No matter how much you dress up whatever AI service has gaslit you into believing it’s sentient, generative AI is inherently limited, impossibly expensive and economically unviable. Its services cost too much to run, its progenitors have no path to profitability, and no amount of rigged benchmarks and anecdotal examples of theoretical engineering teams that are “10x’d” can make up for the fact that you can’t measure the cost of an LLM-driven task or its return on investment.
Anyone claiming that you have to “measure AI’s ROI differently” is attempting to con either you or themselves. While it’s tough to measure the ROI of a particular worker or project, most workers and projects don’t increase your operating expenses by anywhere from 10% to 100% under the vaguest of promises that you might be “doing the future.” AI is calamitously expensive and, despite years of promises of it getting cheaper for both those running AI services and its customers, costs have only ever increased.
I think that’s by design. AI labs want their costs to be high so that they can continue growing at ridiculous rates, all so that they can keep feeding money to their hyperscaler compute partners who then invest that money right back into them, creating further reasons to keep buying NVIDIA GPUs, so that NVIDIA can then invest that money back into either AI compute providers (who OpenAI and Anthropic pay) or the AI labs themselves.
Concepts like “efficiency” or “cost reduction” run counter to the greater narrative of AI’s voracious sprawl of data center capex and still-theoretical AI revenue. If OpenAI or Anthropic were to seek profitability or sustainability (assuming these things were possible), that would create less demand for AI compute, which would mean less demand for Azure or Google Cloud or Amazon Web Services or CoreWeave or Oracle Cloud Infrastructure, which would in turn mean less demand for NVIDIA GPUs.
The problem with this marvelous plan is that at some point there had to be an honest transaction — real, honest, sustainable demand based on a reliable product that people liked paying for because they understood its value. Right now, AI revenues are either chaotically experimental or so thoroughly-subsidized that labs are giving away hundreds of dollars a user in the hopes that at some point said user might want to pay even more money for measurably less value, the kind of proposition you make when you think your customers are fucking idiots.
You must log in or # to comment.
the kind of proposition you make when you think your customers are fucking idiots
To be fair, a huge proportion of customers – especially AI customers – are fucking idiots.
The ceos for sure
The people inflating the bubble have to know what they’re doing, right? It’s hard to come up with an economic scenario where the market doesn’t crash and burn as a result.
I can’t shake the feeling that two key motivations for bleeding all this cash are A. For moving more wealth to the top (which, y’know, is already happening in umpteen different ways beyond AI), and B. For quickly building surveillance infrastructure, seeing as how eager the Pentagon/DHS is about AI
For moving more wealth to the top
https://en.wikipedia.org/wiki/Dot-com_bubble
Some companies did survive the dot-com bubble popping and went on to become dominant in what was left of the market.
We’re looking at a desperate struggle of AI companies convinced they’re going to be the ones that make it, being encouraged by other companies milking them for everything they can before the crash.
A company that sells RAM, GPUs, etc will of course sell every last unit they can to AI companies – they want to move as much product as possible (at inflated prices) before reality kicks in and demand drops. And they’ll have a hungry consumer market to fall back to after making a shit-load of money off of the AI rubes.
The people building data centers are rushing to get them built for the same reason. They don’t give a shit about anything that happens after they’ve been paid.
I suspect part of it is that every company wants to be the one to outlast the competition and become the de facto AI company everyone uses. They want to be the 90’s Microsoft of AI. To this end, they’re all setting fire to giant piles of money.
Maybe I’m wrong but even if one clear winner emerges, I don’t see AI becoming that big of a market. They’re trying the heroin dealer approach of giving away cheap or free product so they can get people hooked and then jack up the prices on a bunch of people who will lie, cheat, and steal to get their next fix. Except AI slop isn’t heroin, it’s just slop.
Slop that produces inaccurate results and then tells you it did no such thing.
For about $2000, I picked up a Mac Studio with 96 gigs of RAM, which is effectively all VRAM thanks to Apple’s architecture. It doesn’t have the raw number-crunching power of the big GPUs but with all that space, you don’t need to worry much about the size of the model until they start getting really big, so it’s pretty easy and flexible.
I’m able to do basically everything that others are doing with AI, entirely locally. It generates text and writes code (it’s no Opus, but probably on par with the best of a year and a half ago), images, videos, songs, all that stuff (those last few are garbage but it’s basically the same level garbage that the cloud models are making). And I have total privacy, will never get a surprise price hike or lose access to a model I like, and know exactly how many bottles of water I’m using to cool it (zero).
It’s not heroin, it’s weed. There will be a market for it, but, like, you can also just make it yourself. 90% of what people want will just get done on device. I can’t see any way this turns into a trillion dollar industry.
Slop that produces inaccurate results and then tells you it did no such thing.
I use Claude at work and local LLMs at home, and they all produce good code. The Kagi agents work pretty well at searching for information that would have taken me 30-60 minutes to find. Although, I generally favor thinking models because they are good at edge cases.
You have to know how to provide good context for the situation, like examples, prior art, documentation, etc. Many people have a hard time even expressing an idea to a group of humans. Imagine your (pointy-haired) boss shows your department a picture in the next meeting:
B: “Go make this thing!”
D: “What thing?”
B: “Here. This thing. Make something just like it for our company!”
D: “Well, we can’t just copy it outright. What color should it be? What do we want to improve on? How do we tie it to our existing software?”
B: “I don’t know. That’s your job to figure out, right?”
That’s how most people treat LLMs. Garbage in, garbage out.
Well there you go. Another argument in support of AI being not nearly as big of a deal as the fanboys would have people believe.
If you blame the users (e.g. Steve Job’s “you’re holding it wrong”) it rarely ends well.
Depends on what you mean by “fanboys”. No, the hype isn’t as real as these rich assholes make it out to be. LLMs aren’t going to replace all human workforce everywhere ever, like some of these techbro dipshits quickly find out.
But, I’m still going to use technology that produce a tool in two minutes what would have normally taken me two hours to do. Sure, I have to code review it, but that doesn’t take nearly as long as the work itself.
I heard someone describe LLMs as useful in the context of being handy plugins which tallies with what you’re saying. In certain situations they’re useful. The problem is that they’re being completely misrepresented and sold as being able to do things that they can’t do consistently or well so the hype and astronomical sums of money being thrown around is problematic.
I use them for a bit of coding leverage, and they require a fair bit of… not hand holding, but explicitness, maybe. Otherwise they wander down the path of statistical averageness, and the average code they saw during training was shit.
So they’ll happily serve up a pile of inefficient dogshit, complete with working tests and docs and all of that. And then they’ll happily refactor it at your suggestion to slowly turn it into something that’s resource friendly and generally secure and generally expandable.
But there’s no way for them to do that by themselves if you don’t have the basic domain knowledge to guide them in the right direction. You’ll just get average code, and once you’ve been in the game for a few decades, you realise that nice, efficient, quality code is the rarity, not the norm.
I’ll add ( C ): drive the price of general-purpose and performant hardware to the moon so everyone outside of the upper-class doesn’t have a choice about whether to use cloud vs self-hosted anything, which is just another way of describing who has control over your experience and what you can do with the hardware you bought. I mean, Apple already makes it difficult to run whatever you want on your own system (and blocking certain things entirely), even before getting to the whole app store walled garden thing, so how long before Microsoft pulls the trigger on making TPM enablement mandatory for OEMs, with a very short list of allowed certificates? At that point, you’ll have to choose between “good-ish hardware, but app store apps only” and “meh hardware, full control, but no access to anything useful because everything requires hardware attestation of some kind”.
That’ll work until they create for themselves new competitors that are good enough at previous (normal) market prices. And you don’t need incredible high foundry tech for that, so it’s closing up on them eventually.
Fair point! Counterpoint: Western countries tend to use national security to push back against effective competition (see also: EVs), so I don’t think it’s strictly a technical or knowledge issue, and that while eventually there will be competition, whether or not we can legally buy it might matter more.
(Whether security concerns are valid is a separate point outside the scope of my comment: on one hand, Wi-Fi cameras you buy are basically botnet members, on the other hand the allegations against (IIRC) supermicro were never (haven’t checked in a while, could have changed but I’d probably have heard already) substantiated so the whole “embedded surveillance hardware” theory is still a theory, meaning we should expect the quality of the firmware to be inconsistent. Not exactly something new: consumer BIOS/EFI tends to be buggy, manufacturers focus on the enterprise versions)
I think that folds pretty well into both points A (reserving personal hardware for the wealthy) and point B (the ability to mandate software installations is a valuable mass surveillance touch-point)
I don’t disagree! I think it’s worth mentioning separately if only to show the mechanics of how it would play out, each step is too distant from any other step otherwise.
