We should ask, when will AI make a discovery on its own? For instance, computers should be able to understand numbers, and run analysis on numbers. Computers have complete access to every fact that humans know about numbers. So numbers should be the first place that we should expect to see genuine innovation from AI. This is a simple test for the moment that AI is able to make original contributions to our society: when can AI come up with a new thesis about numbers, and then build an original proof, something that can be published in the major, peer-reviewed math journals.
Until AI can do that, we have to admit that it's not really aware or sentient or any of the other more ambitious things that have recently been claimed for it.
Can AI teach us anything new about the pattern of prime numbers?
Can AI develop an original proof for the shape of shadows in high dimensional spaces?
Can AI creatively prove a new limit to mathematics?
There are 2 researchers in AI who deserve more attention: Kenneth O. Stanley and Joel Lehman. They wrote a great book: Why Greatness Cannot Be Planned. They look at the limits of utility functions and explain the importance of novelty. As an antidote to some of the hype around AI, I strongly recommend this book:
>This is a simple test for the moment that AI is able to make original contributions to our society: when can AI come up with a new thesis about numbers, and then build an original proof, something that can be published in the major, peer-reviewed math journals.
>Until AI can do that, we have to admit that it's not really aware or sentient or any of the other more ambitious things that have recently been claimed for it.
We have to admit no such thing, that is an absurdly high bar. The vast majority of humanity has not produced an original mathematical proof worthy of being published in a peer-reviewed math journal, and realistically it isn't possible for the vast majority of humanity to do so. Nevertheless, we are essentially all sentient/aware. "If it can't generate new and novel math that can pass peer review, it's not aware or sentient" is a moving of the goalposts so far and fast it should be giving you windburn.
The difference being, probably the vast majority of humans _could_ publish in a journal if they devoted their life to it 100%.
Additionally the AI does not get bored or frustrated, which is probably one of the biggest impediments (other than money) that most people would have to such an endeavour.
If the AI had to do this while also doing all the other things humans do at the same time and constrained to the power of a human brain, then yes it would be unrealistic.
>The difference being, probably the vast majority of humans _could_ publish in a journal if they devoted their life to it 100%.
I believe that, yes. That's why I said "realistically", because of course the vast majority of humans currently in existence actually cannot in a reasonable timeframe (keep in mind ChatGPT has been around for 8 months), no matter how much you incentivise them. And maybe if there were 7 billion AIs on the planet, 0.00174533469% of them could produce a publishable paper on mathematical theory throughout an average 60-80 years of life - I don't believe it, but we have nowhere near enough knowledge about current AI systems to say for sure right now.
My point isn't that an AI couldn't generate a novel mathematical proof - eventually I'm certain one could, and we should definitely work towards it. My point is that it is absolutely absurd to say that an AI isn't intelligent if it can't generate a novel mathematical proof, because if that standard was applied to humans it would mean 99.9982546653% of us aren't intelligent.
Yes but that's not the question either. Chatgpt can probably publish in a journal already. The question is whether it can make impactful work. This is very unclear if most people could do even if working on it 100%
So what if you cant either. You as a human at least possess the self-direction and innate will to direct your own actions and thoughts in some direction. Does AI? No it doesn't. It literally does nothing unless directed by human-set parameters into doing so. For now, regardless of technical abilities, this makes AI far from anything that can easily be defined as sentient.
We evolved visual and hand dexterity for considerably more than two millions years (more like hundreds of millions, I don't care to go find out when hands first evolved but the neural crest was 550 million years ago), and we need many, many more than "only few examples" to be able to draw hands, let alone a sentence. This is something you could only possibly say if you have never tried to draw a realistic hand. There is a reason that long before AI image generation was publicly talked about, many artists joked all the time about how they could draw everything except hands correctly. Hands are particularly difficult to draw. If anything, it is genuinely interesting and maybe worthy of research as to why hands are so hard for us to draw and why they are so hard for these DL networks to draw, and if the reasons are related.
I am of the opinion that AI is neither truly artificial in nature nor intelligent, in the way that we imagine intelligence.
But AI is capable of doing the things you mentioned, perhaps not on that scale just yet, but certainly in principle.
The reason being, that transformer AI in LLM models is actually just an engine for parsing human intelligence.
As the engine improves, it will appear “smarter”, but it is still just parsing its way through the n-dimensional memetic matrix that is human language and culture. …. Just like we do.
Unless there exists a superintelligence expressed in that data set, AI will not express superintelligence in the way we would expect.
AI does express superintelligence though. In its ability to carry on coherent conversations with thousands of people simultaneously on a diverse range of subjects and create documents and code at the speed of conversation.
Right now it is hobbled by limitations of the parsing engine and an inflexibility of not being able to aggregate new knowledge, but those things are improving and being worked on, just not ready for public access yet.
The problem with LLMs, which are really impressive, is that they have too many parameters. This makes them fragile, as they lack regularization.
IMHO, a very interesting research route is to combine small mechanistic models with big networks such as LLMs, where the latter play the role of intuition.
Research by Tenenbaum from MIT, and other similar groups, is heading in this direction.
The "inteligence" level is extremely hard to gauge overall since humans typically have a very deep understanding of a handful of topics, with lots of repetition and knowledge distillation while learning.
Whereas current LLM training focuses on getting as much diverse data as possible and training for very few epochs, so the inteligence we get is wider than any human could ever hope to achieve, but is also shallow in its understanding of each topic. Combining all fields of knowledge together does have its benefits in solving certain problems though.
High schoolers come up with new ideas all the time. AI is just generic scenarios that are mathematical average of what you’d expect. It’s really reductionist to say people are that, betrays a certain cynicism about humanity
Those new ideas are hardly ever better than a string of words probabilistically drawn from a bag. Saying that AI can't think the same things we do betrays a kind of human exceptionalism that's been a part of AI criticism forever. The bar is always one step higher, like some kind of mathematical function that approaches infinity as capabilities improve.
So are you then saying that nothing humans do is derivative?
How many books and movies are just retelling one of a handful of original stories in a different way?
That's all ai does, it builds upon existing works just life we do.
It wrote an entire very accurate book outline for a book about langchain a tool it knew nothing about. I only fed it very basic info, not the whole docs. Clarified a few things it got wrong etc...
Many of the claims against it is that it can't do x thing that "some" group of humans can do even if it's someone many others also can't do.
I guarantee you, it can write a better proof or get closer than I ever would.
IMHO one of the things it really lacks is a purpose or drive. Without a desire and without rewards for learning shit, it'll only learn the basics it needs to formulate an answer.
Curiosity, intrigue, desire, even if we could fake that, might lead to some interesting things. Also adding senses and multi modal things.
AI already does 'innovative' work in the art field. It makes new images, new things that have not been digitally painted before. I think that making new proofs or new other intellectual things is something that can be solved just by making better models.
Most humans are not intelligent by the GP proposed measure. Whatever happened to the Turing test? It's core concept was the holy grail, the inability to know if you are speaking to a human on the other side of your text messages or not was an unimaginable apex. We have now not only conquered that summit, but blown right past it.
I've started to think of LLMs as not so much AI as collective intelligence. An LLM aggregates a huge amount of human-generated information and thinking into one convenient semi-intelligent entity, without doing much really original thinking of its own (so far).
But this alone is potentially profound. Better ways to be collectively smarter could itself accelerate change. Vernor Vinge's famous essay "The Coming Technological Singularity" wasn't just about AI; he also suggested collective intelligence as a way the singularity could happen.
I have a feeling that these ad-hoc bars for AI to clear are extremely similar to Plato's definition of a human ("a featherless biped"), in that they look for what features a human/intelligence has, but don't incorporate the flip-side.
Hence the large amount of Diogenian refutations; including plucked chickens, chess computers, visual classifiers, generative AIs, LLMs, and now, I guess, proof generators (which already exist in some form or another) that could loudly proclaim "behold, a human!".
Unless we rigidly define what intelligence actually is, how can we even hope to correctly identify one?
If the word intelligence is too overloaded or obscure, it is best replace that word use with component words, often new themselves. Intelligence can redefine itself.
I have a theory that there is a kind of dual-think going on around AI 'hallucination'. Specifically that the only meaningful difference between imagination and what people are calling hallucination is whether or not the outcome is useful.
Complete lay-person viewpoint here of course, outside of toying with some neural networks back in the day.
Most hallucinations I’ve seen “make sense” or “look right”. I guess that’s a certain type of creativity. And it’s not like common sense ideas have never been profitable..
I think the difference is more to do with the fact that 'hallucination' is passed off as reality (whether it's ChatGPT confidently telling you that Abraham Lincoln had a MySpace page, or that weird guy on the train telling you that there are spiders in the seat cushions).
People are usually able to distinguish between their imagined scenarios and the real world.
> computers [AI] should be able to understand numbers, and run analysis on numbers.
That's not how any of this works!
"Human brains are made of neurons, so humans must be experts on neurons."
Large Language Models are all notoriously bad at simple arithmetic ("numbers") for the same reasons humans are. We cheat and use calculators to increase our numeracy, but LLMs are trained on human text, not the method used to generate that text.
They can see (and learn from) the output we've generated from calculators, but they can't see the step-by-step process for multiplying and adding numbers that the calculators use internally. Even if they could see those steps and learn from that, the resulting efficiency would be hideously bad, and the error rate unacceptably high. Adding up the numbers of just a small spreadsheet would cost about $1 if run through GPT 4, but a tiny fraction of a cent if run through Excel.
There have been attempts at giving LLMs access to calculator plugins such as Wolfram Alpha, but it's early days and the LLMs are worse at using such tools than people are.
I'm not an expert on this by any means, but from what I've read about the four-color theorem, essentially all of the insight in the proof came from human mathematicians. Mathematicians figured out an ingenious way to reduce to the problem to carrying out a a large number of tedious computations, and the computer was used only to carry out those computations and thereby complete the proof. This seems quite different from a computer coming up with a proof by itself.
I recall that DeepMind’s AI discovered a new type of Sorting Algorithm. Sorting is one of the most “trafficked” area of CS research, so I would say it’s a true discovery.
Certainly computer-aided proofs have been a thing for a while; I wonder where one draws the line between "the AI made a new proof" and "we build a system that proved X"? Which I guess really gets at the more basic question "what is an AI"?
I think one line is: you ask the computer about a problem in a declarative way (e.g. Z3) and the computer came up with a solution or a description of where it is blocked by the problem with potential areas of attack.
It doesn't even have to discover anything new to be a compelling proof of concept. All it has to do is discover something we already know without having been fed the answer in some way.
Today's AIs can't do this, because the entire basis for their intelligence is having been fed all the answers humanity has, and regurgitating those back to us in a somewhat more flexible and adaptive way than a search engine.
The current AIs powered by LLMs intend to "talk/think like ordinary humans do".
There might exist some practical applications for such AIs that might have economic value, but doing highly innovative things is not among these.
Doing highly innovative things rather means subverting the current state of art in a very clever way. If you think of people who have this property, you will likely immediately think of some ingenious smartass who is nearly always right, but insanely annoying to the people surrounding him because of this know-it-all attitude.
Would such an AI be possible to create? I don't know, but let's assume it is.
What should be obvious is that such an AI would need entirely different techniques to develop, but let's again assume that this problem has been solved.
What would a business model for such an AI look like? You clearly could not sell API access to it, since such an AI would demand far too demanding in the learning requirements for its users (discussion partners if implemented as a chatbot); look in the mirror: how many post-graduate level textbooks about some scientific topic (in particular math or physics) did you read in the last months?
So, such an AI would only make sense in the basements of some big corporation or three-letter agency, where there AI is commanded by some insanely brainiac users who have gotten a yearslong training to develop the actual intellectual capacity and scope of knowledge to actually understanding a glimpse of the AI's ideas. This glimpse then "trickles down" into innovations where no one has the slightest idea about where their true origin is (they fell into someone's lap).
The machines invented so far haven't done "highly innovative things" all by themselves, and yet people doing innovative things often find their machines useful. I expect organizations consisting of both humans and machines will still be pretty important for a while.
> The machines invented so far haven't done "highly innovative things" all by themselves, and yet people doing innovative things often find their machines useful.
This statement is near to being tautological. The central test is rather whether people who do "highly innovative things" become more productive in doing so not just by the mere fact that the AI removes some yak shaving from their work.
Otherwise, you could simply argue that people who do "highly innovative things" also find
- a housecleaner
- a personal secretary
- using a word processing program instead of a typewriter
But they do find them useful. Personal computers were a pretty important invention, too. More recently, the web and "smart" phones (which aren't actually smart) resulted in major changes to organizations. We work differently now.
I'm not actually sure what argument you're making, though? It seems like you're saying that only a certain kind of technological innovation would count for some purpose, but I don't know what purpose you're interested in.
In my answer to your answer (https://news.ycombinator.com/item?id=36937760), I argued that the fact that people doing innovative things often find their machines useful says nothing about that AIs are capable of doing innovative things.
Yes, I read that, but it's unclear what your assumptions are or what you value. I guess having AI that is "highly innovative" in the same way that some people can be innovative is something you value, but it's not that clear to me why that's important.
It’s the exact opposite trope in my experience. The ingenious people that are always right are invariably courteous, polite and a pleasure to be around. Those a little bit lower on the intelligence rung are usually the ones that feel the need to be contrarians and generally disruptive to “prove” their intelligence.
> The ingenious people that are always right are invariably courteous, polite and a pleasure to be around.
My experience in both academia and business was/is very different: to those who are insanely competent, you typically had (in some sense) to "prove" by diligence and intelligence that you are "worthy" their time.
I can understand that attitude really well: otherwise a lot of people would insanely waste their time, and the respective people would get nowhere.
Arrogant smartasses often put more effort into having their genius recognized than they do in being correct, and that often... works. Loud and aggressive people take as much credit as they can and they make sure everyone knows, so they generally appear more competent than the rest.
Sometimes they really are outstanding on their own (e.g. Linus Torvalds), but just as often, it turns out that yes, they are highly competent, but the true insights actually come from their underlings.
I could scarcely have predicted the rapid breakthrough pace of innovation that got us where we are in 2023, and I dare not try to predict as this author is trying to do what will be hard or impossible about the innovation we’ll see in another 2, 5 or 10 years.
People are terrible predictors of the future, and especially terrible in emerging fields and novel areas of research. It’s astounding how much confidence people continually posses in their predictive capabilities despite their truly dismal track records in predicting the future.
People just blindly extrapolate. Things that are down: math and language skills, transportation speed, life expectancy. Things that are up: manufacturing costs, mental health conditions, incarceration rates. Batteries and rockets are better, but well below expectations. CPU performance is well below expectations. Parallel computing improvements are satisfactory, I suppose.
The amount of AI woo on this comment thread (and on this site in general) is absurd. I see various defenses of AI intelligence that trot out the (at lest for now in the case of present technology) tired arguments that 1. humans also create derivative work just like AI and rarely display original creativity, and 2. That AI does create all kinds of new iterations of previous inputs from distributed sources.
One particularly absurd comment further below even states as follows:
>The vast majority of humanity has not produced an original mathematical proof worthy of being published in a peer-reviewed math journal, and realistically it isn't possible for the vast majority of humanity to do so.
None of these demonstrate anything close to sentience or real intelligence in present LLM and AI systems. The most fundamental reason why is obvious: AI IS NOT self directed. It has no innate reasoning, sense of self or motivation of any kind to direct its own actions and "thoughts" towards any particular goals, emotional or otherwise.
This alone is a simple, straightforward bar that distinguishes current AI as a very elaborate autofill system from genuine sentience, even if the sentience applies to a perfectly ordinary human that also never publishes their own math proof or creates a work of art.
Also, to further dispel the argument about most humans not doing these things either and also largely being derivative thinkers who base their notions on an aggregation of what others previously created: So what? Humans at least innately can, as a species perform original feats, even if not all do so and some require much more effort than others to do so. Even among humans who never achieve something orignally or cleverly their own, self direction and motivation are pretty much universally present. AI lacks these, no matter how you dress it up or how much visual and media koolaid you drink. And you have to be deeply, blindly emotionally fixated on AI to argue otherwise, because no existing evidence supports your notions.
As for AI passing the Turing test, this only shows that the Turing test was an unwittingly bad measure of AI sentience. Done.
Human level AGI is hard to achieve but there are loads of smart people working on it so there's a good chance of progress. If it comes to be it's bound to be transformative, I figure.
> We could be headed off in the wrong direction altogether. If even some of our hurdles prove insurmountable, then we may be far from the critical path to AI that can do all that humans can.
humans can go insane and start putting memories and thoughts together in ways that don't make sense in any logical or mathematical way
perhaps Skynet will be born the first time someone tries voltage-glitching an AI-enabled padlock
"Why many human beings believe/predict transformative artificial intelligence is really, really hard to achieve according to what we've thought of so far" would be a more accurate title.
There is no underlying theory of this, as it relates to intelligence.
To all intents and purposes modern AI is pattern matching and good statistical inferences against a semantic or other contextualising model. Sometimes it includes a good trick like shifting phase space, or adopts approaches from linear programming or genetic algorithms or whatever, but back-propagation aside.. What it lacks is "discriminiation" in the sense of telling good facts from bad, and "inductive reasoning" which is so easy to say, but remarkably hard to do (hence, how long it takes us to do it beyond the trivial) -This has to be externally sourced. Systematic bias creeps in. Thus artists are sueing because this amazing "new art" is so highly derivative it points strongly to the hands whose works were "copied" (and I use that word deliberately) to make it. Time after time I see experts point out the GPT productions are word-vomit good, to style but lack substance and are often just plain wrong.
The stories about porn detection training to learn % image hues of skin-tone and little else come to mind. Face recognition is bloody good. I am seriously impressed by face and other specific search term models in image analysis. But its also lacking "the how" here.
Code is somewhat unique in being highly proscribed. That GPT does a good job of finding code examples in SQL does not point to it doing a good job of inductive reasoning of the applicability of maritime law to a specific problem in common carriers, or how velcro works, but causes more deaths than buttons because soldiers need to be silent. It can't "know" these burdens, it can only correlate (or not)
I think it's remarkable but the kind of implications in "transformative" are something people should not expect to see, until there is at least some attempt at science here. For now, its pretty much "shake the box and see what sticks" with some good theories about tuning the models, but no inherent claims to understand "why it works" and most importantly, how it relates to consciousness or intelligence as we implement it in wetware.
I really want to see people expose theory. How does it work? Not "moar GPU" or "Moar data" but how it actually forms output, without highly specific weighted trained processes from humans. What I hear is that training AI on AI outputs is fruitless. To me, thats as good as test as any: when you can use at scale an AI product to inform the training of an unrelated AI system I will be impressed.
Chat Bots are not impressive. TL;DR the turing test is not actually a determinant of anything except "humans can be fooled"
Thats how I see it. I am not in field. Happy to be corrected by people who are. I think people like Hinton are being extremely careful in their choice of words when they speak about AI, and people like Kurzeweil are not. I pay attention to what Geoffrey Hinton says.
> What I hear is that training AI on AI outputs is fruitless. To me, thats as good as test as any: when you can use at scale an AI product to inform the training of an unrelated AI system I will be impressed.
The results exceeded the researchers' expectations. The team attributes them directly to data quality, as the title of the paper suggests: "Textbooks is all you need," in reference to Google's research on the Transformer breakthrough ("Attention is all you need").
If I have the right context, they selected HIGHLY QUALIFIED inputs. Which probably meant their training set was good as a seed to another system because it had been pre-groomed.
Until AI can do that, we have to admit that it's not really aware or sentient or any of the other more ambitious things that have recently been claimed for it.
Can AI teach us anything new about the pattern of prime numbers?
Can AI develop an original proof for the shape of shadows in high dimensional spaces?
Can AI creatively prove a new limit to mathematics?
There are 2 researchers in AI who deserve more attention: Kenneth O. Stanley and Joel Lehman. They wrote a great book: Why Greatness Cannot Be Planned. They look at the limits of utility functions and explain the importance of novelty. As an antidote to some of the hype around AI, I strongly recommend this book:
https://www.amazon.com/Why-Greatness-Cannot-Planned-Objectiv...