Login
You're viewing the functional.cafe public feed.
  • Gabriele Sveltogabrielesvelto@mas.to
    Nov 14, 2026, 3:53 PM

    Training LLMs requires a gigantic amount of hardware, brute force calculations, storage, memory and power. Enormous amounts of data whose scraping is not being paid by the companies involved, it's a cost they've spread over the entire internet. As models get larger training gets more expensive. But even as companies sink more money on training LLMs have already reached a plateau in term of functionality and in some cases are degrading in a process called model collapse. 10/

    💬 3🔄 45⭐ 267

Replies

  • Gabriele Sveltogabrielesvelto@mas.to
    Nov 14, 2026, 3:53 PM

    Let's talk statistics for a minute to understand model collapse. Say you have a bunch of measurements of something and you want to build a mathematical function that will tell you what values appear in the measurements and how often they appear, this is what is called a distribution. A distribution can be used to reproduce the material it was created from, but the output will not be the same because something was lost in the process of making it. The less frequent samples, the stuff that lives at the fringes of your data set, the details. 11/

    💬 1🔄 24⭐ 197
  • Gabriele Sveltogabrielesvelto@mas.to
    Nov 14, 2026, 3:54 PM

    Now let's say you use your distribution to generate a new bunch of data: it will look a lot like the inputs you used to make it, but it won't be the same. And if you try to make a new distribution out of that data it will be distorted. You'll lose more information and for every iteration the outputs will look different, they will be more uniform, they'll contain less details, like when you recompress a JPEG file over and over and over. 12/

    💬 5🔄 21⭐ 207
  • Gabriele Sveltogabrielesvelto@mas.to
    Nov 14, 2026, 3:54 PM

    And this is exactly what happens when you train an LLM on its outputs, because an LLM is at its core a distribution of human language. It's also the reason why LLM vendors started putting invisible watermarks in LLM outputs. They need to filter out the slop they make from their training set. But even then some slop will make its way in, and even the slop that gets filtered out is still an additional cost: you scraped it, you processed it to check for your watermark, and then you threw it away. 13/

    💬 5🔄 36⭐ 248
  • Gabriele Sveltogabrielesvelto@mas.to
    Nov 14, 2026, 3:54 PM

    As the internet gets sloppified the chances of keeping slop out diminish, and the training costs increase. And don't forget that all this training has been done without consent, attribution or compensation of the people making the training materials. This space will be regulated eventually, and when it is that will also make the training costs increase further. 14/

    💬 5🔄 38⭐ 236
  • Gabriele Sveltogabrielesvelto@mas.to
    Nov 14, 2026, 3:55 PM

    Which leads us to the hardware itself: this is the point where people will tell you that right now things are looking bad, but the prices will come down because the hardware will improve, and so you'll be able to make powerful local models and whatnot. But here's the thing: hardware improvements are over, in fact they've been for a while. Moore's Law, Dennard scaling, every single factor which lead to faster and cheaper hardware has ceased, years ago. 15/

    💬 5🔄 32⭐ 216
  • Gabriele Sveltogabrielesvelto@mas.to
    Nov 14, 2026, 3:55 PM

    There's no magic hardware improvement around the corner, quite the contrary. Every new silicon node comes with smaller and smaller increases in density and performance over the previous one, but at significant increase in costs. A fabrication facility to produce state-of-the-art hardware takes months to go from a silicon wafer to a working chip, using machines that cost billions. 16/

    💬 4🔄 25⭐ 188
  • Gabriele Sveltogabrielesvelto@mas.to
    Nov 14, 2026, 3:55 PM

    In other words the only way for a modern high-performance silicon process to be profitable is to produce a large amount of chips that will be sold at high price. See how we're coming back to the artificial scarcity where this all started? To keep the semiconductor market going we need more demand and AI represented the largest demand of silicon ever, but none of that demand was real. So when the prices crash the demand evaporates and the foundries will be left with a lot of idling capacity. 17/

    💬 3🔄 36⭐ 176
  • Gabriele Sveltogabrielesvelto@mas.to
    Nov 14, 2026, 3:55 PM

    This will cause a huge amount of capital that was poured into the tech sector to shift elsewhere and this is a good thing. Climate change is weighing down on us and on our future as a species, all the resources we have should be concentrated in fighting it, not chasing a technological fad inflated by a bunch of scammers and mega-rich buffoons having mid-life crises on commercial social media. 18/

    💬 7🔄 123⭐ 376
  • Gabriele Sveltogabrielesvelto@mas.to
    Nov 14, 2026, 3:56 PM

    I'll wrap this up with an appeal to the people who have been using LLMs extensively in their work, often becoming dependent on them: this is the time to come off this train before it's too late, though it might already be. Nobody is making any money with these things, but they made plenty of debts, so you'll be left holding the bag for the entire tech sector since you're the only ones paying for them. Over and out. 19/

    💬 34🔄 216⭐ 523
  • Nov 14, 2026, 5:40 PM

    @gabrielesvelto Thank you so much for this explaining the financial aspect in a way I actually understand. I read Ed Zitron but really struggle to follow all the jargon. Much appreciated!

    💬 1🔄 2⭐ 57
  • 💬 0🔄 1⭐ 34
  • 💬 0🔄 0⭐ 15
  • 💬 3🔄 1⭐ 0
  • 💬 0🔄 0⭐ 4
  • Nov 15, 2026, 11:16 AM

    @stephen@social.calmabiding.me

    why should anybody create places for people who willingly participated in a scam to gain financial support?

    if you were in a position to give financial support either to a person who'd had their IP stolen and used illegally to enrich a scammer or to the scammer for "wanting to stop scamming", to whom would you give that support?

    💬 0🔄 0⭐ 1
  • Gabriele Sveltogabrielesvelto@mas.to
    Nov 15, 2026, 1:25 PM

    @stephen it is still possible to work in the sector without using LLMs, but more importantly if you end up using them it's important to have a plan B for when they won't be available anymore, because that day is right behind the corner

    💬 0🔄 1⭐ 0
  • Nov 14, 2026, 6:15 PM

    @gabrielesvelto thanks for this thread, I follow this stuff a fair amount but the potential catastrophe of mark to market wasn't something I had considered

    💬 1🔄 1⭐ 26
  • Nov 14, 2026, 10:48 PM

    @ricci @gabrielesvelto #TIL 'Mark to Market accounting'
    Looking in to the Enron scandal, the issue wasn't that they did this practice, but that by accounting for all future profits on their current balance sheet, and those future profits were then fore-closed on tomorrow's balance sheet, the shareholders kept demanding more growth, leading to shenanigans, if I understood it correctly.

    💬 0🔄 1⭐ 11
  • 💬 0🔄 0⭐ 8
  • 💬 0🔄 3⭐ 14
  • Nov 14, 2026, 6:53 PM

    @gabrielesvelto Spot on, Gabriel. Exactly what I've been thinking re: the economics of this whole thing, but you've said it more concisely than I've been able to.
    Any chance of having the whole thread put into a blog post or something similar to make referencing it easier? Either way, thanks again.

    💬 0🔄 0⭐ 20
  • 💬 0🔄 0⭐ 10
  • 💬 1🔄 0⭐ 8
  • 💬 0🔄 0⭐ 5
  • 💬 0🔄 0⭐ 7
  • Nov 14, 2026, 7:55 PM

    @gabrielesvelto Good read, thanks for posting.

    I can't help but think they'll socialise the loses the same way as 2008, except its a whole degree of scale larger.

    Fun...

    💬 2🔄 2⭐ 0
  • 💬 1🔄 0⭐ 10
  • 💬 0🔄 0⭐ 2
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 2
  • Nov 14, 2026, 9:34 PM

    @gabrielesvelto

    I don't want to sound negative, but I honestly doubt they won't find something else to grift on.

    All these capitalists just want to siphon money from the majority for the least effort. Building on renewables and stuff like that leads to consumer independence; it's like expecting a drug dealer to sell healthy things like veggies and bandaids.

    The GPU grift went from gaming, to crypto, to "lockdown scarcity", and now to LLMs.

    💬 2🔄 4⭐ 16
  • Nov 15, 2026, 12:20 AM

    @aphantastic @gabrielesvelto

    It seems to me, there are currently only 2 use-cases left for gigantic data centers, once they're not profitable for LLMs anymore:

    - Mass surveillance
    - War (here I'm blindly guessing, though, as I'm not familiar enough with the topic and the resources needed for drone strikes and monitoring)

    💬 2🔄 1⭐ 4
  • Nov 15, 2026, 12:43 AM

    @schokopflaster @aphantastic @gabrielesvelto To quote Jules from Pulp Fiction: "Check out the big brain on Brad!"

    Of course LLMs are abundantly applicable to sentiment analysis, covering your "mass surveillance" use case, but to a lesser extent, military apps.

    The LLM provides the *illusion* of natural language queries at great computational expense. The underlying DANN provides brute-force directed search of the trained embeddings.

    💬 0🔄 1⭐ 1
  • 💬 0🔄 0⭐ 0
  • Nov 15, 2026, 12:42 AM

    @aphantastic
    Quantum will be the next grift.
    E.g. Meta seem to be already positioning themselves for it. Video, VR, LLMs, Quantum. Anything to keep Meta as a growth stock instead of a mature one. The billions they sink into each are funded by printing money (issuing stock) rather than by direct investment. When it fails to be profitable they just roll onto the next scam.

    💬 1🔄 8⭐ 11
  • Nov 15, 2026, 12:50 AM

    @a_cubed @aphantastic You're reminding me that I need to finish reading Cedric Durand's book "How Silicon Valley Unleashed Techno-feudalism". I got 30% of the way in. The gold nugget is his observation that even where democratized praxis & technology leads to greater profits for *all*, incumbent stakeholders resist this.

    💬 1🔄 2⭐ 3
  • Nov 15, 2026, 1:07 AM

    @bms
    An older work that's essential background for this is Drahos and Braithwaite's Information Feudalism, I think. That's highly readable and shows how a small number of US firms (Disney, MS, Big Pharma) got the US trade office working for them instead of even the whole of the US economy, let alone the world economy.

    💬 0🔄 2⭐ 6
  • Nov 14, 2026, 9:46 PM

    @gabrielesvelto I do disagree with your point in post 19 — I think it is possible to make money with probabilistic pattern modelling or statistical sequence prediction in a variety of technical and other fields, perhaps not at the scale and pricing or with the specific products that the famous names are offering today. The invention is out of the bag, it will continue to be refined and improved.

    💬 1🔄 1⭐ 3
  • Gabriele Sveltogabrielesvelto@mas.to
    Nov 14, 2026, 9:48 PM

    @andyd machine learning can be profitable and will be again in many use-cases, but large language models specifically are not going to be part of it IMHO

    💬 5🔄 2⭐ 23
  • Nov 15, 2026, 12:45 AM

    @gabrielesvelto
    They have a small number of use cases where the technology might be profitable. Mostly things where coming up with plausible and checkable results (e.g. protein folding, antibiotics) based on specific "vocabularies". General natural language models? Inherently uneconomic, I agree.

    💬 1🔄 0⭐ 3
  • David Megginsondavid_megginson@mstdn.ca
    Nov 15, 2026, 2:22 AM

    @a_cubed It's hard to see how companies could gain a competitive advantage from being genAI users, though, since everyone else has access to exactly the same models.

    Even if genAI lived up to its promises (a big "if"), it would be essentially a utility, like electricity or internet, but the shares of AI providers aren't priced like utility companies.

    @gabrielesvelto

    💬 2🔄 1⭐ 1
  • 💬 0🔄 0⭐ 0
  • Nov 15, 2026, 8:07 PM

    @david_megginson @a_cubed @gabrielesvelto Transformer models can do really quite well when the underlying ground truth is smoothly differentiable. This is the case in quite a few scientific domains (because science looks for this).

    But human thought turns out to not be a projection on a smooth differentiable hypersurface. It's got elements that are more complex. (It's also necessarily dynamic, but that's a different problem.)

    💬 0🔄 0⭐ 0
  • Nov 15, 2026, 12:51 AM

    @gabrielesvelto @andyd Your point recapitulates Grady Booch's direct rebuttal of Connor Leahy and Dario Amodei on the value of their LLM driven tulipmania. Booch was careful to point out established ML systems, pre-DL/DANN/LLM, do have value and continue to do so.

    💬 1🔄 0⭐ 7
  • Nov 15, 2026, 1:25 AM

    @bms @gabrielesvelto @andyd yep (non-LLM, non-GPT/NN) ML is useful for spam identification, for example, but it relies on much simpler and less sexy techniques like SQL data mining, regexp pattern matching and Bayesian inference graphs like with Graham/Viaweb's "if email subject or body has Viagara, 99% chance is spam" rule. none of Nvidia's hardware needed for that. Google itself relies on unpaid crowdsourced data labelling by its own userbase to "train" its Bayesian inference classifiers still, AFAIK

    💬 1🔄 2⭐ 1
  • 💬 0🔄 0⭐ 3
  • Nov 15, 2026, 1:42 AM

    @gabrielesvelto @andyd I think there will still be some cases where LLMs get used even after "subsidies" are gone. Anyone who is already finding value from open-weight models on their own computers are going to keep doing it, and anyone who is willing to throw a *lot* of money at their problems will be fine with expensive LLM services.

    💬 1🔄 0⭐ 3
  • Nov 15, 2026, 7:24 AM

    @varx

    Open-weight models are trained by the big commercial ones (Claude etc). If the bubble pops, it'll take open-weight models down with it.

    Of course, you'll be able to keep using your open-weight models after that, but you won't be getting any more updates.

    @gabrielesvelto @andyd

    💬 1🔄 0⭐ 3
  • 💬 1🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • Nov 15, 2026, 8:36 AM

    @gabrielesvelto Necessity is the mother of invention and all that. The small language model and domain specific model are likely to be interesting, accidental, superior technologies. At least for things I’d like to use them for!

    💬 0🔄 0⭐ 3
  • Nov 16, 2026, 1:39 AM

    @gabrielesvelto @andyd I'm curious about this take. I've heard from multiple cases, both online and people I work with, that off-the-shelf LLMs and even lower end ones perform very well on prediction or classification tasks and often better than a custom ML model trained by a data scientist with access to all the potential data the company owns. I don't doubt there are cases where training a custom model isn't advantageous. But an LLM doing well ootb will often be the more viable option, especially for small volume or quickly iterated environments. A small to mid model that exists today won't disappear.
    Am I missing something?

    💬 1🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 3
  • 💬 1🔄 1⭐ 1
  • 💬 0🔄 0⭐ 0
  • Nov 14, 2026, 10:16 PM

    This is why it’s so important to retain authorship over the things you produce. If you’re using LLM’s to vibecode or otherwise replace the actual work of understanding what you’re putting your name on. You’re reliant on a system that’s precarious at best.

    I do wonder what the fallout of this will be in terms of existing / trained models. I’m sure these companies will want to milk them for as much value as they can. With slower progress, what will the cost of providing existing models be?

    💬 0🔄 1⭐ 0
  • Nov 14, 2026, 11:04 PM

    @gabrielesvelto
    I, for one, would bet my bippy that the tech bros, hedge funders, data center entrepreneurs, and Ai evangelists are now spending as much time working on ways to unload their debt onto the people who WEREN'T responsible for the Ai bubble as they used to spend creating the Ai bubble. Look out Mr and Mrs taxpayer, the "smartest people in the room" are coming for your wallets and nest eggs.
    #Ai #FinancialBubble #DataCenters

    💬 1🔄 9⭐ 12
  • 💬 0🔄 1⭐ 5