AI sucks. Hating it is not enough
16 points by Signez
16 points by Signez
And yet we cannot uninvent it, we cannot boycott it out of existence
I think I don't need to read further.
You see, we don't need to boycott it out of existence. We have to resist, push and fight back just long enough until it all bursts. Quite a lot of us are doing just that, and like every bubble before, this one will burst too.
It's not a revolution. It's not a new era. It's a bubble. Bubbles burst.
The technology is not suddenly going to cease existing when the bubble bursts.
And this is also called out in the post later in the post:
But the problem here is that such a bubble, if and when it happens, will look a lot more like dot-com than NFTs.
I'm not responding to the article, I'm responding to the comment I'm replying to. I agree with that particular assertion from the article. The comment I'm replying oo seems to imply that the technology and its usages will go away once the bubble bursts.
It won't, indeed. But with the money well drying up with the burst, who'll pay for the massive requirements of scraping & training, and everything else?
With cryptocurrencies, the tech stayed, and it is still present, because deploying yet another scam on the blockchain is relatively cheap. Building a model, and running it at scale, is not.
It's not a revolution. It's not a new era. It's a bubble. Bubbles burst.
Too bad you stopped reading, the article addresses exactly that not long after:
Let’s begin with the most popular argument among my peers: that AI is a bubble. That might still be true. Ed Zitron’s critics could be wrong: the dodgy accounting might be too damning, and it could still all come tumbling down.
But the problem here is that such a bubble, if and when it happens, will look a lot more like dot-com than NFTs. Real value, fake valuations.
The speculation stops, the technology stays around. Reshaping our lives in drastic ways, for better or worse, just like the internet.
The thing is, AI works. To some level. And though the current prices are likely unsustainable, the true prices are likely low enough to find loads of customers. Even if there’s no longer enough money to pay for the training of new models, there will almost certainly be able to pay for the use of existing ones. And that, will slowly fund the new ones.
I disagree with the author, and am with Ed Zitron here. There is absolutely no way GenAI can be sustainable long-term. Even the higher, token-based prices are subsidized, and companies are finding those too high (let alone individuals). Old models are also useless, because they don't keep up with the rapidly moving trends. A model that can't help you work with the currently trendy thing is useless. Try using a model from 2022, see if it's of any use today.
Without constant training, and constantly releasing new models, whatever miniscule value there may be in GenAI, it is gone.
Once the bubble bursts, "slowly funding new ones" is not going to be a thing, an old model rapidly decays in usefulness (that's why they're scraping and training new models 24/7), noone will be paying thousands of dollars a month for an obsolete model.
Old models are also useless
They are, yes. But I’m talking about the current models, summer 2026 (or later depending on when the bubble will burst). I hear those are not useless, and they certainly won’t go away.
an old model rapidly decays in usefulness
For use cases requiring fresh information, they sure do. More stable stuff like programming though? No way. Imagine a Claude Mythos trained with data from 2020. I bet it would find about as many bugs as the one we have, because programming techniques and vulnerability classes haven’t changed that much since.
I say that, kinda hoping the burst will be as complete as you predict. I’d like the thing gone. I just don’t think it will. And even if it does, there’s a fair chance someone will find a way to make models more efficient in a relatively near future. Currently they’re glorified Markov chains, and any structure or insight emerges from the statistics of surfaces features (words, pixels…), rather than an actual model of the world that governs it. But if we find a way to encode a model of (parts of) the world itself, similar to how Alpha Go operated on a model of the board, we might reduce costs to a point where the current hype is accurate.
The hope, should such models arise, is that they’ll be specialised enough to be directed to our most pressing problems first, and may even be easier to regulate.
But I’m talking about the current models, summer 2026 (or later depending on when the bubble will burst). I hear those are not useless, and they certainly won’t go away.
They might be useful now, but in 2 years, they'll be old too. Considering how much it costs even to operate them, good luck keeping these models alive for two years, and make them profitable enough to support training new models.
More stable stuff like programming though?
Programming ain't stable. Parts of it are, yes. Other, large parts of it, are not. A model from last year won't know about the hot new post-quantum crypto library that was published two weeks ago.
And even if it does, there’s a fair chance someone will find a way to make models more efficient in a relatively near future.
I wouldn't hold my breath. They've been pushing this AI thing for a fair couple of years now, it's not new tech. If we can trust Ed Zitron, then as demand goes up, so do the cost, so they're not exactly succeeding at making them more efficient. The AI companies themselves saying they need trillion dollars in the next few years suggests they're not expecting to become more efficient either.
How would they'd suddenly become more efficient without money, when they couldn't while being thrown infinite amounts of it at them?
Believing that a magical, more efficient model will appear is a fairytale.
Programming ain't stable.
I might be biased because I tend to focus more on the fundamentals, but… as a C and C++ application programmer, my feeling is that it is. No major change in years. A new technique here and there, but practically no breakthrough that measurably impacts the way I should program.
A model from last year won't know about the hot new post-quantum crypto library that was published two weeks ago.
Can you describe more precisely how including very recent libraries in the LLM’s training set may impact its usefulness? From where I stand, this is surface level stuff. It won’t really influence the responses of the LLM, except when we’re getting right into those libraries respective niches (and those niches are new). Also, we can always ask the LLM to look at it and tell you about the entry points. Just because LLMs are out of date doesn’t mean search engines are.
How would they'd suddenly become more efficient without money, when they couldn't while being thrown infinite amounts of it at them?
I’m not thinking of current companies, but other approaches altogether. I’ve heard Yann LeCun has promising ideas. I do not trust this will pan out. But neither do I trust that it will not.
Not to be the doomsayer, but AI is by far not the first or the worst thing that happened to the world. The current for of its existence was a path with many bad steps there. So in a way its current form of its existence seems more like symptom of many broken things in the world rather than the cause, an amplifier of many diseases, but if we look at other examples there is a pattern of dealing with symptoms in barely working manners.
Law like everything that has any form of "evolution" (so changes over time) has the issue of the world around it adapting so its "arms races" and usually lawyers work faster than the legislative.
On top of that we now live in a world where the experts on regulation often have biases due to being (in one way or another) related to the thing they wanna regulate, so there are biases and incentives that aren't matching those of the general public. In the end that more often than not just leads to solidifying monopolies or duopolies. We see that in so many areas and while I think it's much better than nothing it sometimes also has the side effect of just accepting the root cause (which in this case isn't "AI existing" as some people like to pretend).
I think if we never fight underlying symptoms, ie. the incentives to fuck up the world for everyone (which at least to some degree are artificial) we will just see more forms of people doing that. It sometimes feels like we just blame the next best thing for misery as if the kind of problems is new, when often it's just the next amplified version. A bit like guns didn't create violence but certainly amplified it or how fast food didn't create obesity, but amplified it. We see AI being an amplifier for many things that existed before. I do not think it has to be, just like producing dishes in a fast organized manner isn't the problem, but producing then in a way to be addictive and completely ignoring health is. Forms of transportation are a similar topic. If we had public transport, good roads for things like bikes, but fire-brigade, ambulances and so on still having way to quickly get to you a lot of the problems of present day transportation would be at least greatly reduced. Now transportation is largely just a problem amplifier.
Like with a lot of controversial topic people are talking about different things at times. So "is it bad in general", "is it bad unregulated" or "is it bad with how current incentives are".
Near the beginning of the article, there’s a link to another article warning that AI is much bigger than you think. Now that article, unlike OP, seems to be pro-AI, but that’s not what I’m after. The real problem, is what it recommends we do about it.
While OP acknowledges that the global impact of AI requires collective action, the pro-AI article recommends individual action, and of the worst kind: secure your position, adapt, position yourself for the coming change. As if it would solve anything: either only some people do that, and they will get on top while everyone else get left behind, or everybody follows the advice, and suffer all the same.
That overly individualistic lens is at the root of many, many problems. The chief example being unemployment: "Want a job? Just git gud!", in practice, it just means other people won’t get the job. And yet so many governments go for the individual who is allegedly not skilled enough, or not motivated enough, to get a job, instead of addressing the actual problem and make sure there are more jobs to begin with. (Note: the framework that requires us to have jobs is questionable to begin with, but I’d be veering off topic.)
I’m glad to see an article that doesn’t fall into that trap.
I'm sorry, this reads like "I hate it but I was one-shotted by it". I fail to buy the "trust me, bro." It spends a huge number of words tangentially touching on the financial issues that foreshadowed the downfall of the AI bubble as early as 2024 but ultimately just assumes they don't matter. And my eyes slide off it. And the writing style is not quite as overwrought as Moldbug, but it comes across as informed by that style. I would not recommend this essay to anyone.
I think if you dislike the writing style (I do too, I would agree that it's "overwrought") there are better ways to make the comparison than to compare it to an author who is best known for being a fascist, in the same way that I wouldn't say an artist's style is "reminiscent of Hitler".
Preaching is more effective when done to non-believers. Unless your metric is the volume of clapping, of course.
people say that a lot, but I think building consensus among believers can be a lot more important and effective than trying to convert nonbelievers. consensus is the ground work for action.