I'm on a tour with my new book Enshittification: catch me next in Los Angeles, Calgary and San Francisco! Full schedule here.
When the AI bubble pops, what will remain? Cheap GPUs at firesale prices, skilled applied statisticians looking for work, and open source models that already do impressive things, but will grow far more impressive after being optimized:
The AI bubble companies are scams. They've spend most of a trillion dollars in capital expenditures, and by their own (very cooked and dishonest) numbers, they are grossing a total of $45b/year, industry-wide:
At $45b/year (an inflated number, remember!) it's going to take them a long time to recoup the hundreds of billions of dollars they've spent so far. But they don't have a long time: the massive GPUs that power AI's "foundation models" and cost six- or seven-figures each burn out remarkably quickly. The companies that buy these GPUs claim they'll last five years (and depreciate them over that schedule); however, this is accounting fraud, because in reality, these GPUs have a duty-cycle that's more like two to three years:
To recoup their existing and announced investments, AI companies will have to bring in $2 trillion, more than the combined revenue of Amazon, Google, Microsoft, Apple, Nvidia and Meta:
And they have to bring in that $2 trillion before all those GPUs burn out…which is, again, about 2-3 years.
Or sometimes just 54 days.
AI companies' purchases and R&D expenditures aren't guided by the need to make products that will bring in $2 trillion dollars. AI companies spend money in order to put on a show for investors, to demonstrate that they are very serious about AI. Think of all those GPU-stuffed data-centers as akin to a peacock's tailfeathers: an expensive way to attract mates (or, in this case, investors), by emitting costly signals that demonstrate your power:
https://en.wikipedia.org/wiki/Signalling_theory
Of course, it's far cheaper to pretend to be spending a lot of money than it is to actually spend it, and they're doing plenty of that, too. Meta has promised to spend $72b next year on data-centers. However, Meta's annual free cash flow is $52.1b. OpenAI says it will spend $60b/year on data-centers, which is five times its annual revenue of $12.7b (and the company is losing $9b/year). As The American Prospect's Brian McMahon writes, "How can OpenAI plan to spend five times what it brought in?"
I don't know how many of these giant "foundation models" will still be online after the crash, but I would not be surprised if that number is zero.
So the big question is, what comes next? What will the AI bubble leave behind?
Some bubbles leave nothing or next-to-nothing behind. Enron left nothing behind but the cooling corpse of a CEO who popped his clogs before he could be sentenced to life in prison. Worldcom left behind a CEO who survived long enough to die behind bars…and a ton of fiber in the ground that people are still getting use out of (I'm sending these keystrokes to the internet on old Worldcom fiber that AT&T bought and lit up).
Crypto's not going to leave much behind: a few Rust programmers who've really taken security by design to heart, sure, but mostly it'll be shitty Austrian economics and even shittier JPEGs.
So what kind of bubble is AI? That's the $2 trillion question:
Before I get to that, let me be clear here: bubbles are always bad. As much as I like my 2gb symmetrical fiber, the fact that it exists because a crook stole billions of dollars from everyday people who were only hoping to live a dignified retirement of material sufficiency is terrible. Worldcom CEO Bernie Ebbers deserved what he got, and worse.
The AI bubble is on its way to sucking up a trillion dollars and not all of that money is coming from Saudi royals, hedge fund bastards and Elon Musk's credulous creditors. Plenty of it will come out of the savings of working people who've forced to play the suckers at the table thanks to the replacement of guaranteed pensions with "market-based pensions" that only pay out if you guess right about which stocks to buy:
Those people are going to get wrecked. And so are the rest of us. You don't need to be an AI investor to get wiped out by the AI investment bubble, either. With 30+% of the S&P 500 tied up in seven AI companies' stock, the coming crash will definitely escape containment and crash the whole damned economy.
So the bubble is bad. Really bad. But even so, there will be things we can salvage from it: open source models, skilled programmers, cheap GPUs bought out of bankruptcy for pennies on the dollar. It would be better if we created that stuff without burning the world's economy to the ground and emitting a heptillion tons of CO2, but ignoring the productive residue of the AI crash won't bring the economy back, or suck the carbon out of the atmosphere.
The open source models are a big deal. They're already capable of doing really impressive things, like transcription, image generation, and natural language-based data transformation, running on commodity hardware. I run several models on the laptop I'm typing this on – a computer that doesn't even have a GPU.
What's more, there are a lot of ways to improve these models within easy reach. The US AI companies that threw these models over the transom after irrevocably licensing them as free software had very little impetus to improve their efficiency by optimizing them. Remember, they're spending money as a way to "prove" that AI has a future.
Shipping a model that runs badly – that needs more data-centers and energy to run – is a way to convince investors that it's doing something really advanced (after all, look how much compute and energy it's consuming!). It's a scaled-up version of a scam that Elon Musk used to pull on investors when he was shopping his startup Zip2 around: he put the regular PC his demo ran on inside a gigantic hollow case that he would wheel in on a dolly, announcing that his code ran on a "supercomputer." Yes, investors really are that dumb.
Even modest efforts at optimization can yield incredible performance gains. Deepseek, the legendary Chinese open source AI model, consumes a fraction of the resources gobbled up by the likes of OpenAI. Deepseek's launch was so impressive that it knocked $589b off of Nvidia's stock price the day it shipped:
There are a ton of these open source Chinese models, and they all perform like crazy. China does a lot of AI optimization because US embargoes prevent Chinese AI companies from accessing the most powerful GPUs, so Chinese coders tighten up their code and outperform US companies even though they're using far less powerful computers.
After the crash, everyone will be in a similar position to those Chinese AI optimizers: Chinese companies can't buy advanced GPUs because of the embargo; and everyone else won't be able to buy advanced GPUs because the AI crash will have cratered the economy for a generation.
But there is so much room at the bottom. Optimized models do really impressive things on really cheap hardware.
How cheap? Well, here's hardware hacker Pete Warden demoing a chatbot that you talk to and that talks back to you – and it's running on Synaptics System-on-a-Chip (SoC) that costs "low single digit dollars":
This is basically a little special-purpose Alexa, except it doesn't connect to the internet at all (and therefore doesn't leak any of your data). In Warden's demo, the gadget is a button-sized voice assistant that is meant to be integrated into a dishwasher, which can interpret the dishwasher's manual for you. If your dishes come out dirty or if the drain gets clogged, you press the button, describe your issue in pretty vague terms, and it instantly speaks aloud all the troubleshooting steps to deal with it.
This privacy-preserving, cheap-like-borscht component adds a voice-activated, conversational assistant to a device, sipping power like the clock on your microwave, running on a processor that costs less than a pack of AA batteries. It's seriously fucking cool.
There's going to be a lot of this AI, after the AI goes away – just like there was a lot of the web after the dotcom crash, when, overnight, San Francisco had infinity office-space, servers, and techies going begging.
If you'd like an essay-formatted version of this post to read or share, here's a link to it on pluralistic.net, my surveillance-free, ad-free, tracker-free blog:
Along the same lines as Zero to Lasagne, only with a more humanities bent, TensorFlow for Poets is a guide to getting TensorFlow to run on OS X (with a screencast so you can follow along!) It’s written by Pete Warden, who helped build TensorFlow. Once you’ve got it set up, you might also want to check out the tutorials on the TensorFlow site.
Maybe this will help one of you get set up with experimenting with deep learning and exploring the artistic possibilities. I do think that they’re mostly unexplored, since relatively few people have the combination of skills, time, and artistic bent to explore them. Hopefully tutorials like this will help increase that population.
My first girlfriend was someone I met through a MUD, and I had to fly 7,000 miles to see her in person. I read a paper version of the Jargon File at 15 and it became my bible. J...
The decades have proved that our [nerds'] way was largely right and the critics were wrong, so our habit of not listening has become deeply entrenched. It even became a bit of a bonding ritual to attack critics of the culture because they usually didn’t understand what we were doing beyond a surface level. It didn’t used to matter because nobody except a handful of forum readers would see the rants. The same reflex becomes a massive problem now that nerds wield real power.
La culture geek est devenue mainstream. Les nerds sont devenus plus importants. Ils ont de l’argent, du pouvoir, un statut. Ils sont à la tête des plus grandes et des plus dynamiques entreprises du monde. Et la culture dominante ne se moque plus de nous, mais nous respecte. Travailler dans le jeu vidéo est devenu sexy.
Post by: Pete Warden
When I first released the DeepBeliefSDK for iOS devices, one of the top requests was for an Android version. I’m pleased to say, after some serious technical wrestling, you can now use the image recognition library in your own Android apps! Just download the github repository and run the Android sample code.
