@lucydev @KatS Because I use it every day, and I can see how much it helps. And to be fair, it primarily helps people who get X done, not the doers of X. Just as automated telephones primarily help those who want to make phone calls (by making them cheaper, faster and much more convenient), not the phone operators who helped to make them in the past.

@miki @lucydev Wait, you're comparing art (visual, written or musical) to operating a telephone switchboard?

@KatS @lucydev Yes. At least the kind of art that people want to consume (and not want to make).

@KatS @lucydev Yes. At least the kind of art that people want to consume (and not want to make).

@miki @KatS don't get the wrong idea, this is pure curiosity.

I had the suspicion that the less someone knows about LLMs or ML, the more they think the tech is capable of, but that suspicion must be false, since you're around

@lucydev @KatS The more you know about LLMs, the more "calibrated" you are about where they work (and don't work) right now. People who don't know much about them are either hypesters (mmaking a company of a thousand LLMs and firing all their employees), or LLM deniers. Both are just as crazy.

I also see not just where LLMs are right now, but where they are going. We went from coding agents being basically a joke a year ago, to them semi-autonomously solving (some) complex mathematical problems and being used for boring gruntwork by world-class, fields-medal-winning mathematicians. They can now also solve an extremely complex GPU performance engineering task that Anthropic used as an interview question for the most brilliant engineers in that discipline, *better than any human given the same amount of time*.

They're still much better at small, well-scoped and bounded tasks than at large open-ended problems, but "small and well-scoped" went from "write me a linked list implementation unconnected to anything in my code" to "write me a small feature and follow the style of my codebase." In a year. What will happen in another year? 5 years? 10 years? God only knows, and he certainly isn't telling.

@lucydev @KatS The more you know about LLMs, the more "calibrated" you are about where they work (and don't work) right now. People who don't know much about them are either hypesters (mmaking a company of a thousand LLMs and firing all their employees), or LLM deniers. Both are just as crazy.

I also see not just where LLMs are right now, but where they are going. We went from coding agents being basically a joke a year ago, to them semi-autonomously solving (some) complex mathematical problems and being used for boring gruntwork by world-class, fields-medal-winning mathematicians. They can now also solve an extremely complex GPU performance engineering task that Anthropic used as an interview question for the most brilliant engineers in that discipline, *better than any human given the same amount of time*.

They're still much better at small, well-scoped and bounded tasks than at large open-ended problems, but "small and well-scoped" went from "write me a linked list implementation unconnected to anything in my code" to "write me a small feature and follow the style of my codebase." In a year. What will happen in another year? 5 years? 10 years? God only knows, and he certainly isn't telling.

@miki @KatS so you're betting unfoundedly that the tech is gonna work right one day?

@miki @lucydev How much thought do you give to the externalities of these things? Their less-desirable impact on the world in which we're trying to live?

@KatS look @miki don't get me wrong but any time i've tried using LLMs for my work, which isn't just some fun side project but actual production-running code, LLMs have been way too unreliable. It also resulted in me knowing jack shit about my own code, which is poison for long term maintainability.

Since these models are just statistically determining the next most likely token based on training data and fine tuning, without any actual understanding or thought behind it, I seriously can't see this tech being reliable enough one day. (reliable compared to humans, i don't seek 100% reliable in this case, natural language is too imprecise for that anyways. i would expect "good enough" as "as good as a professional in the given field")

The other part of the equation is the amount of compute and electrical energy necessary to train and operate such a level, and on that level, there's no way in hell that shit is ever gonna be worth it, financially and environmentally.

i'm not expecting the "make job for phone operators easier", i expect the "when i dial a number, it should be at least as reliable and efficient at routing it correctly as a phone operator would be".

you can call me whatever you want, even llm denier if you need to, but autocorrect on steroids isn't worth exploiting other people's work or boiling our oceans.

@lucydev @miki Similar: I'm not a "denier" - I'm utterly hostile to this mission of eliminating human expertise, knowledge and artistry.

This is pretty impressive, given that I don't even like humans all that much.

@KatS look @miki don't get me wrong but any time i've tried using LLMs for my work, which isn't just some fun side project but actual production-running code, LLMs have been way too unreliable. It also resulted in me knowing jack shit about my own code, which is poison for long term maintainability.

Since these models are just statistically determining the next most likely token based on training data and fine tuning, without any actual understanding or thought behind it, I seriously can't see this tech being reliable enough one day. (reliable compared to humans, i don't seek 100% reliable in this case, natural language is too imprecise for that anyways. i would expect "good enough" as "as good as a professional in the given field")

The other part of the equation is the amount of compute and electrical energy necessary to train and operate such a level, and on that level, there's no way in hell that shit is ever gonna be worth it, financially and environmentally.

i'm not expecting the "make job for phone operators easier", i expect the "when i dial a number, it should be at least as reliable and efficient at routing it correctly as a phone operator would be".

you can call me whatever you want, even llm denier if you need to, but autocorrect on steroids isn't worth exploiting other people's work or boiling our oceans.

@KatS

lmao same actually xD

Wanna be friends?

@miki

@lucydev @KatS Autocorrect on steroids is basically GPT-3 tech. There's a lot more that goes into modern LLMs. A lot of the improvements are due to reinforcement learning, where LLMs learn to predict tokens that actually achieve some outcome, E.G. code that passes tests, answer that is judged "good" by a domain expert. There's still token prediction involved of course, but it somehow turns out that token prediction can get better scores than any human at (unseen) math olympiad questions. And people still say it's not in any way intelligent...

@lucydev @KatS Autocorrect on steroids is basically GPT-3 tech. There's a lot more that goes into modern LLMs. A lot of the improvements are due to reinforcement learning, where LLMs learn to predict tokens that actually achieve some outcome, E.G. code that passes tests, answer that is judged "good" by a domain expert. There's still token prediction involved of course, but it somehow turns out that token prediction can get better scores than any human at (unseen) math olympiad questions. And people still say it's not in any way intelligent...

@miki @KatS if i memorize every possible answer to a specific test, i can pass too. doesn't mean i know shit about fuck.

There's no actual thinking or reasoning involved (and no, reasoning models don't actually "reason"), so yeah, an LLM isn't actually intelligent, it just shows how flawed our tests for intelligence are.

To get some actual intelligence, thinking or reasoning involved, I'd reckon we'd have to fundamentally change something in the architecture of LLMs, and use a fuckton more computing resources for a single model, and considering how much energy the current tech already wastes, and the whole shtick that made LLMs (and more broadly generative AI) work in the first place is "we discovered that there comes a point where the output gets better when we throw rediculous amounts of compute resources on the problem", and it's already getting super difficult to run and maintain.

Honestly, either you're unreasonably optimistic, or you've never taken a look at how things actually work under the hood, but I really recommend you to take a closer look at the technology you praise so much.

A couple things you could take a look at (without an AI summarizer, otherwise you'd learn jack shit):

Attention is all you need, which is the paper that sparked all that AI craze and the development of GPT models and The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity
, which takes a closer look and tests reasoning models to infer strengths and weaknesses of reasoning models with all sorts of levels in problem complexity.

Honestly, before you make any claims about where the tech could be and what it could do, you should have a look at how things actually work under the hood and have a rough idea of how things work, otherwise, no offense, you're just talking out of your arse.

@lucydev @KatS I have very specifically said "unseen questions."

If memorizing answers was a viable strategy to pass that test, humans would have done so.

If you still believe that there's no possible use for a tool that can get gold on a never-before-used set of math olympiad question given a few hours of access to a reasonably powerful computer, and that the existence of that tool will have no interesting impact on the world... I don't know what to tell you.

@miki @KatS > If you still believe that there's no possible use for a tool that can get gold on a never-before-used set of math olympiad question given a few hours of access to a reasonably powerful computer, and that the existence of that tool will have no interesting impact on the world...

How reliable is that source? And if that's true, is it really reasonable to bet everything on this, and let this do all your work when a) you end up completely dependent on the tech and b) utterly destroy the environment in that process?

Real world problems may be less complex but might require much more context.

Oh, and don't get me started on accountability. There's a reason why curl is closing their bug bounty program.

@miki @KatS > If you still believe that there's no possible use for a tool that can get gold on a never-before-used set of math olympiad question given a few hours of access to a reasonably powerful computer, and that the existence of that tool will have no interesting impact on the world...

How reliable is that source? And if that's true, is it really reasonable to bet everything on this, and let this do all your work when a) you end up completely dependent on the tech and b) utterly destroy the environment in that process?

Real world problems may be less complex but might require much more context.

Oh, and don't get me started on accountability. There's a reason why curl is closing their bug bounty program.

the whole ai-bro shtick about "ai democritizes art/programming/writing/etc" seemed always so bs to me, but i couldn't put it into words, but i think i now know how.

ai didn't democritize any of these things. People did. The internet did. if all these things weren't democritized and freely available on the internet before, there wouldn't have been any training data available in the first place.

the one single amazing thing that today's day and age brought us is, that you can learn anything at any time for free at your own pace.

like, you can just sit down, and learn sketching, drawing, programming, writing, basics in electronics, pcb design, singing, instruments, whatever your heart desires and apply and practice these skills. fuck, most devs on fedi are self taught.

the most human thing there is is learning and creativity. the least human thing there is is trying to automate that away.

(not to mention said tech failing at it miserably)