Post

Any half decent system (including phone CPUs) with 64Gb of RAM can do 98% of the actually useful things that LLMs do. Why do you think they’ve bent over backwards to sign bullshit contracts that pretend to tie up the world’s RAM supply?

This is the one conspiracy theory I’ll subscribe to. AI companies and companies deeply dependent on AI (like NVIDIA) recognised that their USP was months away from being entirely doable on local devices, and acted to fuck the global supply chain.

Almost all of this - as @edzitron.com has documented extensively - is via letters of intent, which pretend to move around billions of dollars while having almost zero legal meaning. The “purchases” of the world’s RAM won’t happen. But they’ll fuck the pricing of consumer devices for years.

I’ve spent 30 years writing about the games biz, which has fended off countless waves of hucksters and bullshit artists, but none quite as awful as the AI bros selling a future where we’ll all consume slop through a hose with no human creativity on the other end. But like every other bullshit trend…

bullshit trend, it’s ultimately going to enrich a handful of early movers who convinced gullible idiots in boardrooms to buy in. I’m not an LLM refusenik. I use it every day; it’s a great new technical advancement. If you laid off a single worker thinking AI can replace them, you’re a fuckwit.

These models will fundamentally change how we interact with computers. That’s wonderful! But a new paradigm in computer interfaces isn’t a machine god, let alone a global economic upset. And everything OpenAI claims to do will ultimately run on your wristwatch.

Seriously; people don’t get how fucking insane it is that these people are justifying a trillion dollars in datacentre spend when…. I can run a pretty decent Claude Code alternative on my MacBook Pro M2 Pro from three years ago.

Is Claude Code better? Yes! Is it a trillion fucking dollars better? No. And that’s “no” as the CTO of an admittedly small but quite cutting edge company where we have to develop new algorithm stuff and wrap it in cloud infrastructure every day, which is meant to be the best case for these systems.