7 ms·

Why transformative artificial intelligence is hard to achieve

86 points by hunglee2 3y ago

ggm 3y ago
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.
- causalmodels 3y ago
  > 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 phi-1 team did this a few weeks ago.
  - ggm 3y ago
    I'm impressed. Noting this: 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. But, it's still impressive.
iraqmtpizza 3y ago
All string theory research should be handed over to AI so that humans can get back to doing real physics
southernplaces7 3y ago
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.
lkrubner 3y ago
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: https://www.amazon.com/Why-Greatness-Cannot-Planned-Objective/dp/3319155237/ https://www.amazon.com/Why-Greatness-Cannot-Planned-Objectiv...
- jiggawatts 3y ago
  > 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.
- 3y ago
mistermann 3y ago
"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.
jareklupinski 3y ago
> 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
tim333 3y ago
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.
aleph_minus_one 3y ago
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).
- soligern 3y ago
  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.
  - aleph_minus_one 3y ago
    > 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.
aaron695 3y ago
[dead]
happytiger 3y ago
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.
- iraqmtpizza 3y ago
  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.
  - happytiger 3y ago
    That’s quite an astute comment and I thank you.

86 points by hunglee2 3y ago

ggm 3y ago

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.

causalmodels 3y ago
> 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 phi-1 team did this a few weeks ago.
- ggm 3y ago
  I'm impressed. Noting this: 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. But, it's still impressive.

iraqmtpizza 3y ago

All string theory research should be handed over to AI so that humans can get back to doing real physics

southernplaces7 3y ago

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.

lkrubner 3y ago

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: https://www.amazon.com/Why-Greatness-Cannot-Planned-Objective/dp/3319155237/ https://www.amazon.com/Why-Greatness-Cannot-Planned-Objectiv...

jiggawatts 3y ago
> 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.
3y ago

mistermann 3y ago

"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.

jareklupinski 3y ago

> 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

tim333 3y ago

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.

aleph_minus_one 3y ago

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).

soligern 3y ago
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.
- aleph_minus_one 3y ago
  > 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.

aaron695 3y ago

[dead]

happytiger 3y ago

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.

iraqmtpizza 3y ago
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.
- happytiger 3y ago
  That’s quite an astute comment and I thank you.