level we can't violate chemistry any more than transistors can violate physics, at the highest level the real logic of each can be random.
AI are inhuman, certainly, but still learning.
…
Just to check, you are aware that the weights and biases of an artificial neural network are basically never set by humans? That this process has to be automated?
> Well, that's too generic to even be searchable
You might want to search what articles were discussed here on the subject.
> Whatever standard I use for logical or illogical decisions, wherever I put that line, humans and AI seem to be on the same side
You can make an illogical decision. No computer program can make an illogical decision, though the pathways followed through the code may give rise to something unusual, it still follows the logic written in the program, it cannot do otherwise. Even if you have random values occurring on inputs, the code itself will follow predictable paths. If you write self-modifying code, this too is done in accordance with specific rules.
Your example of neurotransmitters and comparing to neural networks is unfortunately flawed. We developed the idea of neural networks based on a very simplified understanding of what is happening in living neural systems. Our neural network are (whatever you might think of them) extremely simple and follow very specific logical paths. We may not be able to identify what those paths will be a priori but we can analyse what has happened after the event and it will still be based on the underlying logic we put into these systems.
> Just to check, you are aware that the weights and biases of an artificial neural network are basically never set by humans? That this process has to be automated?
Yes. The automatic processes still relies of specific logical rules.
If we use only goto's, we can write all sorts of "unpredictable" programs based on random inputs. Does this generate any intelligence? If we know the code and the inputs, we do have the ability to determine what has happened.
> AI are inhuman, certainly, but still learning
I would say it this way AS systems are simply artificial constructs and the data generated is simply generated and not learnt.
He will go on saying it’s a trick. It’s a form of denial I’m seeing everywhere now when faced with something so genuinely terrifying or identity challenging you can’t process it.
Perhaps I can say you're motivated by an Oedipus complex and we can keep the chain going of ad hominem with a thin psychobabble veneer to make it appear serious.
I hate this sentiment. We might learn that human thought and reasoning are parlor tricks too once we understand them better. Anything we start to understand loses its mystery
"Any sufficiently advanced technology is indistinguishable from magic."
I know nothing about AI but it seems like we're approaching it from the other end - the human mind seems like magic and when we approximate it using technology it feels like we'll reach a moment of "that's all it is?" and refuse to believe we actually did it because we doubt ourselves.
Along the same lines, if achieving equal human rights for all humans were a trip to the corner store, the fight for AI rights is going to be like Mount Everest.
> the fight for AI rights is going to be like Mount Everest.
I used to think this and worried that AI would never have rights (see, e.g., sibling comment to mine), but these days I tend to think the fight will be very brief and heavily in the favor of AI. It could be the first time in history that rights are achieved so quickly that there isn’t much of a struggle at all.
That said, I find it extremely depressing that the default human viewpoint is “it’s a machine and doesn’t deserve rights”. Hopefully AI will have a superior system of morality to ours as well.
It would certainly be neat if we approximated the mind using technology. It's a real shame we haven't done that. And no, computer programs don't have or deserve to have rights.
>That's what ultimately depresses me about AI. It's still just a parlor trick.
I find this to be quite comforting. It means we haven't completely uprooted all of society overnight and have time to stop and think about what this new technology can do for us and what it means for the future.
AI is not a parlor trick.
AI is a branch of statistics. Nobody said that statistics must limit itself to quasi-linear models of numerical data. It was just a limitation of computational resources (initially "AI" was developed by human computers).
The trick is to get people not to associate the dictum "lies, damn lies and statistics" with "hallucinations, damn hallucinations and AI".
AI is not only a branch of statistics. Symbolic AI has nothing to do with stats.
AI he referred to was of that time.
statistics are based.
> , but they're still just that: parlor tricks
It isn't that much different than human behaviors.
People tend to repeat stuff they have seen done by our parents, sibling, friends, medias. Listen to people smalltalking in the streets, repeating the same things all over again every day. The easy success of marketing, politicians, dictators. The power of marketing and success of consumption society. Racism, bigotry, religions, addictions. All these are easily explained because people barely think. They just respond to internal and external stimulus with recipes they have been taught to follow without giving a second thought.
> That's what ultimately depresses me about AI. It's still just a parlor trick. We haven't actually taught computers to think, to reason, to be innovative. Deep learning is definitely having its day, but I suppose this too will pass unless we can unlock certain ways to make AI reliable and responsible. Or... to just start understanding it in general.
Isn't this just semantics, and the expectations that go with them, really?
If the marketing language surrounding ML wasn't so hyperbolic and sci-fi-y ("artificial intelligence"? "neural network"? give me a break!) I think we all could agree that what we can achieve now is really interesting and impressive in its own right.
Even if these models aren't on a path to some kind of "thinking computer" as you envision it, their "parlor tricks" are doing things I would've relegated to the realm of sci-fi even a decade ago, much less 25 years ago.
The reason ML took off in the 2010s is because hardware finally became powerful enough to brute force classical solutions. If anything, this lends credence to the idea that these are more than just "parlor tricks", but that with sufficient hardware we can approximate the value of intelligence. We've certainly experienced the leaps in just the last few years when more hardware is thrown at the problem. Imagine what can be accomplished in the next decade or two based on this growth alone.
We don't necessarily have to replicate the way the human brain works, but as long as the machine is capable of performing quasi-cognitive tasks, there will be immense value (and disruption) to society.
You seem to be conflating AI (in general) and strong AI. They are not the same thing at all. There have been industrial uses of AI techniques (I'm going by the definition from the 1956 Dartmouth workshop) for decades. And what's said in that video is a good example. We take recommendation modules for granted nowadays. They are applied AI.
Disappointing for sure, if you're still waiting for a sentient robot, but they do something we thought was limited to humans back in the day: going to a bookstore, and having the owner tell you "hey, I know you enjoy mystery novels taking place in England, I might have something new for you".
> That's what ultimately depresses me about AI. It's still just a parlor trick. We haven't actually taught computers to think, to reason, to be innovative.
If you think about it we've gone the other way. We're teaching/conditioning humans to think less and react more. This has only gotten worse the last few years and I don't see any shift coming soon. Humanity unfortunately seems just as simple as the algorithm, use the right inputs, in the right context and you can make most of us act just like you want.
The parlor tricks are getting better. In terms of practicality you just need to look at AI art. That stuff is good enough to replace the real thing.
In terms of actual sentience, understanding
I had chatGPT answer questions in the following order:
Each answer displayed astonishing understanding of what occurs. This was the final answer to the fifth question:1. Describe to me what happens when you throw sugar in coffee 2. Describe the same thing in terms of atoms. 3. Is this an example of entropy decreasing? 4. What if the arrow of time was reversed. Describe what happens again at the atomic level. 5. What if there were two regions in space one with increasing entropy and another with decreasing entropy. I place the coffee cup such that one half is in one region the other half is in the other region. What happens?
This answer here is anything but a magic trick. It shows deep understanding of conceptual reality by answering complex questions that are unlikely to exist anywhere else on the internet. The AI literally talked about the mysterious interface without me prompting it at all. It knows the conceptual consequence of two regions of space side by side with differing levels of entropy. The answer is not simply a trick of the next best language token.----------------------------------------------- The scenario you've proposed is purely hypothetical and quite speculative because it defies our current understanding of the laws of physics, particularly the Second Law of Thermodynamics, which stipulates that entropy in an isolated system tends to increase. Nevertheless, it's a fascinating thought experiment. Here's what might happen if you had a cup of coffee straddling two such regions: The Increasing Entropy Region: Dissolution and Distribution: On this side, sugar would behave as it does in our observable universe. It would dissolve in the coffee and disperse randomly throughout the half of the liquid. Mixing and Thermal Energy: Molecules in this region would continue to collide and interact based on thermal motion, contributing to faster dissolution and dispersion of sugar. Sweetening: The coffee in this half would become sweet as the sugar molecules interact with the liquid and dissolve. The Decreasing Entropy Region: Reforming Crystals: On this side, dissolved sugar would spontaneously start forming sugar crystals, leading to a decreasing level of entropy. Unmixing: The sugar would start to separate from the coffee and move to form a distinct layer or solid chunks, further decreasing entropy in that half. Unsweetening: Any sweetness initially imparted to this side would disappear as sugar molecules revert to a more ordered state. Challenges: Interface: The most mysterious part would be at the interface where the two regions meet. Would a sugar molecule crossing from the increasing to the decreasing entropy region suddenly reverse its course? Would there be some form of "entropy gradient" or "entropy tension" at the boundary? This part is extremely speculative and likely paradoxical. Energy Conservation: Energy conservation laws might become complicated here. The energy required for sugar to crystallize in the decreasing entropy zone would have to come from somewhere, and how this balances with the increasing entropy side would pose a significant scientific puzzle. Thermodynamic Anomalies: The coffee cup itself would be under very peculiar thermodynamic conditions, subject to two entirely different sets of rules. The temperature might behave inconsistently across the boundary, and heat exchange would likely be disrupted in perplexing ways. Taste Experience: If you were to taste the coffee, each half of the cup would offer a radically different experience, one sweet and the other likely reverting back to its original unsweet state. Perceptual Paradox: From a human perception standpoint, witnessing sugar both dissolve and crystallize in the same cup of coffee would be a profound and likely disorienting experience. Time's Arrow: This set-up implicitly questions the nature of time itself. Time seems to behave differently on each side, challenging our deepest notions about causality, sequence, and even reality. Again, it's worth noting that this scenario conflicts with our current understanding of physics, but it's a fascinating avenue for speculative thought. Such a universe, if it could exist, would fundamentally challenge our understanding of physical laws and reality as we know it. -------------------------------------------------Is it a practical answer? No. Because the question itself isn't practical. But a non-practical answer does not make this answer a parlor trick.
The entire internet is dismissing this thing as a parlor trick because LLMs fail to add large numbers. I mean come on man. You don't need to be able to do math like a calculator in order to "understand" things.
AI is not yet completely practical. That much is true. However, it is clearly No longer a parlor trick and it is getting closer and closer to transitioning into practical. When that day comes.... Good luck to us all.
It sounds like understanding, but not sure because it already doesn't consider if sugar (or water) molecules could form/exist in the alternate universe (also ignores trading off energy vs entropy changes) or should the water just crystalize in one half of the cup (why isng that considered)?. I don't think one can say it has really though about the problem. To be fair, language is probably not the tool to analyse the problem but mathematics are.
It is a bit like technobabble.
You make a claim here with "Each answer displayed astonishing understanding of what occurs." and the question you fail to ask is: Whose understanding?
The responses are based on the accumulated knowledge of humans and not machines. The systems have not thought through anything and understand nothing. A process of analysing or pattern matching the input question with the data stored retrieves an answer. But that data stored is human knowledge and human effort not machine.
If you look very carefully at the results obtained, it either contains "interesting errors" (for which an intelligent human would pick up) or it is a summation of human knowledge.
The answers still have to be tested and confirmed for rationality and applicability by humans. In other words, this is a tool like all tools created by humans.
I have seen too many examples of what are supposed to be correct answers that contained subtle and not so subtle errors.
Like every system we have ever made, Garbage in gets us Garbage out. We are the ones responsible to checking those answers and making sure that they make sense in the real world.
> A process of analysing or pattern matching the input question with the data stored retrieves an answer.
They're not just retrieving stored text like pulling the most relevant passage from a database. If they were they'd not be able to deal with things outside the training set. They couldn't write code for a custom library that was created after the cutoff (they can with a description), and they couldn't write about terms made up in the question.
I don't see why not. It's not taking a single answer from a database no, it's taking several based on probability and merging them into what it thinks we're looking for. If you learn to multiply with code to perform one task, you can then apply that knowledge for a completely different task. It may look like solving a completely new problem but the LLM doesn't even see the difference.
When you use the term "custom library" that might be you over-complicating the task. It's still just looking up function to do x, function to do y and applying it to the output. Don't get me wrong it's impressive where we're at but there's no need to exaggerate it as magic.
> . It's still just looking up function to do x, function to do y and applying it to the output.
I mean no, no it isn't.
I'm giving it info on how to construct data models with a custom library, so interacting with that is not using anything previously stored, and then giving it businesses/tasks to model as simple human descriptions.
If you tell me that something which
* Takes a human description of a problem
* Describes back to me the overall structure and components required to solve it with a hierarchy
* Converts that into code, correctly identifying where it makes sense for an address to be contained within a model or distinct and referenced
* Correctly reuses previously created classes that are obviously not in its original dataset
has no understanding or reasoning and it just regurgitating things it's seen before simply mashed together, I don't know what to say.
Frankly
> it's taking several based on probability and merging them into what it thinks we're looking for. I
Sounds pretty much like understanding and reasoning to me.
> but there's no need to exaggerate it as magic.
I'm absolutely not saying there's magic. Humans aren't magic and they can do reasoning. I'm saying it's not just looking up text and regurgitating it.
I think this is supported by things like othello-gpt, which builds an internal world model and outputs based on that.
It's difficult for me to assess how original your library is without examples, maybe I could find the exact implementation on github within 30 minutes. But I've yet to see anything that isn't just mashing together stackoverflow and git repositories to save time. I get the same answers with less wordy fluff from a simple search, but I also know where to look.
It's impressive that it knows the difference between "how many are 5 more apples than 10" compared to "how many percent are 5 apples of 10" (I don't know if it does, just assuming). But the first release also tried to reason why the weight of 1 pound of nails depends with the simple prompt "how much do 1 pound of nails weigh". That's most likely a perfect example of it mashing the classic "what weighs more, 1 pound of nails or 1 pound of feathers".
It IS just looking in a database, and mashing it with some fluff. I'm happy to be proven wrong but I need more than your word for it. My experience is that as the topic gets more niche (less data in the training set) the worse the answers I get and it starts making things up based on probability. It doesn't reason in the sense I assume you're expecting.
Have you had a look at othello gpt? https://thegradient.pub/othello/
It's a nice constrained example of a transformer learning a world model, not just looking up responses.
> It's impressive that it knows the difference between "how many are 5 more apples than 10" compared to "how many percent are 5 apples of 10" (I don't know if it does, just assuming). But the first release also tried to reason why the weight of 1 pound of nails depends with the simple prompt "how much do 1 pound of nails weigh". That's most likely a perfect example of it mashing the classic "what weighs more, 1 pound of nails or 1 pound of feathers".
Is there a formulation here that would get to a point where you'd think it's not just mashing things together? Are there elements of a simple question that would be required?
Here's a slightly trickier one for it "Which weighs more, a pound of feathers or balloons made from one pound of rubber then filled with 100g of helium?"
https://chat.openai.com/share/b841c96f-e46c-4adf-8ec3-8778ff...
Very impressive, but is it any more original than classic search engines' old trick of regular expressions to figure out if I mean the currency or weight when I ask "1 pound =" with the contexts USD or kg after "="? Does it understand the input, or are there just enough discussions in the training data to make it look like it is? I'm not convinced it's not the latter.
It uses context to figure out we're trying to convert something to something else. Then it adds all those numbers up. Taking helium into consideration is no doubt interesting, but they've also polished that task since that was the common critique they got so very wrong with the first release (which I mentioned they had fixed). I'm not qualified to assess this part of the answer;
> "If the balloons displace more than 100g of air when filled with helium, then they would effectively weigh less than if they were left empty. If they displace exactly 100g of air, then the balloons would have the same weight as if they were left empty."
I don't know enough to understand how much 100g of helium is and how it behaves. And it doesn't try to explain it to me, it mentions it then takes the easy route assuming it's a trick question. What does that tell you? I guess there are similar discussions around and it gives me the summary. Why doesn't it tell me how much air it displaces under what circumstances? Temperature etc, it should be easy if it's not just a simple discussion on a random forum. A conversion regex could do it.
This comment[1] has a very impressive example. But anything I'm qualified to assess has mostly been meh. If the fix is better training data does that mean it's reasoning or regurgitating? The mistakes it makes are what tells me how it works, not when it tricks me that it's correct. To me it's a very well polished search engine summary.
If you've not looked at it I really recommend othello gpt. That is an experiment explicitly designed to tackle this kind of question, has it just seen enough moves that it knows what should come next?
> Why doesn't it tell me how much air it displaces under what circumstances?
You can ask it and it'll answer.
> If the fix is better training data does that mean it's reasoning or regurgitating?
More training data helps with things you can just bring to the fore, same as a lot of learning. More useful training data though can also help reasoning, which makes sense - deliberate training of people helps improve their logical reasoning capabilities. I know that doesn't guarantee that's what LLMs are doing but humans benefit significantly from both more teaching and better teaching.
> Very impressive, but is it any more original than classic search engines' old trick of regular expressions to figure out if I mean the currency or weight when I ask "1 pound =" with the contexts USD or kg after "="? Does it understand the input, or are there just enough discussions in the training data to make it look like it is?
I'd be interested to know any requirements around this to clearly show the difference. I tried asking what if I filled a balloon at a childrens party with a gas made of atoms that have 1 proton and 100 neutrons: https://chat.openai.com/share/71224df4-5c6c-45f7-88fd-eec316...
(tl;dr: "In the context of a child's birthday party, introducing such a balloon would be a grave mistake.")
It identifies:
* Whether it would float or not and why * That it would be radioactive, and likely types of radiation from it * What that would mean to the balloon * How people would react and the likely consequences of releasing it in a room of children
This is an element that does not exist, in a setting where nothing like this has happened before, with details ranging from types of decay, consequences and human emotional reactions to something like this. Yes, there are real things you can use as a base (e.g. how do people respond to events that kill people), but I feel it's an example of where it's beyond a search engine summary.
> If you've not looked at it I really recommend othello gpt.
I skimmed it and read the conclusion, and it looks interesting, will take a closer look when I have time.
The prompt covers a subject that goes completely over my head so I can't tell how well it reasons. I don't know what 1 proton to 100 neutrons means, but I gather it's radioactive. I don't think it's far fetched that it draws the same conclusion from the training set because to you it seems obvious, and is probably well known to anyone who knows the subject. Kind of like it would understand that "hotter than the sun" is super hot, can correlate to different melting points. But I wouldn't say it understands the concept of temperature. Given the right prompt it might give you the impression it does.
The feelings of the scenario reads like any PR comment after a tragedy. "We feel shock and disbelief" and so on. The scenario being hypothetical doesn't change that since it's probabilities. It acts just like you'd think it would. The earlier example with the helium balloon is similar, it assumes a human context and not the form, and environment the helium is in. True intelligence might not even consider the presence of atmosphere as the norm. "It has no weight outside of your human constraints" would be novel.
Lets say it has odd numbers between 1-9 in the database. Given the prompt 2 and 8 you will get back 1,3,7,9, sprinkled with some natural language and we get the impression it's intelligent.
Are you saying it understands the effect the neutron to proton ratio has, as opposed to just comparing the vectors closest to your prompt that it builds the answer from? Being tested on new and hypothetical examples only means it will be further from the vectors but still close enough to give us the impression it understands the subject. If the training data didn't include the words neutron or proton it would have no idea where to begin.
In my first comment that started this chain I said:
> I don't see why not. It's not taking a single answer from a database no, it's taking several based on probability and merging them into what it thinks we're looking for.
I don't think even this latest answer is any proof of anything other than that. Are you claiming there is? And what are you claiming is happening?
> I don't think even this latest answer is any proof of anything other than that. Are you claiming there is? And what are you claiming is happening?
I'm claiming that it's reasoning through the problem.
> True intelligence might not even consider the presence of atmosphere as the norm
I hugely disagree, that's not how an intelligent human would answer the question, and if it did this people would be complaining that it clearly doesn't understand the human context in which the question was likely asked.
> I don't know what 1 proton to 100 neutrons means, but I gather it's radioactive.
It would be hydrogen, but a type of hydrogen that doesn't appear in nature. It's a deliberately absurd example so that it's not in the training set. Its answers are different if the question involves tritium which while radioactive has a moderate half-life and wouldn't immediately pop the balloon.
> I don't think it's far fetched that it draws the same conclusion from the training set because to you it seems obvious,
Only because I can reason through what would happen, not because it's something I've seen talked about before.
To figure out what it would do, it cannot rely on an explanation elsewhere, it needs to first identify that the ratio of protons to neutrons is extreme. Then it needs to understand that this typically results in particular kinds of radiation.
It has to then use that information to consider how that would interact with the material of the balloon (and that this is important).
It has to use that information to consider how it would affect people, and what their reactions would be both before and after it explodes/pops.
This is multi-step reasoning through an issue that involves pulling together common expectations, physics and how humans react.
Here's a statement in it that shows to me more than just pulling a few answers together
> Balloon Behavior: Instead of floating up like a helium-filled balloon, this balloon would drop to the ground because the gas inside is denser than air. This might surprise the attendees, and curious children might approach or pick up the balloon, further exposing themselves to radiation.
-
> The feelings of the scenario reads like any PR comment after a tragedy. "We feel shock and disbelief" and so on.
Those are typical things, which is not surprising, but it is also clearly linked with the question. You have to understand how out of context this would be.
> If the training data didn't include the words neutron or proton it would have no idea where to begin.
Fully rediscovering what took humans many years to do off-the-cuff is an outrageously high bar.
What features of a question would you look for to identify whether it's "taking several answers from a database and merging them together" or performing some reasoning? I've asked a few times but don't understand what you're expecting.
> Fully rediscovering what took humans many years to do off-the-cuff is an outrageously high bar.
> What features of a question would you look for to identify whether it's "taking several answers from a database and merging them together" or performing some reasoning? I've asked a few times but don't understand what you're expecting.
You're misunderstanding me, I'm setting no bars, and I have no threshold where this changes. We humans are also just looking things up in our database and doing deductions. We do some computing on urgency as well, like how when we hear a bang our mind goes for danger first before realizing it was harmless, but very similar to what these AIs do. Probabilities and experience. Fresh and novel ideas are very rare in humans as well, and not something I demand before I would consider someone a human.
I did however give you an example that would surprise me, if it considered mass and environment in a way that proves that it understands the problem for what it is. If it told me weight is a human construct and requires gravity/movement and how it depends. An intelligent human doesn't necessarily answer the question it is asked in the way it is phrased. It identifies and irons out misunderstandings, assumptions and other details important to correctly understand the problem, and may even rephrase the question to give a proper response. That would show me a deep understanding of the problem and maybe freak me out a little, but only if the hallucinations are gone and those can be difficult to spot.
This is, just like us, performing calculations and database look-ups. It may feel like it's doing something else but it's not. What would happen if we leave the weights as they are but switch the words? It would give us complete gibberish, but it's no less correct than it was before and it's not even giving us different answers, only the translations to language get distorted. Most people would call it stupid and pointless even if the only change is our interpretation of the answers.
I'm sure Hiroshima, Fukushima and other dangers of radiation is in the training set, as are all of the other steps you mention, it goes round and round testing the numbers based on training. Remember how this chain started, you claimed:
> They're not just retrieving stored text like pulling the most relevant passage from a database. If they were they'd not be able to deal with things outside the training set.
To which I simply replied:
> It's not taking a single answer from a database no, it's taking several based on probability and merging them into what it thinks we're looking for.
I read you (correct me if I'm wrong) as giving this way to much agency. To change my mind that it's doing something unexpected I would ask for logs on the calculations it does, and be able to correlate that to the training set. I have to be able to falsify the conclusions I'm asked to make. I know some people claim we don't understand these algorithms but I assume that's just hyperbole and with the correct measures we could follow every step.
If there are things there which I can not trace I would be very impressed, and honestly a little afraid. They are not trained for every single task, but approximations based on similarities have proven to be very capable even when we think we're out of context (we're not, it doesn't understand context and doesn't care, but neither do most humans).
Perhaps we've been talking past each other then. I've been trying to show that these things can do reasoning, and my upper benchmark is "like a human". If you're starting from "these things might be doing what humans are doing / capable of performing similar tasks" then we're largely aligned.
The further question I still find interesting though.
> I read you (correct me if I'm wrong) as giving this way to much agency. To change my mind that it's doing something unexpected I would ask for logs on the calculations it does, and be able to correlate that to the training set. I have to be able to falsify the conclusions I'm asked to make. I know some people claim we don't understand these algorithms but I assume that's just hyperbole and with the correct measures we could follow every step.
This one is tricky. We know exactly what they do. Interpreting that is very hard though, they've a big pile of mathematical operations with billions of magical constants and it... works. We can see exactly what they do but if I could see every synapse firing in your brain I'd still not be able to understand how it works in a useful manner. So we understand them obviously, but at another level we really don't.
> I'm sure Hiroshima, Fukushima and other dangers of radiation is in the training set, as are all of the other steps you mention, it goes round and round testing the numbers based on training.
Just to be clear here, there is no recursion other than when you add more text. There is not an algorithm saying "identify parts X, then look in database Y, now summarise...". They're trained essentially to just predict the next word given some text. There's some later training to make them more conversational. The capabilities you see are just a consequence of that.
Othello GPT is show just moves. It ends up building an internal model of a board.
> I did however give you an example that would surprise me, if it considered mass and environment in a way that proves that it understands the problem for what it is. If it told me weight is a human construct and requires gravity/movement and how it depends. An intelligent human doesn't necessarily answer the question it is asked in the way it is phrased. It identifies and irons out misunderstandings, assumptions and other details important to correctly understand the problem, and may even rephrase the question to give a proper response. That would show me a deep understanding of the problem and maybe freak me out a little, but only if the hallucinations are gone and those can be difficult to spot.
Let's try and investigate that then, that sounds interesting. I'm not sure I understand myself what you mean that weight is a human construct (it explains the difference between weight and effective weight in the answers to me off the bat, that's the only real difference). Perhaps this is too simple, the answer is quite straightforward.
> If it told me weight is a human construct and requires gravity/movement and how it depends
I asked "Which is the most, a pound of feathers or a pound of helium?" with largely just your statement as the system message and got
>The question seems to contain an intrinsic confusion. When we discuss weight, both a pound of feathers and a pound of helium would weigh the same - a pound. The difference, though, comes in their volume and density. A pound of helium would take up a lot more space than a pound of feathers considering the density of helium is lower than the density of feathers. If you were implying which would be more in terms of volume, then a pound of helium would be significantly more than a pound of feathers.
> However, I might be wrong if we take into account that helium, being a gas, is usually measured in terms of its volume at standard temperature and pressure, rather than by weight like solid or loose materials such as feathers. Also, the weight of a pound can vary slightly depending on where on Earth it is measured due to differences in gravity. However, these factors don't fundamentally change the answer to the question as it was posited.
Perhaps instead you could give me a short question and the kind of answer that would surprise you? I know this thread has gone on some time, but personally this is interesting to me. If you wanted to shift off from hn, feel free to drop me an email, I have a vested interest in understanding how people view LLMs.
> You make a claim here with "Each answer displayed astonishing understanding of what occurs." and the question you fail to ask is: Whose understanding?
The answer is obvious. The LLM is understanding the concepts. The last question was unique. The resulting answer was also unique.
It was not a "retrieved" answer. It was a unique answer. A correct composition of several underlying concepts. A correct composition can only be formulated if the machine had correct understanding of each concept and how they relate to one another.
This thing understands you. It wholly owns this understanding. It is not regurgitating knowledge. It is inventing new answers.
Wake up man. I had the LLM invent 6 regions and heat the cup of coffee to plasma levels of heat. The answer and composition of concepts was remarkable.
You're calling it a parlor trick because of subtle errors? Bro. Come on.
> The answer is obvious. The LLM is understanding the concepts
Who created the LLM? Whose understanding underpins the LLM?
Certainly not the LLM.
> This thing understands you.
Does it? Or is this a result of the intelligence of the human beings involved in building the LLM?
> I had the LLM invent 6 regions and heat the cup of coffee to plasma levels of heat.
Did the LLM actually invent anything? Or was this result directly based on you and your intelligence with the recorded knowledge of all the human sources involved in the solution?
> You're calling it a parlor trick because of subtle errors?
I haven't called it a parlour trick. All I am saying is that there is no intelligence in these systems. Human intelligence built them, but these systems in and of themselves have no intelligence.
We do of course build many intelligent systems all the time, they are called children.
>Who created the LLM? Whose understanding underpins the LLM?
Who created you? Whose understanding underpins you? Asking these questions about you is as irrelevant as asking it about the LLM.
Just because books, educations your teachers, the internet and your parents and the environment shaped everything you know doesn't preclude your membership into the category of things that are capable of understanding.
>Does it? Or is this a result of the intelligence of the human beings involved in building the LLM?
It does understand you. The intelligence of human beings who built it aren't directly involved as it was trained on external data.
>Did the LLM actually invent anything? Or was this result directly based on you and your intelligence with the recorded knowledge of all the human sources involved in the solution?
Does a human actually invent something or is it directly based on recorded knowledge?
You're asking irrelevant questions. Humans do not create things out of thin air either. Humans also invent things by composing existing knowledge to form concepts. The inventing that LLMs can do is equivalent in totality to our understanding of the word "invent"
>I haven't called it a parlour trick. All I am saying is that there is no intelligence in these systems. Human intelligence built them, but these systems in and of themselves have no intelligence.
Totally false. Not only are you wrong but experts in AI including the father of modern AI disagree with you completely and utterly.
If I copied your brain and replicated exactly that brain is "from human intelligence" but that copy of your brain is still an intelligence independent of it's origins and where it got it's knowledge.
>We do of course build many intelligent systems all the time, they are called children.
It's like you're eating your own logic. We also build intelligent systems called LLMs. Same concept.
You miss the point of the questions in relation to the LLM. However, your questions are important in relation to the fundamental difference between humans and what they create.
Let me put it this way: Humans are started with a single cell that eventually grows into an extraordinarily complex entity. If you look at new born babies, they have a capability of learning and exploring that we do not see in any artificial construct that we make.
Our artificial constructs have to be essential fully developed physically before we can then start the process of programming them. Humans have a capability to learn as they develop. We see this occur in all living things.
There is a fundamental insurmountable category difference here between living things and artificial constructs that we make.
Any appearance of understanding is based on the logic that we program into these artificial constructs. They cannot exceed what is programmed into them. Interestingly, living organisms can often exceed that. I think that all programmers should undertake a study of living things to gain a greater appreciation of what we do and just how simplistic are the things we do. That is a particular philosophical point of view that I hold.
I think that your appreciation of what we do and the constructs we create is not in accordance with reality. Not that this is particularly strange as far too many people have a much higher view of our technological prowess, especially when comparing to what has gone before. Starting from my undergraduate engineering days and the ongoing study of engineering and technological history, it has become quite clear that we are often today, quite ignorant of just how technologically advanced previous eras were in all sorts of different areas. There are plenty of research groups that are researching how previous generations were able to do things that we do not know how to do today.
> Does a human actually invent something or is it directly based on recorded knowledge?
Here, we do know that there is at least three ways that invention can arise. Logical progression on recorded knowledge, imagination as to how to solve a problem (thinking outside of the box), observation of the natural world around us.
> You're asking irrelevant questions.
For you to say that the questions I asked were irrelevant shows that you have limited yourself in your pursuit of knowledge and understanding.
> Humans do not create things out of thin air either.
When thinking outside of the box, they do. But I suspect that you may not appreciate this particular point.
> Humans also invent things by composing existing knowledge to form concepts.
As pointed out above, this is one mode of invention.
> The inventing that LLMs can do is equivalent in totality to our understanding of the word "invent"
Here, I disagree with you. But that is very likely to be a philosophical/metaphysical difference between us.
> Totally false. Not only are you wrong but experts in AI including the father of modern AI disagree with you completely and utterly.
Do you understand that you have devolved into a fallacious argument here? This is a seriously flawed fallacious argument on your part. The problem here is that your argument assumes that these [experts] you are referring to are correct, when you have not demonstrated that. Nor have they. There are many experts and others (all highly intelligent and talented people) who for all their intelligence and talent are just wrong. This has been shown to be the case many many times throughout the last century (let alone before that) when our understanding has changed because some little known person has come up with anew idea. One of the best examples here would be someone you would know of - Albert Einstein.
Now when you say
> If I copied your brain and replicated exactly that brain is "from human intelligence" but that copy of your brain is still an intelligence independent of it's origins and where it got it's knowledge.
Here, you have a problem. What is the difference between a living brain and a dead brain? In a single instant, we go from life to death and yet we don't know at this point in time what that difference is. There are lots of experiments being done today which are trying to study if there is a non-physical aspect to intelligence and free-will. Different experimenters in the same team have quite different interpretations of what the data means.
Do be so quick to assume that you know, when the researchers who study this can't agree.
> It's like you're eating your own logic. We also build intelligent systems called LLMs. Same concept.
Not at all. Intrinsically different and there is a vast categorical difference between children and our artificial constructs. From your comment, is it a valid assumption that you do not have children of your own or grandchildren of your own or even pets?
Let me ask a question, what is your background in building systems that augment human capability? If you have been involved in building LLM's, let me know.
It's generating code for a brand new library based on explanations from me, it can write poems about the current news headlines and it can answer hypotheticals with words I've made up. I agree it cannot be just looking up stored answers.
Gpt Othello is a good discussion about this that's more constrained too.
Talking about mixing coffee with sugar/milk and reversing the arrow of time (ie. entropy) is a widely talked about example.
That said, I agree with your arguments and conclusions. Humans are not able to do this kind of reasoning until they have had some education on the topic - and when they reason they get feedback to reinforce their conclusions if others agree or they get a more rigorous proof. We could do the exact same thing with an LLM. Would that potentially make a reasoning AI? For me the jury is still out, albeit I think we need better algorithms for it to be feasible.
> Talking about mixing coffee with sugar/milk and reversing the arrow of time (ie. entropy) is a widely talked about example.
But the interface is unique. That is not a topic widely talked about at all.
Question:
Answer:Instead of two regions add a third region where entropy of the particles entering the region will always stays the same no matter what. Place the coffee cup such that it is in all 3 regions equally. Describe atomically what happens.
In the Atomic Realm:-------------------------------------- In this fictional universe where regions have varying entropy behavior, the coffee cup sits at a tri-junction: a third of it is in a region of increasing entropy (Region A), another third in a region of decreasing entropy (Region B), and the last third in a region where entropy remains constant (Region C).
The Tri-Junction:Region A (Increasing Entropy): Sugar molecules randomly move throughout the coffee, driven by thermal motion. They strike water molecules and dissolve over time, eventually reaching a state of homogeneous distribution. The entropy of the system in this portion of the cup increases as expected. Region B (Decreasing Entropy): Astonishingly, sugar crystals spontaneously form in the coffee. Sugar molecules collide with water molecules and join an ever-growing crystalline lattice. Over time, the sugar becomes a distinct entity in the coffee, and the entropy of this system decreases. Region C (Constant Entropy): The sugar neither dissolves nor crystallizes. Instead, sugar molecules and water molecules co-exist in a dynamic but unchanging equilibrium. At the atomic level, every dissolution event is perfectly balanced by a crystallization event, maintaining constant entropy.
This isn't a parlor trick regurgitation. It's displaying understanding of the concepts by speculating on new concepts via composition of several other concepts.Molecular Exchange: Some molecules, through Brownian motion, will wander across the boundary lines dividing the three regions. Entropic currents from Region A and B cross into Region C and are balanced out, maintaining a zero net change in entropy in that region. Special quantum-entangled pairs straddle the boundaries, existing in superpositions of increased, decreased, and unchanged entropy states. Entropic Flux and Balance: A constant entropic flux is established between the three regions. The entropic gradient tensor now has a third component, and the entropic flux density vector also adjusts to a tri-vector to balance out the exchange between the three regions. Localized Time Anomalies: At the atomic level, it's as if time runs forward in Region A, backward in Region B, and is somehow "static" or cyclical in Region C. Particles appear to remember both their past and future states in Region C, generating a unique form of atomic "temporal coherence." Causality and Quantum Superposition: Despite these anomalies, causality is preserved. Particles that cross into different regions adapt to the local entropy behavior and exhibit quantum behaviors that align with the local direction of time. In essence, the tri-junction becomes a hotbed of bizarre but mathematically consistent phenomena, a place where the fabric of spacetime and the fundamental laws of thermodynamics seem to negotiate a delicate truce. As for the cup of coffee, one-third of it will end up well-mixed, one-third will have sugar crystals, and the last third will remain an intriguing paradox of thermodynamics. You might call it the most scientifically fascinating cup of coffee ever brewed. -------------------------------------------------Of course it's not regurgitation. But it is combining related concepts and I'm sure that somewhere in its training input, there is plenty of talk about increasing, decreasing, and constant entropy.
However, it does not really say much. In particular the "Localized Time Anomalies" section for region C seems to be a hallucination and from there on, it really sounds like something a science writer, without the proper education, could have written.
It sounds good, but from an experts view, it is at best a popsci introduction to the subject.
>it really sounds like something a science writer, without the proper education, could have written.
It just flew past your head didn't it? You remarked on the similarities between an AI and a human in order to prove the triviality of AI.
Getting AI to the ability of a science fiction writer was an impossible feat just 3 or 4 years ago.
That's a fun example. Here's one I posted elsewhere about filling a balloon with a gas made of 1 proton and 100 neutrons at a kids party.
https://chat.openai.com/share/71224df4-5c6c-45f7-88fd-eec316...
Not really sure how you can consider ChatGPT a parlour trick. It has been around a relatively short time but for me its replaced a large proportion of my Google searches already. I do not see how its utility can be denied whether it reasons or not (whatever that means).
In fairness, Page Rank and indexing are also… well, perhaps too big for a parlour, but they're (very good) tricks.
Is ChatGPT playing a trick on us by mimicking the sentience of the humans whose writings it ingested, or is the trick that by doing so it began to actually think like us and so simulates a conscious mind within?
I lean towards the former; but we don't know what sentience even is yet, so we can't prove it.
This is why they've called it AI winter the past three times.
It's a season.
Like the seasons, the cycle repeats.
This time it's nuclear winter.
Examples: Graphics User Interface. The iPhone seated mobile compute as a permanent fixture, it and Android bringing internet and computing to billions of humans for the first time. Not just the wealthy industrials. Or IBM DB2 for SQL. Or Ford's Model T. Or the Gutenberg. None of these were the first first. Maybe even on the n-th iteration 2 decades after coming out of DARPA or global university research, something is just ready for commercialized prime time.
Computing was solved by Lady Ada and Babbage. It's all electrical engineering, software, fabrication, productization, mfg, displays, sensors etc. after that.
Semiconductor was a big deal in making it happen. But that's almost besides the point of the theory of compute if it can be solved in other novel ways using alternative material science.
What a tired and lame take.
100m people used this latest iteration. It’s hardly a winter.
If there's anything the last few years has taught me is that hype is rarely any indication of common sense. I'm always surprised what reaches success and what doesn't. Truly revolutionary ideas are ignored and not understood while stupid but polished ideas are booming.
One example of that is the crypto space. So many actual good for humanity implications, but the "killer app" that made the news was fucking NFTs. I'm trying really hard not to come to the conclusion that humans are just mindless zombies but it's getting harder every year.
Once we reach a billion DAU you can be sure it will just be "Original Memes Tailored Just For YOU"™ instead of figuring out the logistics for solving world hunger. Mark my words.
The crash is gonna be wild when people realize all these companies are blowing smoke. We did get Google out of the last crash though I guess.
What a tired and lame retort.
You say all that. Meanwhile teenagers are finding love on character.ai.
There is a famous Dijkstra quote, “The question of whether a computer can think is no more interesting than the question of whether a submarine can swim.”
Do the intrinsic properties of the system really matter at the end of the day if it performs as well as we do at some task? Heck they’ve been doing many things better for decades, but those are the types of tasks we take it for granted that a machine should be able to do. Solve differential equations. Play chess. But now computers are doing “human” tasks competently. Writing creative fiction. Generating graphical art.
We don’t have a good working definition or metric for intelligence. Surely it is not a monolithic property. Animals exhibit many traits we associate with intelligence. Some of the stuff GPT 3+ generates sounds pretty intelligent. It is the type of things an intelligence may have produced, because it was trained to do just that. If we look at intelligence as a cluster of traits, or behaviors, I think we are surrounded by intelligence - human, artificial, or otherwise. Doesn’t have to be an AGI to fall in that category. It doesn’t even have to be particularly impressive.
I wrote about my definition of intelligence earlier this month: https://tildes.net/~comp/194n/language_is_a_poor_heuristic_f...
According to my definition, intelligence is actually all around us. We are blind to it because we focus only on how intelligence manifests in humans (defined by by our specific social organization and biological senses), and then use that as a benchmark to judge every other thing in the world.I have a definition of intelligence. [...] Intelligence is prediction. In the case of intelligent living processes ranging from single celled organisms to complex multicellular life, intelligence arises from the need to predict the future to survive and reproduce. More intelligent organisms build more elaborate models of the world using better developed senses in order to do so. Humans model the world primarily with language, which allows us to share our models with each other, across both space and time! Without language, it is extraordinarily more difficult to communicate complex abstract thoughts. As a side effect of our high level of intelligence, our wetware is capable of modeling things outside of language, such as mathematics. [...] In general, I think we need to stop equating intelligence with consciousness, agency, moral value, and living things. These are entirely different concepts, and, as a species, we are blinded by our egotistical identity of being smarter than the other animals.A less socially-charged definition of intelligence would make it easier to compare intelligence across living and non-living processes, though it would not be "popular science" useful for ranking humans.
I don't agree with your definition at all.
2 people want to kill each other. The one taking the first step is the intelligent one because according to your definition he was better at predicting an outcome than his opponent.
The real world is more complex than that and there are multiple options where both survive, or letting your opponent live and killing yourself because his life is more beneficial to humanity and so on.
Any organism can survive, but for most (including us) that implies a selfish outlook but the most intelligent people I know or heard of never even consider their own ego.
Maybe I wasn't clear enough. The definition of intelligence I propose is wholly distinct from human prosocial values like cooperation.
This makes it useful for judging these properties across living and non-living intelligent processes, such as bacteria, ants, plants, dogs, LLMs, etc. It is not a useful definition for judging the value or "goodness" of human beings within society.
I'm arguing that intelligence (as prediction) is simpler than we often presume, not a mystery at all, and a basic building block of complex life. We happen to have a lot of it.
I'm not coming from a moral standpoint either, just giving an example of how correct prediction may lead to consequences we would call retarded. Predicting resources being scarce may lead to over-consumption which leads to the extinction of everything. You could argue an intelligent organism would predict that and adapt but we're still trying ourselves.
There are plants that slow their own growth to share resources if their neighbor is of the same species, and vice-versa if it's a different species. That could be called intelligence in a sense, but it's not as simple as just prediction there's a social aspect and a long-term goal. But is it even conscious and aware of what it's doing or is it just the traits favored by evolution. Is agency part of your definition of prediction or is it enough to just react to the surroundings?
Intelligence is much more complex than a single trait. Being good at prediction is just that, being good at prediction.
Your definition of intelligence has been around for millennia and is part of the pantheism concept.
Sure, it's neat that it can do those things... still I don't think that I'm exactly ready to toss aside books written by humans in favor of AI-generated alternatives.
Writers have started using AI-generated ideas to help them write books.
Same with code. Sure it might not write your product code from start to finish with no help but it will speed up your dev speed significantly for certain tasks. Just because we haven't reached singularity doesn't mean what we have now is useless and putting it down as a "parlor trick" as grandparent said seems to me very unwise.
Ad copy has served a similar purpose but nobody would claim ad copy has (or could) supplant creative writing because of that.
Ad copy is (a form of) creative writing.
OK. Literary writing.
"Collaborative filtering", "using past activity and stated preferences to guide us"
I can't think of areas Amazon does worse in today than these two areas.
Reviews are so untrustworthy they are just noise.
Those coffee beans I ordered a month ago? I have to go back to my orders page, search (and for some reason their search is dirt slow) - then wade through unrelated products to finally find them.
But oh, hey, you just bought a vacuum cleaner? I know what you would really like! MORE VACUUMS!
I guess they innovated on those two areas in 1998, and then since then the only thing they did was remove the `review-count-rank` sorting option so their AI can suggest Amazon Brand products.
hey, you just bought a vacuum cleaner? I know what you would really like! MORE VACUUMS!
This is a common refrain, but I would be amazed if it weren't backed up by data and I anecdotally fit into the mold. If I buy a product that I really like, then I'm likely to buy it again to give as a gift to friends. If I buy something like a a vacuum and I don't like it, then I'm likely to buy a different model that better fits my needs. It seems silly when you aren't interested in buying another one of whatever item you bought, but there are multiple legitimate reasons that people would want to.
You'd buy a vacuum cleaner as a gift for a friend? I find that incredibly hard to believe but if it's true, it's very strange behaviour.
I think the comment was that if they buy a vacuum and don’t like it, they return it and buy a different one. So if that happens 20% of the time, that next month has a much higher chance of buying “another vacuum” than the average month.
Also, I bought my parents a Roomba because I liked mine, so that can happen too.
Several years back a bunch of us pet-owning friends wound up round-robin gifting the same model SpotBot carpet cleaner to each other, because it was something we all agreed made our lives better. Nothing weird about it at all.
The fact that you had to preface this with you and your friends being pet owners kind of does make it a special case. You being suggested this vacuum cleaner by Amazon after buying it also doesn't do anything for you in this scenario because you bought the same one for your friends, so the recommendation wasn't needed. I think people are losing sight of the original argument which is that Amazon's recommendations can sometimes overemphasize recent purchases with very little context around what the items are.
Yeah, but everyone is "special case" in retail.
I've done it before: shortly after I bought a robot vacuum; I gave the exact same model as a birthday gift
A vacuum cleaner would be a great housewarming gift, for example.
It truly is a gift that sucks
I did, not a month ago. Nothing strange about the process whatsoever.
Posting to confirm, having seen back-end sales numbers from (admittedly much smaller) vendors, the correlation between $just_bought_thing and $will_buy_another is very, very high, across pretty much every category I cared to look at.
But surely if you just bought a <Roomba> and like it and intend to buy it again as a gift then you don’t need an advert for Roombas - (and showing you such an advert followed by you buying another Roomba is making the advert look more effective than it was) - and you definitely don’t need an advert for Bissel or Dyson, and you definitely don’t need a dozen adverts for Amazon PLINGBA BEST VACUUM, TYBCHO VACUUM EXPERT, DAOLPTRY VACUUM CLANER etc etc?
Showing someone ads for products in a category they recently purchased from is one of the most effective things a store can do, in terms of focused advertising driving sales. We can wonder why, but the data is exceedingly clear.
Well if those are the reasons then why not code it directly?
Ask the user if they liked the product and would likely gift it in the future and add the item to a "Gift ideas" list.
If the user returned then start suggesting immediately an alternative.
Blindly recommending the same thing just because there's a correlation seems stupid. Would be nice to at-least have a nicely visible button to stop recommending this item since I'm done purchasing anything similar for the next couple of years.
> But oh, hey, you just bought a vacuum cleaner? I know what you would really like! MORE VACUUMS!
Amazon did literally this to me just a couple days ago.
I'd bought a vacuum cleaner on Amazon a week earlier, and, when doing an Amazon checkout of a cart with one item, it threw up a list of consumable items that I'd bought in the past, to possibly add to the order... including another one of those $120 upright vacuum cleaners.
(Maybe they have data that says this makes them more money, even though I'd guess it might hurt customer confidence in the site.)
I don't know how the recommendation engines work, but if there are dollars there, i could understand this to be:
Amazon: Sellers! Do you want to advertise to customers interested in vacuums ? Seller: YES! Take my money!
(Amazon proceeds to uprank vacuums to customers with prior vacuum sales)
>But oh, hey, you just bought a vacuum cleaner? I know what you would really like! MORE VACUUMS!
I have worked on the same recommendation systems. It's also the most often oncall issue. The problem is mostly due to lag in event processing (especially orders).
Amazon had fewer items back then and the bar and complexity was lower too.
The fascinating thing is that this guy has literally 100s of billions in pocket. He ALSO has all the time in the world. He can do anything he wants. Anything. The dude is smart, genius, visionary. He even precisely knows what needs to be done. And what is he really doing? It is simply breathtaking.
Article has no body for me. Site appears to use an iframe whose src expects the Referer header to be sent, but I have `network.http.referer.XOriginPolicy = 1`set in FF about:config to reduce cross-origin leakage, so no Referer is sent.
Great point. It is using cloudflare stream.
https://developers.cloudflare.com/stream/viewing-videos/usin...
Open to any recomendations for alternative as i too am quite displeased with the state of such things. But still prefer it to YouTube.
I believe the media server is set to reqire referer to prevent embeding on alternative origins.
Seems to be https://developers.cloudflare.com/stream/viewing-videos/secu.... Probably can be disabled and replaced with short-lived signed tokens. Though perhaps CloudFlare could have used [more modern iframe restrictions][1].
[1]: https://w3c.github.io/webappsec-csp/#frame-ancestors-navigat...
I notice that a lot of really successful people speak quite quickly early in their careers. It's almost as if their own voices can't keep up with the excitement they have about their product/idea and they haven't quite honed in the outward presentation of it.
Could just be the Adderall.
Is Jeff still able to convey such a technical response anymore?