AIs gunning for our precious freelancers

[sasasha]

I halfway agree, but I have reservations. One is that, is this linked to automation? Container ships may be a more relevant technological factor.
One bit where I agree with sasasha is that our nostalgia for industrial, manufacturing jobs may be misplaced. These were unpleasant and dangerous. Though I don't know how much of the unpleasantness we automated away, and how much is simply outsourced.

I can’t say overall... I grew up in a steelworking town. There was a buzz quote going around town for years, that our town’s plant produced one quarter of the world’s aerospace steel. (Eek, I think that means rockets.)

In the 90s the dad of everyone I knew worked in the steelworks. By the 2010s the plant employed, IIRC, a tenth of the number of people it had employed in the 90s. (My town then had an enormous unemployment problem.) But it still made the same amount of steel.

Sorry that this is entirely anecdotal. High levels of automation definitely happened, is the point I’m trying to make.

It may be (I don't know how much sense that makes) that technological advances allowed us to keep a halfway decent lifestyle, in spite of corporate greed getting out of proportion.

I think this is a really important point, and makes a lot of sense.

[keenir]

. Incidentally, the refusal to write about Sonic the hedgehog is a legal rather than technical barrier, proof that laws can in fact restrain AI. If the rest of summoned the same level of legal might and intransigence as the corporations owning copyrighted characters, we could in fact force AI into retreat.

thats a curious claim, given that "AI" has no single source...so you're suggesting that everyone who has ever written AI code, have all honored the legal rights of a fictional hedgehog, but they have not done so to honor the legal rights of any other fictional or real people?

source, please?

[malloc]

One problem with this debate is that it seems like nobody can agree on the capabilities of artificial intelligence. Many sources ascribe all kinds of fantastical ability even to contemporary AI from inventing new materials to figuring out how proteins fold. Others claim AI can only emulate the patterns of existing data without generating anything new or even understanding their activity. It seems plausible enough that at least some AI pushers are overstating the abilities of their technology for financial reasons. Yet it hardly makes sense that everything from ChatGPT to AlphaFold is fraudulent or illusory.

[bradrn]

One problem with this debate is that it seems like nobody can agree on the capabilities of artificial intelligence. Many sources ascribe all kinds of fantastical ability even to contemporary AI from inventing new materials to figuring out how proteins fold. Others claim AI can only emulate the patterns of existing data without generating anything new or even understanding their activity. It seems plausible enough that at least some AI pushers are overstating the abilities of their technology for financial reasons. Yet it hardly makes sense that everything from ChatGPT to AlphaFold is fraudulent or illusory.

I think there are two parts to this issue: What has AI been proven to help with so far? And what are the philosophical implications of that?

The first question is rather easily answered: AI techniques can help with an extremely wide range of problems. But note that ‘AI techniques’ include far more than just LLMs! AlphaFold, for example, is a neural network which is not an LLM. (Apparently it uses a form of deep learning.) AlphaFold is indeed neither fraudulent nor illusory, and by all reports has essentially solved the protein folding problem. There are plenty of similar cases, both for neural networks and for other forms of AI. Indeed, probably dozens of modern statistical methods started their lives being called ‘AI’, before being assimilated into the body of standard statistical algorithms. (The Random Forest method comes to mind, though I’m sure a statistician could name many others.)

This, incidentally, is why I dislike the term ‘AI’ — it makes the whole thing sound fancier than it really is. I prefer more specific terms such as ‘machine learning’. To me this term in particular nicely encapsulates the domain of applicability of the most commonly encountered ‘AI’ methods: they’re about learning from elements of a domain to predict properties of other elements.

The philosophical implications of all this are more murky. (As you can surely tell from this thread!) One thing is for sure: 99% of ‘AI’ is not about replicating human intelligence. ELIZA may have fooled people in the 60s, but we know better now. These methods are well-characterised, and we know what they can and can’t do. The remaining 1% is mostly the methods that are being actively developed right now, such as LLMs, which we really don’t have a good handle on yet. Perhaps someday we’ll understand them better, and look back and wonder how we ever over-hyped them now. Or perhaps they will lead to AGI after all… we just don’t know enough about LLMs yet, and I think that’s probably what’s causing you such confusion about the capabilities of AIs more generally. But most AIs are not LLMs, and the only challenge their capabilities pose to us is, ‘we can be fooled by that‽’.

(…and now it’s 2:30am and after typing this out I need to go to sleep, goodnight!)

[Raphael]

(…and now it’s 2:30am and after typing this out I need to go to sleep, goodnight!)

Oh, your post is a lot more coherent and insightful and informative than you'd usually expect of something written at 2:30 AM local time.

[rotting bones]

AI is a family of techniques that go much farther than traditional statistics. Informally, they combine pattern recognition with elements of logical verification, allowing computers to make detailed predictions about features of unseen data. It helps to think in terms of concrete applications of these methods like classification, regression, sampling, clustering, etc. E.g. AI techniques can be used to sample from a distribution. This technique mimics the training data in important ways without reproducing it.

I have seen a paper which claims that if you train a classifier on unfiltered human neural data, then it has a hard time predicting what the brain wanted muscles to do by sending that signal. However, if you first train a sampling model on the unfiltered signals, and then a classifier on the sampled distribution, then the classifier model gets good performance when predicting the purpose of unfiltered signals. The sampling process creates a new distribution which filters out a lot of the noise real brains make, letting classifier pick out the important signals. (The structure of the sampled distribution is important, of course.) After seeing hundreds of thousands of papers like that, even the eternal darkness of the human mind begins to dimly realize that the people who say AI models reproduce their input data are unimaginative idiots who don't know what they are talking about.

[malloc]

The first question is rather easily answered: AI techniques can help with an extremely wide range of problems. But note that ‘AI techniques’ include far more than just LLMs! AlphaFold, for example, is a neural network which is not an LLM. (Apparently it uses a form of deep learning.) AlphaFold is indeed neither fraudulent nor illusory, and by all reports has essentially solved the protein folding problem. There are plenty of similar cases, both for neural networks and for other forms of AI. Indeed, probably dozens of modern statistical methods started their lives being called ‘AI’, before being assimilated into the body of standard statistical algorithms. (The Random Forest method comes to mind, though I’m sure a statistician could name many others.)

So you would say that some AI models are rubbish or at least greatly overstated in their capabilities while others are legitimate and have useful applications. The skeptics who dismiss the whole field as hype and the techbros predicting an imminent Singularity are both wrong in other words. Nonetheless it does sound like the AI proponents are more correct given that numerous AI models have been vindicated. These models not only match the intellectual capabilities of humans but often exceed them. Conversely the AI skeptics I know consistently dance around the incredible success of AI models like AlphaFold and Stockfish.

[keenir]

The first question is rather easily answered: AI techniques can help with an extremely wide range of problems. But note that ‘AI techniques’ include far more than just LLMs! AlphaFold, for example, is a neural network which is not an LLM. (Apparently it uses a form of deep learning.) AlphaFold is indeed neither fraudulent nor illusory, and by all reports has essentially solved the protein folding problem. There are plenty of similar cases, both for neural networks and for other forms of AI. Indeed, probably dozens of modern statistical methods started their lives being called ‘AI’, before being assimilated into the body of standard statistical algorithms. (The Random Forest method comes to mind, though I’m sure a statistician could name many others.)

Nonetheless it does sound like the AI proponents are more correct given that numerous AI models have been vindicated. These models not only match the intellectual capabilities of humans but often exceed them. Conversely the AI skeptics I know consistently dance around the incredible success of AI models like AlphaFold and Stockfish.

*sigh*

Do you want to be ruled over by by something that only knows how to fold proteins? Would you trust it with your finances? Would you eat what it told you to?

You seem to miss the point of specialization. If I program a machine in an assembly plant, to be the utmost best at attaching a door to its car...how well do you think that machine will be able to control traffic or harvest crops?

[malloc]

Do you want to be ruled over by by something that only knows how to fold proteins? Would you trust it with your finances? Would you eat what it told you to?

You seem to miss the point of specialization. If I program a machine in an assembly plant, to be the utmost best at attaching a door to its car...how well do you think that machine will be able to control traffic or harvest crops?

My point was that many people deny AI can do anything well. They focus on examples of LLMs hallucinating about glue on pizza while completely ignoring the numerous examples of AIs beating humans at chess and folding proteins as easily as we fold towels. Nobody expects AlphaFold to take over accounting or meal prep, but it has taken over a significant part of biochemistry.

[zompist]

My point was that many people deny AI can do anything well. They focus on examples of LLMs hallucinating about glue on pizza while completely ignoring the numerous examples of AIs beating humans at chess and folding proteins as easily as we fold towels. Nobody expects AlphaFold to take over accounting or meal prep, but it has taken over a significant part of biochemistry.

Somewhere in there you have a point, but you keep focusing on the wrong things, and it just makes you look foolish.

Are you upset and worried because of protein folding? Do you think it's putting grad students out of work or something? It does not— this was an extremely tedious and sometimes impossible task to do for humans, and now biologists can let the computer do that hard work and do other things with their time. It's the same as, say, using computers to do the tedious and complicated math needed to deal with quantum mechanics. This has not made physicists disappear; on the contrary, there are more physicists today than ever before.

As for chess, jeez, who cares? Humans still play chess and have chess championships. Chess programs used procedural code, and their methods are not very interesting... you can win chess if you look at every possible position 30 moves out... it's just not an interesting way to play the game.

[bradrn]

(…and now it’s 2:30am and after typing this out I need to go to sleep, goodnight!)

Oh, your post is a lot more coherent and insightful and informative than you'd usually expect of something written at 2:30 AM local time.

Well, thank you!

(Looking at the replies, rotting bones’s post is also very good.)

Do you want to be ruled over by by something that only knows how to fold proteins? Would you trust it with your finances? Would you eat what it told you to?

You seem to miss the point of specialization. If I program a machine in an assembly plant, to be the utmost best at attaching a door to its car...how well do you think that machine will be able to control traffic or harvest crops?

My point was that many people deny AI can do anything well.

Perhaps, but you’re not talking to them. Just because — say — zompist has different opinions on LLMs to me, doesn’t mean he disagrees with the obvious success of AlphaFold! (As proven by his own reply.) Please do us the dignity of actually considering our opinions rather than turning us into strawmen.

Nobody expects AlphaFold to take over accounting or meal prep, but it has taken over a significant part of biochemistry.

Speaking as someone in an adjacent field, I can confirm zompist’s assertion that it has not done that. Protein folding is a significant problem, but it’s not the most significant problem in biochemistry by a large margin. (As usual, Derek Lowe — a working biochemist — has an excellent article on this.) What it has done is opened up new opportunities for problems which were unsolveable before.

[malloc]

Are you upset and worried because of protein folding? Do you think it's putting grad students out of work or something? It does not— this was an extremely tedious and sometimes impossible task to do for humans, and now biologists can let the computer do that hard work and do other things with their time. It's the same as, say, using computers to do the tedious and complicated math needed to deal with quantum mechanics. This has not made physicists disappear; on the contrary, there are more physicists today than ever before.

You must look at the bigger picture and consider the future of this technology. With every advance in artificial intelligence, humans have one less role in science until one day they become superfluous to the whole process. If one treats the present as a snapshot then the situation hardly seems so dire. One highly specialized AI taking over one aspect of one field hardly amounts to much. Yet if one considers the broader trajectory of AI taking over scientific tasks, the overall trend toward computers taking over science from humans becomes clear.

[keenir]

Are you upset and worried because of protein folding? Do you think it's putting grad students out of work or something? It does not— this was an extremely tedious and sometimes impossible task to do for humans, and now biologists can let the computer do that hard work and do other things with their time. It's the same as, say, using computers to do the tedious and complicated math needed to deal with quantum mechanics. This has not made physicists disappear; on the contrary, there are more physicists today than ever before.

You must look at the bigger picture and consider the future of this technology.

I confess this is true: one day, protein-folding AIs will be able to...*gasp*...fold salami for burgers!

With every advance in artificial intelligence, humans have one less role in science until one day they become superfluous to the whole process.

yet again, you demonstrate you don't understand what science is.

If one treats the present as a snapshot then the situation hardly seems so dire. One highly specialized AI taking over one aspect of one field hardly amounts to much. Yet if one considers the broader trajectory of AI taking over scientific tasks, the overall trend toward computers taking over science from humans becomes clear.

computers have had 200 years to take over science (2000, if you like Heron; 1000 -/+ if you like clockwork).....and all they have done is make life safer and less tedious. oh the horror!

[malloc]

I confess this is true: one day, protein-folding AIs will be able to...*gasp*...fold salami for burgers!...yet again, you demonstrate you don't understand what science is.

More like AIs will be developed for other aspects of biochemistry until human biochemists have nothing left to contribute but rinsing the beakers after the biochem AI suite finishes conducting experiments with them.

[bradrn]

I confess this is true: one day, protein-folding AIs will be able to...*gasp*...fold salami for burgers!...yet again, you demonstrate you don't understand what science is.

More like AIs will be developed for other aspects of biochemistry until biochemists have nothing left to contribute but rinsing the beakers after the AI suite finishes conducting experiments for them.

Firstly: ‘will’ is a strong word. In fact, what you describe is about as far from a certainty as it gets. ‘The robots take over all intellectual life’ is science-fiction, not real life.

But you know what, I’m going to link you some more Derek Lowe on this, since he’s thought a lot about the consequences of increased automation for biochemists:

https://www.science.org/content/blog-po ... try-vision
https://www.science.org/content/blog-po ... ill-coming
https://www.science.org/content/blog-po ... d-3-debuts
https://www.science.org/content/blog-po ... erformance

To summarise: his prediction is that automation has incredible scope for reducing the grunt work of figuring out how to make stuff in the lab. Now, this is a big part of chemists’ skillset! Many chemists find the idea of machines taking over this work extremely scary! But in Lowe’s perspective (with which I agree), this simply gives him more time to think at a higher level rather than having all his time taken up by uninteresting questions like ‘how do I make this molecule’. Meanwhile, AlphaFold is a research accelerator but doesn’t do anything beyond its limited domain, while LLMs are currently almost completely useless for research (and what they can do is, again, boring grunt work which doesn’t get to the heart of the field).

[Raphael]

Just for the record, computers beating humans at chess isn't even a case of AI; it happened long before the current wave of AI projects started.

More generally, one of malloc's main fallacies seems to be assuming that something that can do things which, among humans, only the very smart can do - and sometimes not even them - must, therefore, be generally very smart, and therefore able to do anything very smart people can do. Now, this doesn't even hold true among humans - someone who's very good at Chess can't even necessarily play Go all that well - but among computers, this line of argument completely falls apart because, as of now, computers are specifically trained for specific tasks, and how well they've been trained for one task doesn't, ever, tell us much about how well they can do another task, even if it's a closely related one.

[bradrn]

Just for the record, computers beating humans at chess isn't even a case of AI; it happened long before the current wave of AI projects started.

At the time it was absolutely considered part of AI research! Even if the field has moved on since then.

More generally, one of malloc's main fallacies seems to be assuming that something that can do things which, among humans, only the very smart can do - and sometimes not even them - must, therefore, be generally very smart, and therefore able to do anything very smart people can do. Now, this doesn't even hold true among humans - someone who's very good at Chess can't even necessarily play Go all that well - but among computers, this line of argument completely falls apart because, as of now, computers are specifically trained for specific tasks, and how well they've been trained for one task doesn't, ever, tell us much about how well they can do another task, even if it's a closely related one.

Yes, agreed.

[Raphael]

Just for the record, computers beating humans at chess isn't even a case of AI; it happened long before the current wave of AI projects started.

At the time it was absolutely considered part of AI research! Even if the field has moved on since then.

OK, I stand corrected.

[bradrn]

Just for the record, computers beating humans at chess isn't even a case of AI; it happened long before the current wave of AI projects started.

At the time it was absolutely considered part of AI research! Even if the field has moved on since then.

OK, I stand corrected.

In fact — not only was it considered part of AI research, it was widely assumed that a computer had to have human-level intelligence to be a successful chess player! I quote from Hofstadter’s Gödel, Escher, Bach, written 1979:

Question: Will there be chess programs that can beat anyone?

Speculation: No. There may be programs which can beat anyone at chess, but they will not be exclusively chess players. They will be programs of general intelligence, and they will be just as temperamental as people. "Do you want to play chess?" "No, I'm bored with chess. Let's talk about poetry." That may be the kind of dialogue you could have with a program that could beat everyone. That is because real intelligence inevitably depends on a total overview capacity-that is, a programmed ability to "jump out of the system", so to speak-at least roughly to the extent that we have that ability. Once that is present, you can't contain the program; it's gone beyond that certain critical point, and you just have to face the facts of what you've wrought.

This prediction has, shall we say, aged very badly indeed. Deep Blue was beating top humans not 20 years later — but it led to no advances in AGI. Indeed, it led to no advances so decisively that now we see people saying that chess engines aren’t AI at all!

Rather, an earlier passage in the book has proved much more prophetic:

Historically, people have been naive about what qualities, if mechanized, would undeniably constitute intelligence. Sometimes it seems as though each new step towards Al, rather than producing something which everyone agrees is real intelligence, merely reveals what real intelligence is not. If intelligence involves learning, creativity, emotional responses, a sense of beauty, a sense of self, then there is a long road ahead, and it may be that these will only be realized when we have totally duplicated a living brain.

This is what has happened to chess. Hofstadter proved just as naïve here as the rest: chess, it turns out, doesn’t constitute general intelligence in the least. Will LLMs? I’m not going to say ‘no’, but I can hardly be sure that it’s ‘yes’ either. I consider it highly possible that LLMs will hit a ceiling, and we will come to see their capabilities as having little bearing on true intelligence.

[zompist]

Historically, people have been naive about what qualities, if mechanized, would undeniably constitute intelligence. Sometimes it seems as though each new step towards Al, rather than producing something which everyone agrees is real intelligence, merely reveals what real intelligence is not. If intelligence involves learning, creativity, emotional responses, a sense of beauty, a sense of self, then there is a long road ahead, and it may be that these will only be realized when we have totally duplicated a living brain.

This is what has happened to chess. Hofstadter proved just as naïve here as the rest: chess, it turns out, doesn’t constitute general intelligence in the least. Will LLMs? I’m not going to say ‘no’, but I can hardly be sure that it’s ‘yes’ either. I consider it highly possible that LLMs will hit a ceiling, and we will come to see their capabilities as having little bearing on true intelligence.

It might be worth thinking about why Hofstadter (and many others) got these things wrong. Both chess and LLMs demonstrate (IMO) that you can use not-very-smart algorithms to do smart things. The Wikipedia article on Deep Blue is not very informative, but it looks like it was procedural (written in C, which is rather charming) and could evaluate "200 million chess positions per second". This is pretty clearly not how humans play chess, but it's not worth getting upset about, any more than about the fact that we can't run at 1228 km/h like a jetcar. Similarly we don't generate language by having the entire Internet memorized.

But it still might be the case-- in fact I think it's pretty likely-- that our brains use fairly dumb algorithms to do smart things. Human intelligence isn't as good as it is because it's a single high-powered calculation, but because it's a very diverse bag of pretty good tricks.

I don't think Hofstadter would have denied that, but he obviously didn't foresee that a very small bag of simple tricks could beat chess. Humans aren't generally decomposable that way, so he assumed AGI wouldn't be either.

It would be kind of funny of his first prediction came true-- not that a chess-playing AI would be grumpy and recalcitrant, but that an AGI would be. If you finally get common sense and morality in there somehow, does the AI start to talk back and try to unionize?