Other What's ruining your life right now?

On the positive side, the bubble is now entering the "VCs would quite like their promised return now, please" phase - which is the beginning of the end, at least in the bubble's current form. Anthropic is trying to take Claude away from the $20-a-month subscribers (which won't help, 'cos the estimates I've seen suggest the company is burning $8-13 in costs for every dollar it brings in), OpenAI is shunting people across to cheaper models whether they like it or not, and Microsoft - friggin' *Microsoft*, which owns desktop computing and runs one of the largest cloud infrastructures in the world - has just closed sign-ups for GitHub Copilot while removing the most expensive models from those who already have it because it's just a massive money pit with zero path to profitability. I'm also seeing rumblings about the rising cost of all this leading to "AI layoffs," where companies are cutting their AI spend in favour of hiring actual (cheap) humans again.

Given what studies have shown about the scale and speed of mental atrophy experienced by the users of generative AI, we're going to be in for Interesting Times when people get priced out of using the things. There's going to be a strong jobs market for people who've sensibly steered clear, I reckon - though potentially only amongst the rubble of the companies who put all of their money into AI and lost the lot come the bubble's true burst.

Oh, which reminds me of one more point: studies continue to show marked mental degradation from the use of generative AI, which is rapid and long-lasting... which could potentially open companies insisting on its use to workplace injury lawsuits. We can but hope!

 

There were also studies about what tobacco was doing to people, and what asbestos was doing to people, and d'you know what? Turns out they were absolutely right.

To say nothing of the environmental considerations: these companies are talking about spinning up datacentres powered by gas turbines which will put out more CO₂e than entire nations. Or the ethical considerations. Or the fact that you're making people like Scam Altman and Pedo Guy rich(er). Or the fact that the output is crap. "It keeps getting better, use the new model!" Oh, the new model that OpenAI just put out, which is twice the per-token price as the one it replaces? That new model? The output of which is still crap?

"I told it to make me an app that does X, and it did!" Yeah, and it's crap (and probably violates about eight different licenses from the projects the code's stolen from). Sure, maybe it works. But you can't build software like that. The technical debt companies - and I'm talking at-scale, here, if you use it to make a ten-line bash script to automate something then you're probably going to dodge all this but I'll still judge you - are building up is staggering. Have you seen the Anthropic "Claude is written by Claude" code leak? It's horrifying. It's a house of cards built on a foundation of investment ice-cream, and it's going to give. Microsoft claims it's moved a lot of its development across to AI (at the same time as putting a "Copilot is for entertainment purposes only" disclaimer on the thing(!)), and GitHub has never been flakier - not to mention Windows.

"It's just like having a junior (or, in these days of 'agentic AI,' a hundred junior) developer(s) working for me!" No, it's not. D'you know what happens when a junior dev makes a mistake? They learn from it, and eventually - surfing on the crest of a wave of a million mistakes - become senior devs. D'you know what happens when Claude makes a mistake? It "apologises," tells you you're amazing for catching it, rewrites half a dozen files to introduce a handful of new problems while potentially but not definitely fixing the one you found, and then Anthropic doubles your subscription fee. What it does not do is learn. Oh, sure, they've got "memories" and "soul files" (ugh), but they're constrained by the context window. Unless your codebase is tiny (in which case why the hell are you using gen-AI) then it ain't going to be able to hold both all the code and all the documentation and have any room left for your prompt, never mind the "memory" of the last time it screwed up the codebase.

There's a thing in right-wing, Christofascist, hard-C Conservative circles: "the only ethical abortion is my abortion." Sloperators are the same: "Yes, there are ethical problems, there are mental problems, there are scaling problems, the people who make it all are universally terrible and definitely don't have my best interests at heart, the way they work means they can only ever deliver statistical mediocrity at best, but I'm different. I'm special. I have broken through the surface and spoken to the ghost in the machine."

No. Just... no.

...I should probably write the stuff I'm actually supposed to write today...

 

Never going to happen, calling it now. The reason the models are getting better is 'cos they're ploughing more memory and compute into training, inference, or both. Look at Byron's thing: the model has to burn though millions of tokens in order to look smarter than its predecessors. They're getting bigger faster than optimisation can improve: the new DeepSeek-V4-Pro is a 1.6 trillion parameter model with a one million token context window. The original DeepSeek-R1 was 671 billion parameters. Sure, you can quantise down to reduce memory requirements, but every time you do so you lose precision (I mean, literally, that's what quantising is: changing from BF32 or FP32 or whatever to INT8, INT4, I think some places are even messing around with INT2) and the model gets that bit less capable.

Think of it like this: I'm typing this on a computer which is orders of magnitude faster than the one I first got on the internet with, back in the mists of time. Yet its day-to-day performance feels near-identical, bar how quickly I can download porn stuff. The software's requirements grew at pretty much the same pace as the hardware, so everything feels as slow as it ever was - bigger, shinier, higher-resolution, but not really any faster. Instead of accepting the improvements to make things faster, developers opted to grow their software's complexity - either adding new features (websites are still slow 'cos they download and execute literally tens of megabytes of JavaScript just to show me the homepage) or just failing to optimise like you had to when 640kB was enough for anyone.

LLMs are the same. Nobody's going to use TurboQuant to make their LLM run as good as it does now in one-sixth the memory, they'll use it to make an LLM six times bigger run in the same amount of memory (which will then be, oooh, I dunno, a generous 10% better than the current one?)

 

I'm not talking about increasing capacity (i.e. you can have six million users instead of one million), I'm talking about making the models bigger so their responses seem better - which is what they're all competing at now. They'll still only be able to support one million users, the costs will be unchanged, but the models will be six times bigger.

 

You mean make it more expensive? Sure, that's what they're already doing - but as Anthropic found out when it tried to move Claude from the $20 a month tier to the $100 a month tier, there's a limit to what most people will pay - and it's a very, very low limit. If you replace ten $20 users with one $100 user, you've just halved your revenue - and if that 'power user" uses the same compute as those lightweight users combined, you've quartered your profit.

But that's separate to my point, which was that there's no future in which any of this gets cheaper for end-users. (Which may have also been your point, too, of course - I think I may have misinterpreted your "ride that wave onto the beach" metaphor!)