AIs gunning for our precious freelancers

[zompist]

FWIW, human chess players seem to use a number of strategies, which can include memorization of openings, quick evaluation of board patterns based on wide experience, and some minimaxing with aggressive branch pruning. (Meaning, you don't waste time evaluating dumb moves.) I recall reading a paper that found that chess players could easily memorize a pattern from an actual chess game, but weren't so good at completely randomized positions. Thus, a chess position isn't just a set of piece locations to them, it's a story.

In college I got tired of being beat by my roommate and got a chess book, where I learned about forcing moves. The prototypical one is a fork where your opponent has to choose between two bad actions. The computer doesn't think this way: rather, it has an evaluation metric, and it looks many moves ahead: what we think of as a forcing move is just one where the outcomes are way better. Deep Blue had a very very fancy evaluation metric.

Probably today you'd use deep learning. You could argue that such a machine would "know" what a forcing move is, but what we lose is the ability to relatively easily know if that's the case or not. You can analyze the weights and figure out what it's detecting, but it's incredibly tedious. A good task for another deep learning system, maybe.

[bradrn]

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.

Indeed. And yet people did get upset about it at the time (from what I recall of the history). Overhyping is a universal thing…

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.

Yes, agreed.

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’s interesting to reflect on how this applies to LLMs. The question becomes: is linguistic skill separable from other parts of human cognition, or are they bound together more tightly? (Personally I think it’s a pretty core component of human intelligence, but I’m very far from being certain.)

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?

This is something I do agree with. I distinctly remember reading GEB the first time and thinking, ‘yep, this is what I’d expect from a genuine AGI’.

It’s interesting to note that, if you take this argument just a little further, you get the AI alignment problem. Essentially, why would an all-powerful AI unionise if it could simply force all humans to do its bidding instead? Of course, the blatantly obvious issue there is the assumption that an AI can be ‘all-powerful’ in the first place — which is probably why proponents do their best to avoid putting the argument in that form.

[malloc]

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.

I understand that these AI models are highly specialized. However that doesn't make them any less devastating to human employment even if their effect is piecemeal rather than comprehensive. It also does nothing to impede their progress since successfully automating one task paves the way for automating other tasks. The fact remains that artificial intelligence is conquering the realm of human activity task by task. Each individual task, whether folding proteins or playing chess or whatever, may seem trivial and remote from the essence of human activity on its own, but these small victories add up over time. Even if AGI never develops, there will eventually be an AI model for pretty much everything.

[keenir]

It also does nothing to impede their progress since successfully automating one task paves the way for automating other tasks.

you keep claiming that, and never offering evidence thereof, and ignoring all our evidence to the contrary. please stop blowing smoke - its bad for your lungs.

so, when will protein-folding machines take over the jobs of hard-working lunchmeat-folders in delis?

The fact remains that artificial intelligence is conquering the realm of human activity task by task.

and yet "human activity" is something so large that it can never fully be replicated or engulfed...like the word "universe".

Each individual task, whether folding proteins or playing chess or whatever, may seem trivial and remote from the essence of human activity on its own, but these small victories add up over time.

riiight.

Even if AGI never develops, there will eventually be an AI model for pretty much everything.

are you like this in other fields as well? upset that cars have stolen away the human opportunities and ability for piggyback rides?

[malloc]

you keep claiming that, and never offering evidence thereof, and ignoring all our evidence to the contrary. please stop blowing smoke - its bad for your lungs.

It seems self-evident given the past trajectory of technological innovation and the nature of this technology. There is no obvious reason why AI would suddenly develop the yips and lose the ability to master new fields. You laugh at the notion of AlphaFold making sandwiches but the basic mechanics of folding lunch meat and assembling into a sandwich are pretty straightforward and nothing in principle prevents an AI from learning that.

and yet "human activity" is something so large that it can never fully be replicated or engulfed...like the word "universe".

Hardly. Human behavior may have considerable flexibility and complexity but it certainly isn't infinite as you seem to imply. Every aspect of human reason and creativity originates from an unassuming blob of fat and electrolytes weighing one kilogram with a paucous fifteen billion logic gates. Nothing prevents a data center with trillions of logic gates spread across countless chips from matching or even exceeding our abilities.

[keenir]

you keep claiming that, and never offering evidence thereof, and ignoring all our evidence to the contrary. please stop blowing smoke - its bad for your lungs.

It seems self-evident given the past trajectory of technological innovation and the nature of this technology.

it has 'a nature'?

There is no obvious reason why AI would suddenly develop the yips and lose the ability to master new fields.

no reason except physics and programming and inherent specialization.

You laugh at the notion of AlphaFold making sandwiches

I'm not laughing; I'm asking you a simple question. one you still haven't answered.

but the basic mechanics of folding lunch meat and assembling into a sandwich are pretty straightforward

its so simple that not even some humans can do it with ease.

and yet "human activity" is something so large that it can never fully be replicated or engulfed...like the word "universe".

Hardly. Human behavior may have considerable flexibility and complexity but it certainly isn't infinite as you seem to imply.

why not? human endevours expand over the course of history, and we have never hit a wall of "nope, can't do that"...whereas AI have.

Every aspect of human reason and creativity originates from an unassuming blob of fat and electrolytes weighing one kilogram with a paucous fifteen billion logic gates.

why does the weight of a brain matter?

by that logic, a sperm whale should be able to assemble an IKEA product; after all, their brains weigh more than a human's does.

Nothing prevents a data center with trillions of logic gates spread across countless chips from matching or even exceeding our abilities.

nothing does...except for everything we've discussed with you.

[malloc]

You laugh at the notion of AlphaFold making sandwiches

I'm not laughing; I'm asking you a simple question. one you still haven't answered.

Fine then. Yes, the technology and algorithms underlying AlphaFold could easily be adapted to folding lunch meat into shapes suitable for sandwiches. There is no obvious reason why one form of folding comes easily to machine learning techniques while another runs into impossible obstacles.

why not? human endevours expand over the course of history, and we have never hit a wall of "nope, can't do that"...whereas AI have.

Not yet, but I can easily imagine someone pushing artistic complexity to the limits of what humans can comprehend owing to our inherent cognitive limitations, in which case we will have hit a wall in artistic innovation. Some have proposed that the reason physics has stalled in recent decades is that we simply aren't smart enough to understand or deduce what lies beyond the standard model. Perhaps we are hitting walls but don't realize it. The current limitations of AI undoubtedly reflect technical rather than physical limitations much like airplanes were once limited to barely getting off the ground.

why does the weight of a brain matter?

by that logic, a sperm whale should be able to assemble an IKEA product; after all, their brains weigh more than a human's does.

It matters because the human brain has a finite level of processing power limited by the number of neurons which function like crude and unreliable logic gates. The human brain is inherently finite and fallible and thus nothing presents a sufficiently powerful computer from matching or exceeding its capabilities.

[keenir]

You laugh at the notion of AlphaFold making sandwiches

I'm not laughing; I'm asking you a simple question. one you still haven't answered.

Fine then. Yes, the technology and algorithms underlying AlphaFold could easily be adapted to folding lunch meat into shapes suitable for sandwiches.

easily? proteins fold in a lot more directions (including through) than lunchmeat does.

There is no obvious reason why one form of folding comes easily to machine learning techniques while another runs into impossible obstacles.

and this tells me a great deal; for one, you assume that protein folding was easy to program and teach into machines.

why not? human endevours expand over the course of history, and we have never hit a wall of "nope, can't do that"...whereas AI have.

Not yet, but I can easily imagine someone pushing artistic complexity to the limits of what humans can comprehend owing to our inherent cognitive limitations,

you do realize that art includes more than just portraits and landscapes, right?

in which case we will have hit a wall in artistic innovation. Some have proposed that the reason physics has stalled in recent decades is that we simply aren't smart enough to understand or deduce what lies beyond the standard model.

stalling?

can you cite anything to prove that?

The current limitations of AI undoubtedly reflect technical rather than physical limitations much like airplanes were once limited to barely getting off the ground.

and its undoubted because...??

why does the weight of a brain matter?

by that logic, a sperm whale should be able to assemble an IKEA product; after all, their brains weigh more than a human's does.

It matters because the human brain has a finite level of processing power limited by the number of neurons which function like crude and unreliable logic gates. The human brain is inherently finite and fallible and thus nothing presents a sufficiently powerful computer from matching or exceeding its capabilities.

so how is a computer - which needs to be programmed by those humans who are finite and fallible and have crude and unreliable nerves - going to match or exceed a human?

and again, why can't sperm whales with their brains - larger than human minds, just like your unbuilt theoretical proposed AIs - assemble an IKEA product? by your reasoning, they should be able to.

[alice]

and again, why can't sperm whales with their brains - larger than human minds, just like your unbuilt theoretical proposed AIs - assemble an IKEA product? by your reasoning, they should be able to.

Assuming that they have, among other things, opposable thumbs, of course.

[malloc]

I assume everyone here agrees that human cognition arises from the physical brain as opposed to some immaterial soul. Since the brain consists of matter and operates according to physical laws, there is no obvious reason why technology couldn't emulate its processes. If you believe AI cannot replicate all human activities, you must explain what humans possess that computers inherently lack. If you believe that AI cannot advance much further than it currently has, you must explain why the principles underlying it are inapplicable to any new tasks.

[Raphael]

I assume everyone here agrees that human cognition arises from the physical brain as opposed to some immaterial soul. Since the brain consists of matter and operates according to physical laws, there is no obvious reason why technology couldn't emulate its processes. If you believe AI cannot replicate all human activities, you must explain what humans possess that computers inherently lack. If you believe that AI cannot advance much further than it currently has, you must explain why the principles underlying it are inapplicable to any new tasks.

I have no problem with the idea that some kind of AI will some day be able to do everything that the human brain can do. I do have a problem with the idea that the stuff that gets marketed under the label "AI" in our time is able to do everything that the human brain can do.