“We’re in an LLM bubble,” Hugging Face CEO says—but not an AI one

The risks of AI investment in manufacturing and other areas are less clear.

See full article...

 

[dobrien75]

When the AI bubble pops, there will be collateral damage.

The take I heard yesterday is that the data centre builds are largely unusable for other computing tasks. Super expensive in build, maintenance and running costs, and barely usable for other things besides LLM processing

 

[WCR-790]

At last. The Roomba was AI in it's time, and didn't need a football field of GPUs to navigate my living room. Facial recognition, now about 20 years old, was AI to those who developed it, but doesn't need $100B of data centers to function. Machine learning, neural networks, those things have more applications than just LLM. As one of the developers of the Roomba wrote 35 years ago: "Elephants Don't Play Chess".

 

[jdale]

Funny, since someone in the last thread was trying to argue that LLMs aren't overhyped nonsense.

You know what's super cool? "AI" that figures out how to fold hundreds of thousands of proteins in the time it used to take to figure out, like, five.

You know what is decidedly less cool? Lying chatbots that pretend to be human and services that make fake music and art based on the works of countless thousands or millions of actual, real, human, and thoroughly uncompensated artists.

Just taking down the LLMs wouldn't do anything about other forms of generative AI, e.g., for creating music and art and video.

Is generative AI for images, music, and video also operating at a financial loss, so it's unsustainable? I don't know. Image generation can run locally, so there could be viable products even without being propped up by VCs. I don't know whether that's really feasible for music and video.

 

[Missing Minute]

with all the recent comments about Bezos and other techbro CEOs picking names for projects/products/companies that seem to tempt the gods of nominative determinism or expose a near sociopathic misunderstandings of the source material


why ?
View attachment 122532

(sorry if I missed the explanation)

 

[MilesArcher]

He's not wrong, but it doesn't matter. When the Dotcom bubble burst, all software companies took a hit. Heck, all companies took a hit since there was a huge decrease in the money supply. Same thing is going to happen next year. The fact that there's other value out there is immaterial. The other value is no where near the insane valuations and revenue expectations for the LLM companies, the hyperscalers and Nvidia.

 

[poke 532810]

with all the recent comments about Bezos and other techbro CEOs picking names for projects/products/companies that seem to tempt the gods of nominative determinism or expose a near sociopathic misunderstandings of the source material


why ?
View attachment 122532

(sorry if I missed the explanation)

The name is based on this emoji: , technically known as “Smiling Face with Open Hands”. It’s even shown at the top of the page if you go to https://huggingface.co/. However, I think your version does a superior job of capturing the essence of generative AI.

 

[Granadico]

My manufacturing company has been pushing us to use AI for the past year for no reason, and just today i heard someone mention about it being a bubble. A week or two ago my mom who can barely use a smartphone talked about investing in AI. We're definitely "slowly then suddenly" phase.

I can only hope that the coming recession doesn't turn into a full-blown depression, and that it brings PC component prices crashing so i can buy a gaming PC to hold me over in between going to food banks.

 

[graylshaped]

It's not a bubble because it doesn't work (or because it does), the promise was not a working AI for the masses, the promise was infinite money for investors, which would fail even if everything went perfectly. The bubble won't pop because AI doesn't work, it will pop because it never made the amount of money it was (unrealistically) expected to.

“Both” is also valid here.

The unicorns and rainbows functionality promises were needed to justify the financial math that with each and every circular investment gets even more jaw-droppingly dumb.

 

[zogus]

Just taking down the LLMs wouldn't do anything about other forms of generative AI, e.g., for creating music and art and video.

Is generative AI for images, music, and video also operating at a financial loss, so it's unsustainable? I don't know. Image generation can run locally, so there could be viable products even without being propped up by VCs. I don't know whether that's really feasible for music and video.

Pulling AI tasks from the cloud doesn’t necessarily make it economically viable. It just means you pay for your own server and maintenance instead of paying Amazon to do it for you. It makes sense in some situations, but not in many others. Among the people in the production industry I’ve spoken to, the motivation for local image AI is usually not cost, but risk of data theft.

 

[Marlor_AU]

There are plenty of “AI” applications that aren’t going anywhere. Transformers and neural networks are no less powerful than they were before this “bubble”, and are still improving at a decent pace.

For example, the use of ML for denoising, upscaling, sharpening, subject isolation, etc. is here to stay. Autoencoders are a powerful and effective tool.

 

[Geebs]

ChatGPT delivered semantic search. I would argue Google still hasn’t delivered semantic search. Most of the rest of consumer AI is just a better UI, sometimes on top of real ML advances. The rest is just piffle, puffery, and some huge VC losses in 2026.

Which still raises the question of what use is semantic search when the material being searched is 99% slop generated en masse by the exact same tools? LLMs haven’t delivered search, they’ve incapacitated it.

 

[Madestjohn]

The name is based on this emoji: , technically known as “Smiling Face with Open Hands”. It’s even shown at the top of the page if you go to https://huggingface.co/. However, I think your version does a superior job of capturing the essence of generative AI.

forcing itself down our collective throats ?


thanks for explaining


though still bewildered by the idea of naming your company the written description of an emoji

that’s like trying so hard to be ‘clever’ you give yourself a lobotomy

 

[noxylophone]

My manufacturing company has been pushing us to use AI for the past year for no reason, and just today i heard someone mention about it being a bubble. A week or two ago my mom who can barely use a smartphone talked about investing in AI. We're definitely "slowly then suddenly" phase.

I can only hope that the coming recession doesn't turn into a full-blown depression, and that it brings PC component prices crashing so i can buy a gaming PC to hold me over in between going to food banks.

I recall some story from ‘29 that some New York broker sold when his shoe-shine boy gave him stock tips.

 

[concernUrsus]

I have used chatGPT as a better Microsoft Word spell/grammar check. It does help with sentence flow. Beyond that it is a glorified Google search (better than Google search but probably not for the cost).

 

[star-strewn]

Honestly, the fact that "AI" is now synonymous with "generative AI" and "LLM chatbots" is one of the most infuriating aspects of this latest boom. As a software engineer, I have passing familiarity with machine learning and traditional "AI" heuristics like task planning, resource allocation, and multi-agent coordination. These AI techniques have real value in certain domains.

It's disappointing that by calling "generative AI" just "AI" in the public sphere and overhyping it to death, these unrelated areas of AI research will inevitably be caught in the bubble pop.

 

[A Man A Plan A Canal Panama]

Been saying exactly this for many years. LLMs are a hoax, and they are casting shade on forms of "AI" that can actually work. There have been multiple AI Winters and there's another one coming. Unfortunately this time it's tangled with VC and PE scam teradollars which threaten to ruin the economy and ecology of Earth

 

[EricM2]

Funny, since someone in the last thread was trying to argue that LLMs aren't overhyped nonsense.

You know what's super cool? "AI" that figures out how to fold hundreds of thousands of proteins in the time it used to take to figure out, like, five.

Yes, that's super cool, but no sociopathic multi-billionaire or hedge fund will senselessly throw billions at it, because there's only business cases that promise to make them just a few dozen Dollars per Dollar invested. Like curing cancer or so.

You know what is decidedly less cool? Lying chatbots that pretend to be human and services that make fake music and art based on the works of countless thousands or millions of actual, real, human, and thoroughly uncompensated artists.

Also true, but the business case "replace workers with chatbots, lying or not" was advertised to them to make them thousands of Dollars per Dollar invested. So that's where most of the investment ended up.

The pop of the LLM bubble will be a likewise over-reaction like the current bubble, hurting everything AI, hyped or not. So I guess it's important for entities like hugging face to build some mental distance in the mind of investors between themselves and LLMs.

 

[Zeppos]

It is a tough bubble. Media is openly reporting that it is a bubble. But we all keep stampeding towards the wall.

Too many are still in denial. Too much invested?

 

[Missing Minute]

we can agree there’s froth at the foundation‑model p&l while also noting the application layer is already monetizing and the engineering curve keeps cutting cost. i’m not betting on a basilisk; i’m betting that “agents + tools + falling $/token” keeps compounding.

How is falling $/token going to make it more likely that it will be possible to have sufficient profit to earn back the money invested?

 

[Carewolf]

He should replace “AI” (a marketing term since it’s inception I. The 60s or 70s, can’t remember), with ML, and then I’m OK with that statement.

Thing is that ML is everything invented before LLMs, and LLMs got tacked “AI” moniker exclusively to hype them.

But LLMs arent really learning and are no more ML than they are AI.

 

[Murry Wilson]

imagine this creep turning his baseball cap around backwards and thinking: "perfect"

 

[zogus]

Reminder that people were saying tech stocks were a bubble in late 1998 and then the Nasdaq fully doubled in price during 1999-2000.

Nasdaq was at 2,192 in Dec 1998, and peaked at 4,696 in Feb 2000. But it crashed after that and did not achieve 2,192 again until late 2005, and 4,696 until 2015. The question is--are we at Dec 1998, Feb 2000, or somewhere else?

 

[Marlor_AU]

It is a tough bubble. Media is openly reporting that it is a bubble. But we all keep stampeding towards the wall.

Too many are still in denial. Too much invested?

Bubbles don't pop until the fear of losing out in a crash overcomes the fear of missing out in a boom.

In other words, the extent of likely losses needs to outweigh the promises of potential gains. When huge up-sides are promised, this can take an awfully long time to occur.

And AI/ML really does have a lot of promise as a transformative set of technologies... just not in the form many people think, and not at the time-scales that optimists suggest.

 

[JonathanSmith]

Yes, that's super cool, but no sociopathic multi-billionaire or hedge fund will senselessly throw billions at it, because there's only business cases that promise to make them just a few dozen Dollars per Dollar invested. Like curing cancer or so.

The biggest name in protein-folding neural networks is google, afaik. I'm not defending the chatbots, but the real work IS still trucking along in the background.

 

[ibad]

That's what I've been saying! it's an LLM bubble, not a general AI bubble! Totally agree with him... but.... as others have said, LLMs are a huge part of AI these days and when they go down it will affect the rest of the AI industry, as well as the whole economy.

Still... I do believe that LLMs do have legitimate use cases and demand. They will just not lead to AGI or provide the commensurate returns. As long as LLM companies get efficient in terms of compute and maybe shift their business models, possibly to include seedy crap like product placement or adult content, they will have a viable product and some profits. But it's not gonna be AGI and any investors that were counting on that level of return within 5 years will get hosed.

 

[Yangguang]

But who will be using the specialized models? Humans? Of course not. It will be the general models using specialized models using even more specialized models. Models all the way down...

 

[Fred Duck]

As one of the developers of the Roomba wrote 35 years ago: "Elephants Don't Play Chess".

A clever reference to the fact that in the hit experience chess, the bishop was originally an elephant.

Also, have you seen Ray Villafane and Sue Beatrice's sand sculpture of Chessie Trunkston?
https://digitalsynopsis.com/design/elephant-playing-chess-with-mouse-sand-sculpture/

 

[Northbynorth]

ML models has a lot of use cases. My field has used them for 30 years.

LLMs write nice text and have better grammar than most people, but they are used way outside their home territory.

I was asked recently if it was feasible to replace our huge documentation and safety instruction system with queries to an LLM. My answer was - Is it ok if same/similar queries/prompts seldom give the same result? He went silent.

 

[Yeay right]

So yeah, maybe I'm just one more normie who will someday be consumed by the basilisk, but I'll believe in the digital sampo when I see it and not before.

Basilisk is already dead. Companies are already ready to burn all their cash and bet their existence on reaching AGI, no looming threats are required.

a near sociopathic misunderstandings of the source material

And torment nexus is another reference that is so overused that it might as well get retired.

I am not sure why capitalism system takes so weird turn. Is it just a mad scramble to find a new market because increasing share in existing ones is too hard and not profitable enough?

 

[WereCatf]

But who will be using the specialized models? Humans? Of course not. It will be the general models using specialized models using even more specialized models. Models all the way down...

They're being used all the time quite literally everywhere and have been for at least a decade now. Just think of e.g. how many IP-cameras these days come with object detection features? Yeah, those use these AI, but not LLMs.

As an aside, I'd love to learn how to train custom object detection models. The tech is enormously useful and requires far less computing power than LLMs, like e.g. all of my IP-cameras' object detection routines are running on a 7th-gen Intel i7's iGPU without any issues.

 

[alexnode]

There is a bubble, and there and I agree, I had a discussion with a financier friend about 3 years ago, and our opinion was aligned that in the first phase we needed hardware, second phase software/ models, third phase and the most expensive...data and professional training.
The bubble is around hardware and the insane amounts of LLMs around... while it should have been on data at this stage, doctorate level data with armies of people working on them. I could see more interest on documenting plants, with photos, SEM scans, timelapses of growth, soil and chemical data etc than putting another 5M gpus in a datacentre.
This is why I prefer a bit deep mind who are doing project based models. Focus and fill the gaps.

 

[macosandlinux]

So did Apple dodge the bullet - being late and or cancelling some of the hyped AI features? I'd say yes, also their PR speak was always ML, not AI/LLM, and that will also help them weather the storm better than all the others.

 

[Nop666]

He's not wrong, but the problem is that all the current hype is about useless LLMs, & when that bubble pops, it's going to poison the business climate for all the actually useful AI projects.

 

[Nop666]

I think that there is some question-begging here. Just because LLMs are improving in their performance does not mean that there is not an economic bubble around them, because even the advanced models haven't yet shown an obvious path to profitability.
But leaving aside the bubble question (which is really subject of the article): The belief that the LLMs will soar to superintelligence and then these superintelligences will gift us with money and other fine things makes some pretty strong assumptions. I might feel better about the notion if the LLMs were trained on data that had itself been generated by superintelligences. It would surprise me very much if, by simply increasing compute and data, we were able to develop a system that could simply design, say, a robot with a working proprioceptive system or a neural prosthesis.
So yeah, maybe I'm just one more normie who will someday be consumed by the basilisk, but I'll believe in the digital sampo when I see it and not before.

Of course Roko's Basilisk is even more imaginary than the prospect of any LLM becoming an AGI.

 

[Wheels Of Confusion]

Generative-AI heavy CEOs this week: "Yeah we're in a bubble and it's about to burst and clean out a lot of companies, but not MY company!"

 

[Nop666]

LLM's in their current form, are in a bubble. Maybe there will be some breakthrough that happens where they stop sprouting utter garbage 69% of time (Yes, that's a made statistic just like LLM's like to hallucinate.) LLM's will only start being useful when they're around 99.9% - 100% correct. Without being able to trust the product and that breakthrough coming soon, I can't see LLM's lasting and bubble will pop.

ML and other parts of 'AI' are much more useful and purpose built, and there lies the problem. It's hard to market something that's a niche product to the masses.

The problem with LLMs, at least in how they are currently built, is that they really are just stupidly expensive & complicated auto-complete machines; they're incredibly flexible in how they can twiddle with words, just like a super-dooper version of the Eliza program from the 1960s, but also like it, they don't actually understand anything in any real way.

 

[Nop666]

The big problem with the current 'AI' bubble is that it's 100% based on it's proponents' (primarily OpenAI) belief, due to ignorance, that throwing enough resources at LLMs will trigger emergent intelligence, despite the obvious* fact that LLMs aren't the kind of tech that could possibly do that. If it were possible, the machines are big enough now that it would've happened, but it hasn't.

* WEll, obvious to anyone familiar with the history of AI research, at least.