As with almost all technology, AI tech is evolving into different architectures that aren’t wasteful at all. There are now powerful models we can run that don’t even require a GPU, which is where most of that power was needed.
The one wrong thing with your take is the lack of vision as to how technology changes and evolves over time. We had computers the size of rooms to run processes that our mobile phones can now run hundreds of times more efficiently and powerfully.
Your other points are valid, people don’t realize how AI will change the world. They don’t realize how soon people will stop thinking for themselves in a lot of ways. We already see how critical thinking drops with lots of AI usage, and big tech is only thinking of how to replace their staff with it and keep consumers engaged with it.
iopq@lemmy.world 1 week ago
It was also inefficient for a computer to play chess in 1980. Imagine using a hundred watts of energy and a machine that costed thousands of dollars and not being able to beat an average club player.
Give it twenty years to become good. It will certainly do more stuff with smaller more efficient models as it improves
kayohtie@pawb.social 1 week ago
If you want to argue in favor of your slop machine, you’re going to have to stop making false equivalences, or at least understand how its false. You can’t make ground on things that are just tangential.
A computer in 1980 was still a computer, not a chess machine. It did general purpose processing where it followed whatever you guided it to. Neural models don’t do that though; they’re each highly specialized and take a long time to train. And the issue isn’t with neural models in general.
The issue is neural models that are being purported to do things they functionally cannot, because it’s not how models work. Computing is complex, code is complex, adding new functionality that operates off of fixed inputs alone is hard. And now we’re supposed to buy that something that creates word relationship vector maps is supposed to create new?
For code generation, it’s the equivalent of copying and pasting from Stack Overflow with a find/replace, or just copying multiple projects together. It isn’t something new, it’s kitbashing at best, and that’s assuming it all works flawlessly.
With art, it’s taking away creation from people and jobs. I like that you ignored literally every point raised except for the one you could dance around with a tangent. But all these CEOs are like “no one likes creating art or music”. And no, THEY just don’t want to spend time creating themselves nor pay someone who does enjoy it. I love playing with 3D modeling and learning how to make the changes I want consistently, I like learning more about painting when texturing models and taking time to create intentional masks. I like taking time when I’m baking things to learn and create, otherwise I could just go buy a box mix of Duncan Hines and go for something that’s fine but not where I can make things when I take time to learn.
And I love learning guitar. I love feeling that slow growth of skill as I find I can play cleaner the more I do. And when I can close my eyes and strum a song, there’s a tremendous feeling from making this beautiful instrument sing like that.
spicehoarder@lemmy.zip 1 week ago
Oh my God, that’s perfect. It’s kit bashing. That’s exactly how it feels.
iopq@lemmy.world 1 week ago
Stockfish can’t play Go. The resources you spent making the chess program didn’t port over.
In the same way you can use a processor to run a completely different program, you can use a GPU to run a completely different model.
So if current models can’t do it, you’d be foolish to bet against future models in twenty years not being able to do it.
frezik@lemmy.blahaj.zone 5 days ago
Buy any bubble memory lately?
I have a book from the early 90s which goes over some emerging technologies at the time. One of them was bubble memory. It was supposed to have the cost per MB of a hard drive and the speed of RAM.
Of course, that didn’t materialize. Flash memory outpaced its development, and it’s still not quite as cheap as hard drives or as fast as RAM. Bubble memory had a few niche uses, but it never hit the point of being a mass market product.
Point is that you can’t assume any singular technology will advance. Things do hit dead ends. There’s a kind of survivorship bias in thinking otherwise.
kayohtie@pawb.social 6 days ago
I think the problem is that you think you’re talking like a time traveler heralding us about the wonders of sliced bread, when really it’s more like telling a small Victorian child about the wonders of Applebee’s and in the impossible chance they survive to it then finding everything is a lukewarm microwaved pale imitation of just buying the real thing at Aldi and cooking it in less time for far tastier and a fraction of the cost.
bridgeenjoyer@sh.itjust.works 6 days ago
Its because the tech bros have 0 empathy or humanity. Llm slop is perfect for them.
Deflated0ne@lemmy.world 1 week ago
Show me the chess machine that caused rolling brown outs and polluted the air and water of a whole city.
I’ll wait.
iopq@lemmy.world 1 week ago
Servers have been eating up a significant portion of electricity for years before AI. It’s whether we get something useful out of it that matters
Deflated0ne@lemmy.world 6 days ago
That’s the hangup isn’t it? It produces nothing of value. Stolen art. Bad code. Even more frustrating phone experiences. Oh and millions of lost jobs and ruined lives.
It’s the most american way possible that they could have set trillions of dollars on fire short of carpet bombing poor brown people somewhere.
CorvidCawder@sh.itjust.works 6 days ago
Not even remotely close to this scale… At most you could compare the energy usage to the miners in the crypto craze, but I’m pretty sure that even that is just a tiny fraction of what’s going on right now.
jj4211@lemmy.world 6 days ago
It probably would have if IBM decided that every household in the USA needed to have chess playing compute capacity and made everyone dial up to a singular facility in the middle of a desert where land and taxes were cheap so they could charge everyone a monthly fee for the privilege…
outhouseperilous@lemmy.dbzer0.com 1 week ago
Not the same. The underlying tech of llm’s has mqssively diminishing returns. You can akready see it, could see it a year ago if you looked. Both in computibg power and required data, and we do jot have enough data, literally have nit created in all of history.
This is not “ai”, it’s a profoubsly wasteful capitalist party trick.
Please get off the slop and re-build your brain.
iopq@lemmy.world 1 week ago
That’s the argument Paul Krugman used to justify his opinion that the internet peaked in 1998.
You still need to wait for AI to crash and a bunch of research to happen and for the next wave to come. You can’t judge the internet by the dot com crash, it became much more impactful later on
outhouseperilous@lemmy.dbzer0.com 1 week ago
No. No i don’t. I trust alan Turing.
Dangerhart@lemmy.zip 1 week ago
It seems like you are implying that models will follow Moore’s law, but as someone working on “agents” I don’t see that happening. There is a limitation with how much can be encoded and still produce things that look like coherent responses. Where we would get reliable exponential amounts of training data is another issue. We may get “ai” but it isn’t going to be based on llms
iopq@lemmy.world 1 week ago
You can’t predict how the next twenty years of research improves on the current techniques because we haven’t done the research.
Is it going to be specialized agents? Because you don’t need a lot of data to do one task well. Or maybe it’s a lot of data but you keep getting more of it (robot movement? stock market data?)
Dangerhart@lemmy.zip 6 days ago
We do already know about model collapse though, genai is essentially eating its own training data. And we do know that you need a TON of data to do even one thing well. Even then it only does well on things strongly matching training data.
Most people throwing around the word agents have no idea what they mean vs what the people building and promoting them mean. Agents have been around for decades, but what most are building is just using genai for natural language processing to call scripted python flows. The only way to make them look coherent reliably is to remove as much responsibility from the llm as possible. Multi agent systems are just compounding the errors. The current best practice for building agents is “don’t use a llm, if you do don’t build multiple”. We will never get beyond the current techniques essentially being seeded random generators, because that’s what they are intended to be.
jj4211@lemmy.world 6 days ago
It might, but:
iopq@lemmy.world 6 days ago
Stockfish on ancient hardware will still mop up any human GM
jj4211@lemmy.world 6 days ago
Umm… ok, but that’s a bit beside the point?
Unless you mean to include those 1980 computers, in which case stockfish won’t run on that… More than about 10 year old home computer would likely be unable to run it.
jaykrown@lemmy.world 1 week ago
Twenty years is a very long time, also “good” is relative. I give it about 2-3 years until we can run a model as powerful as Opus 4.1 on a laptop.
iopq@lemmy.world 1 week ago
There will inevitably be a crash in AI and people still forget about it. Then some people will work on innovative techniques and make breakthroughs without fanfare
Image
frezik@lemmy.blahaj.zone 6 days ago
We really need to work out the implications of the fact that Moore’s Law is dead, and technology doesn’t necessarily advance on a exponential path like that, anyway.
The cost per component of an integrated circuit (the original formulation of Moore’s Law) is not going down much at all. We’re orders of magnitude away from that. Nodes are creating smaller components, but they’re not getting cheaper. The fact that it took decades to get to this point is impressive, but it was already an exception in all of human history. Why can’t we just be happy that computers we already have are pretty damned neat?
Anyway, AI is not following anything like that path. This might mean a big breakthrough tomorrow, or it could be decades from now. It might even turn out not to be possible; I think there is some way we can do AGI on computers of some kind, but that’s not even the consensus among computer scientists. In any case, there’s no particular reason to think LLMs will follow anything like the exponential growth path of Moore’s Law. They seem to have hit a point of diminishing returns.
Semi_Hemi_Demigod@lemmy.world 6 days ago
The 19th and 20th centuries saw so much technological advancement and we got used to that amount of change.
That’s why people were expecting Mars by the mid 80s and flying cars and other fanciful tech by now.
The problem is that the rate of advancement is slowing down, and economies that demand infinite, compounding growth are not prepared for this.