When the Bough Breaks
Baby and All
Unless you’ve been living under an extremely dense rock for the past couple of years, you’ve probably noticed the hype around Artificial Intelligence and how it’s going to revolutionize everything from your coffee maker to the very meaning of what it is to be Human. And, honestly, there is some truth to the claims, though this revolution may not be exactly as advertised.
Assuming, however, that you have not been living under such a rock, none of this is news to you. But tech trends can be confusing and complex, and this one is certainly no exception. I don’t know everything but I have been sort of paying attention as a tech professional, hobbyist, and even as a bit of an AI accelerationist myself.
If you’re wondering why everyone is suddenly talking about AI even more than usual this week, I think I may be able to help untie the knot.
The Story So Far
In late 2022, a then little-known Silicon Valley tech firm released a product called ChatGPT, and it really was impressive. Essentially, ChatGPT — and any one of the thousand or so systems that do basically the same thing — was (and is) a chat bot that is very easy to confuse with an all-knowing oracle: you ask it a question, it answers. You can have conversations with it about anything (as long as it isn’t too “controversial”, which is defined by whoever happens to be calling the shots), and these discussions can go pretty deep. I like to use it, honestly. I don’t really do anything serious with it, but it’s fun to bounce ideas off of it or brainstorm or whatever. It’s like an interactive notebook. Fun.
Anyway, people were astonished by this interactive notebook and immediately started imagining all the things they could do with it. Mostly that meant convenient tasks like stripping ads and bloat out of recipe pages, boiling down large blocks of text to a few key bullet points, tricking it into giving you instructions to make meth, making up D&D campaigns, that sort of thing. And, of course, a lot of tech bros convinced themselves the AI was sentient and going to turn us all into paperclips, but that’s another article.
ChatGPT obliterated the charts in terms of user engagement. It was the fastest web app ever to reach 100 million users, for example. For the first few months, OpenAI had trouble maintaining and expanding their servers just to keep up with demand. This frenzy raised a lot of eyebrows in bigger companies, like Microsoft, as usually, this kind of popularity is an indication that something truly momentous is happening. So in poured billions of dollars, with big players making deals with other big players, smaller players getting snatched up by the likes of Apple, and so on.
The Tech industry, which by 2022 was really feeling a drought in new breakthroughs after a brief surge during the pandemic, was absolutely ravenous for this new technology. Ever since then, AI has been the buzzword of every new product, feature, white paper, and tech convention. The hype built to a fever pitch — AI was going to do your laundry, AI can recognize mushrooms, AI can do the job of 15 developers, AI can tell you when your coffee is just right. AI was slapped into every gizmo, gadget, and appliance in sight. And every quarter, it just got bigger, and better, and faster, and cheaper, and more ubiquitous. Artificial General Intelligence - the sci-fi moment where machines “wake up” and perform just about any task as well as or better than any human - was months away.
The hype was really causing a whole lot of activity. And hype really only has one downside: it’s hype. It isn’t actually real. Hype is a lot of promises about what could, might, or should happen. It isn’t what is happening. So what was happening?
This generation of AI tools is broadly known as Generative AI, or “GenAI”. In extremely oversimplified terms, the way it works is by collecting and organizing metric fucktons of data, tagging it with the help of human programmers and operators, and “learning” from patterns in that data. Then, you pipe in some other data, like a sentence or an image or something, and it generates output that looks like it would mesh well with what you gave it, if that had been part of the training data.
There’s a fly in the ointment, isn’t there.
The good thing about this approach is it works wonderfully. The bad thing, of course, is that loads of things that aren’t true could be true, in some alternate timeline where that was what we put in the training data. The tools create content that is plausible, not content that is correct. These AI systems have absolutely no barometer for truth other than statistical likelihood. And if you’ve ever worked with statistics, you know that way lies madness. They’re also insanely easy to trick into bypassing safety protocols and content guidelines, which I guess is important for weirdos who prefer not to have a giant brain telling every mad scientist how to recreate smallpox in a lab or something.
At the same time, pressure has been mounting to make these things actually do something that tech companies could charge for. So far, most of the big AI platforms’ products have been impressive but expensive. And for average users, they’re little more than novelties.
2024 was the year of changing all that — it was the year of AI breaking into the mainstream. Microsoft, Apple, Google, Amazon, and all the other big firms decided it was time to prove that AI was valuable to users. And the way they did that was to .. uh.. put AI in every platform and charge people for it, sometimes sneakily. Okay, almost always sneakily. This generally had the effect of making people downgrade subscriptions or just ignore it,.
But they also began to refocus their long-term plans around AI. Enormous data center projects were begun to build even better models, because surely this problem where AI has no concept of truth can be solved by just making it bigger, right? Nvidia repositioned itself from a company that makes graphics chips for PC gamers to one that makes enterprise-grade graphics chips for tech stock gamers. Apple redesigned their entire product line so people could summarize email and talk to Siri more gracefully.
Increasingly, AI has become not just the next big thing, but the centerpiece of the American, and therefore global, technology industry. And all of that hinges on AI making good on its promises.
And what were those promises, again?
Better coffee! Faster scripting! Right? No. Actually all the stuff I’ve mentioned so far about AI being there to make our lives simpler, faster, and more convenient was smoke and mirrors. That was all a lot of marketing to get us — workers — to stand around twiddling our doodads. The real promise — often unspoken, but more often couched in plausibly deniable HR lingo — was never made to us. The real promise AI is meant to fulfill is much grander and much more dystopian.
With the advent of Artificial Intelligence, Big Tech has been selling the idea that someday soon, business owners would be able to eliminate human labor from their equations. No more strikes. No more labor negotiations. No more sick leave, paid holidays, health care, family emergencies, or messy human beings gumming up the machines. They could simply have an idea, feed it into a magic computer, and boom! Production.
Tech has been promising the people that they can be more productive, while promising our employers they could simply stop paying the payroll. Sure, it was going to be expensive, especially up front. But it was a sure thing, and it was coming soon, and it would fundamentally alter the economic calculus of human society forever. Finally, a world with none of those useless eaters. Just mai-tais and yachts and private jets until the end of time, while the swarming masses of poor displaced workers and their snivelling families just rotted in a gutter somewhere next to what’s left of the ice caps and rain forests.
Whenever anyone pointed this out, companies like OpenAI just did some hand-waving about Universal Basic Income and said “don’t worry, we’re a non-profit and we’re working for humanity”. But then again, they transformed into a for-profit as soon as they thought AGI was inevitable, and I haven’t heard too much about this UBI thing for quite a while. Have you?
Enter the Dragon
So things looked dire. And don’t get me wrong — things still look pretty grim. But just this past week, something unexpected (for Big Tech, anyway) happened. A Chinese cryptocurrency farm that moonlights as an AI company called DeepSeek released a new AI model called R1. And it has sent the Big Tech bros into an absolute nose dive.
Why? Because R1 meets or exceeds the performance benchmarks of o1, OpenAI’s most advanced model? Well, yes, but not only that. Just another competitor, OpenAI could deal with. There are already plenty of more or less equivalent models competing with o1. The problem with R1 is:
It’s free. Not just as in beer, but as in speech. R1’s training weights are open-source, which means anyone who wants it can simply download it. They can continue training it, or fine-tune it, or whatever they want. For zero dollars.
It’s portable. You don’t need a massive data center to run R1. It comes in a few “sizes”, but even the fully functional version can run on commodity hardware in your house. Running it would take a pretty expensive computer, but not an unattainable one. You don’t need Nvidia’s fancy $10,000 GPUs or OpenAI’s data centers or billions of dollars of advanced networking gear and cooling infrastructure.
It’s offline. That is - you don’t have to even connect to the internet to use R1. You can run it entirely sandboxed in an offline environment. No subscriptions, no accounts, no data sharing, no big-tech analytics, nothing. Of course, you can also use the version hosted by DeepSeek so you don’t actually have to buy any hardware at all. Also for free.
Did I mention it’s free?
Ultimately, the conundrum for Big Tech boils down to this: they sunk hundreds of billions of dollars into this tech that still hasn’t even proved its own worth, and now there’s no incentive for anyone to buy it. Even if we wanted this tech (which most of us don’t anyway), and even if we want enterprise-grade physical reliability and data security, it’s infinitely cheaper to just use R1 and build and run it ourselves than to even consider using any option from Google or OpenAI or Anthropic that is anywhere near expensive enough to offset their costs.
So… that’s it? It’s so over?
Nothing is ever really over. I mean, hell, did you know they’re still making cassette tapes? There are still newer models on the horizon from OpenAI and their ilk. But it’s getting increasingly difficult to justify the costs associated with those when China can easily waltz in and deliver something just as good months later.
It also demonstrates that the main strategy Big Tech has had to avoid this outcome — namely, maintaining a “moat” that sets their own products apart sufficiently from competition, especially foreign competition — is simply not working. If China can match o1 in a few months for a fraction of the cost anyway, we might as well just wait for their next model instead. After all, even OpenAI’s flagship models still can’t tell the truth, by definition. As long as all we’re doing is puttering around with corner cases and novelty features anyway, we’re not really missing out on anything.
And this is the crux: the AI boom has been fueled mostly by FOMO from the beginning. If we remove the desperation to get ahead simply due to the impossibility of staying ahead, then there’s no boom.
As for eliminating human labor as a meaningful economic metric any time in the near future, well, ask how much of a threat AI is to the people currently building those massive data centers. They might lose their jobs because AI failed, not because it succeeded.