Technology

Sorry, you're correct, that was a typo, I'll edit.

The point of comparison is still correct, about it being equal to the total compute of London. I just wrote the number wrong.

That's part of it, but the bigger part is that the enterprise business has completely failed to manifest.

Transformer model based AI is, at best, a fun toy or a minor convenience. For example, as a coding assistant, it functions as a fairly effective and speedy search engine, so long as you have the skill to check the output. But it also shortcuts the research process for you in a way that makes you more likely to overlook any potential downsides in the solution it offers, or perhaps miss a better solution you might have found if you did the research yourself. It's handy for stuff that's low stakes, but not reliable enough for anything that really matters.

And as a replacement for a coder, it just sucks, producing bug riddled, insecure code that it lacks the ability to debug effectively, meaning you still have to employ coders to fix its output.

This is the story with every potential entreprise application; it can't be trusted enough to replace the expensive humans it's supposed to replace, and it doesn't actually make the expensive humans substantially more productive when it assists them (this has been studied; while a lot of people will anecdotally claim AI makes their job easier, in practice the numbers show that it mostly either slows people down or makes no difference).

As a customer service agent, it's a dangerous liability, with a habit of outright hallucinating answers, and vulnerable to prompt injections that could allow for all sorts of dangerous shenanigans if you give it the ability to actually start making decisions about refunds and rebates.

Replacing troublesome, expensive humans was always the real sales pitch. Every other feature these charlatans advertised was just part of a scam to keep us troublesome, expensive humans from complaining too loudly about getting replaced. When Sam Altman tells us that AI is going to cure cancer and solve global warming, he knows it's a lie. But he also knows that "I want to take all your jobs" isn't exactly great PR, so he has to invent reasons why this is actually a good thing.

But the real sales pitch hasn't come through, and it's increasingly apparent that it won't. You don't invest hundreds of billions of dollars into compute to sell a slightly "better" search engine, or an app that shows you a short video of you kissing your crush. Those are not products that are going to give that kind of return on investment.

I'm not saying that transformer model AI will never have any practical uses. But the big practical use, the one that was driving a truly unprecedented wave of tech investment; that isn't happening. AI is not going to come and take all our jobs. Or, at least, this version of it isn't. Microsoft have made the smart play by jumping off the runaway train before it reaches the bridge. They're going to get fucked by this, but not nearly so hard as the people who don't know when to fold.

Even if it was getting slim enough to run on consumer grade hardware it wouldn't undercut the SaaS model.

I think the real thing, as others have mentioned, is that it's not reliable enough and everywhere I've seen it's been difficult to move the latest models past just being chatbots.

If AI was actually going to replace every human worker on earth, even with huge efficiency gains you'd still want to build out absolutely ungodly amounts of compute capacity.

Also, it should be noted that while Deepseek has demonstrated that its possible to substantially reduce the compute requirements for transformer based models, doing so relies heavily on a "Good enough" approach that moves the results further away from being enterprise capable. It's not a cut and dried solution to the backend costs of running these models at the scale that investors want to see them running.