The Gist: AI, The Sound and Fury

This is a much more tentative analysis then you usually see from me. That's because I am wandering around inside a landscape where I am very much a stranger in a strange land- thinking about AI. I wanted to write it mostly because it would help me sort out my own ideas about this technology. And I thought I'd post it to you, most esteemed Gist reader, because my tentative conclusions seem to be different from most of those I've seen discussed.

Let's try to start with some context.

Accidental Empires

OpenAI, a fairly obscure company in the global economy, launched a project called ChatGPT in November 2022. It was a downloadable app that connected back to the central OpenAI servers. I call it a project and not a product because it was derived from an earlier bit of research work on creating a language model that OpenAI had developed called GPT-3. If you were launching this thing now as a mass market product, you would not stick yourself with calling it ChatGPT.

Because that is a terrible name.

The thing was, at the time, they weren't expecting ChatGPT to be a mass market product. This was put out in the same spirit of science projects as their earlier research. But then it turned out they had a hit. Five days after the release of ChatGPT, it had 1 million users. In fact, it was so successful that it began to bend the entire technology industry around it. When you become the fastest adopted consumer product in tech history, you attract attention from the money people.

A Box of Tricks

I suppose we should point out here that the AI industry has been developed through hugely extractive and abusive behaviours. Everything written, drawn or pictured on the open internet- all these words here that you're reading, for example- has been scraped, copied and used to train the models so they can guess what is the most statistically likely next word to be produced. Here's The Verge reporting what Mark Zuckerberg said about the commercial worth of the works of Warhol, the writings of Zadie Smith and even these Gists;

“I think individual creators or publishers tend to overestimate the value of their specific content in the grand scheme of this,” Zuckerberg said in the interview, which coincides with Meta’s annual Connect event. “My guess is that there are going to be certain partnerships that get made when content is really important and valuable.” But if creators are concerned or object, “when push comes to shove, if they demanded that we don’t use their content, then we just wouldn’t use their content. It’s not like that’s going to change the outcome of this stuff that much.”

This is a piece of verbal slight of hand, of course, as is almost everything said by the compulsively mendacious husks of humanity granted billionaire status by the quirks of tech history.

If the principle is that you cannot train your AI model without taking and using without consent or payment, um, everything ever made then it suggests that either any AI model is inherently illegal or any income you make is barely yours. And what Zuckerberg does here is try to atomise that problem, which stands at the root of this industry as a commercial venture. He seeks to ignore or minimise collective responses, be they through laws, Treaties or mass action litigation, and isolate an individual creator.

What contribution to the entire corpus of human writing have you made with this one article, or artwork, or work of imagination you're complaining about? "In the grand scheme of this", 'this' being Zuckerberg's description of human imaginative endeavour since history began, sure, hardly anyone has made a definitive difference. But every voice added to that chorus is a gift and adds value to the whole. And, crucially, it is that entire chorus which the AI training process needs to function.

Is it any wonder that this industry is so hostile to the concept of unions? Their model of collective action to assert rights is a threat to its very existance, which relies on individual profiling and targeting of content at atomised users. Anything which sees individuals coming together and acting in unison, whether in their workforces or in the wider public through democratic representation, is inimical to its ability to maintain the power imbalance which has made them so much money.

Making money in the imagined AI future of the tech oligarchs is inimical to democracy. But is making money even possible?

Mad Money

Here is where we point out that the economics of AI are flatly insane. None of this makes sense as an investment. It is pure casino gambling. Every person using ChatGPT costs OpenAI more money then they pay them. Every free and paying user asking how to make a fluffy omelette or getting it to draft an email to their kid's teacher drains their resources. Now multiply that by all the money invested in all of the competing and ancillary AI companies- Google's Gemini, Meta's Llama, Anthropic's Claude and all the companies whose name you've never heard of building and filling data centres with Nvidia chips. Everything spent on this entire gold rush has been to mine gold that costs its owner more than they can charge for it.

Sure, say the investors, they're losing money every time somebody uses their service. But, as the old dot com bubble joke goes, they'll make it up on volume.

The investors' gamble is that everything will settle down later into a few big winners who will eventually control and, in a real sense, own most of the global economy. If AI is everywhere and nothing can work without one of these AI providers then eventually, the thinking goes, they can put up the prices to anything. This is a dream of a rentier economy not seen since French sans-culottes systematically guillotined the heads of the last one.

Imagine the Apple App store, where they take 30% of every transaction, but apply it to the entire global economy.

That's the excuse for the vast investment. Not just the hope of living in the courts of a handful of new Sun Kings. But the fear of missing out on, not the Next Big Thing, but the Last Big Thing. If you are a pile of money, the AI bubble is an event of existential FOMO.

The extent of this money bonfire is only barely perceptible from where we are, because unlike the finance boom, it is based on capital expenditure building actual physical assets like data centres. And in the main those are being built in the US. The effect of this centralisation of investment is that the world's second largest economy is being distorted. (paywall, but you can read Wired magazines for free through the Libby app, if you use your local library card)

But we should just say again here that none of this makes any sense whatsoever. Firstly, because of the existence of China. But secondly, because it is based on a presumption that the AI industry will end up like the modern internet advertising industry, with two or three winners making impossible amounts of money from globally-centralised systems.

And, like Alanis Morrisette before me, I'm here to remind you that event was a complete anomaly in the history of how tech has worked for the last 50 years.

It happened because social networking and search-based online advertising both entwined at a moment when it wasn't possible to make them occur locally in a compelling way. The Intel Inside chips in people's home computers weren't up to it. The copper wires that connected homes to the internet weren't up to it. And the software to make it a good experience wasn't up to it for anyone normal. (Of these three, it is the last issue which has proved most stubborn).

The actual run of technology remains driven by Moore's Law and its variants. Chips get smaller, cheaper and more powerful over time. What used to need a mainframe to process now lives in your pocket, trying to show you videos of cats on Roombas.

The Centre Cannot Hold

"The Eighth Congress of the Russian Communist Party decides that the existence of a unitary centralized Communist Party with a unitary Central Committee directing all the work of the party in all parts of the RSFSR is essential."
- Eighth Congress of the Russian Communist Party, On the Organizational Question. March 23, 1919

Moore's law makes decentralisation of purely processing-intensive activities simply a matter of time. All those billions spent on data centres and all that water and power being used to train models are to produce a non-recurring benefit. Training a model needs vast, boggling, amounts of power, only delivered at an advanced speed by the bleeding edge of the relevant type of chips (now made by the world's most valuable company by share price, Nvidia). But once it's done, it's done.

The second stage in using an AI model is called inference. This is the bit where the model that has already been trained responds to queries or new info you feed it. And this bit is far less demanding. You can use last-generation chips in your data centres. Or a large firm or organisation could even buy a set of boxes to run it in-office. Or, if you don't need that much complexity in your responses, you can run a pre-trained model on your own laptop and chat away to it.

Cory Doctorow, the tech writer and activist, has even pointed to the possibility of tiny chips that know just enough to read complex manuals and documents acting as helpful assistants to trouble-shoot everyday appliances. He's termed this the Productive Residue of the AI Bubble. All that training is done. We can't recover the assets used for it. We may as well see what we can do with the models after their business models collapse, in the same way as we continue to use the fibre optic cables that were laid at lunatic cost during the dotcom bubble of the 2000s.

There is no natural moat protecting centralised AI systems, once training models are being commoditised after they are produced.

Imagine a world where samurai swords are created by master metal smiths, honing their skills for years, forging them through tears over months of hard spiritual and physical labour, finally holding the folded and hammered steel to the dawn's light to capture the essence of the ineffable in the blade.

Now imagine this same world has the cardboard box with Duplicator written on it in crude marker from Calvin and Hobbes. As soon as the completed sword exists, you can just toss it in and make as many identical copies as you like.

That's the AI industry.

Social Media had network effects to keep the balance in favour of centralisation. Advertising and search had the Google Flywheel, where searches informed ads making them much better than anyone else could match.

AI has nothing to stop it from becoming just another feature that runs on your computer or phone except whatever the time will be between someone finding lasting value for it and however long it takes for Moore's law to let that use live on your own device.

Everyone has been laughing at Apple's complete inability to make Siri better, because they have been nowhere in the AI race. But, in the final shakeout, they will still be in a position to incorporate already trained AI models into their devices as they grow more powerful but won't have spent billions on creating them.

We who remain

The strangest thing about this AI bubble is how it manages to throw such a harsh light on the existing, sometimes hidden, fault lines in society. In its push for more concentration of wealth by rent-extracting, it just reveals how much our economy has already been allowed to run on that model.

In its hostility to recognising people's rights to collective action through legislation or banding together to insist on their rights it reveals the hostility to democracy that has been an undercurrent to the tech industry's growth for years.

Let's leave the last word to one of Mark Zuckerberg's individual, overvalued content creators. Taken as a whole, the AI bubble is;

"but a walking shadow, a poor player,
That struts and frets his hour upon the stage,
And then is heard no more. It is a tale
Told by an idiot, full of sound and fury,
Signifying nothing."
- William Shakespeare, Macbeth, Act V, Scene V