Need to let loose a primal scream without collecting footnotes first? Have a sneer percolating in your system but not enough time/energy to make a whole post about it? Go forth and be mid: Welcome to the Stubsack, your first port of call for learning fresh Awful you’ll near-instantly regret.
Any awful.systems sub may be subsneered in this subthread, techtakes or no.
If your sneer seems higher quality than you thought, feel free to cut’n’paste it into its own post — there’s no quota for posting and the bar really isn’t that high.
The post Xitter web has spawned soo many “esoteric” right wing freaks, but there’s no appropriate sneer-space for them. I’m talking redscare-ish, reality challenged “culture critics” who write about everything but understand nothing. I’m talking about reply-guys who make the same 6 tweets about the same 3 subjects. They’re inescapable at this point, yet I don’t see them mocked (as much as they should be)
Like, there was one dude a while back who insisted that women couldn’t be surgeons because they didn’t believe in the moon or in stars? I think each and every one of these guys is uniquely fucked up and if I can’t escape them, I would love to sneer at them.
(Credit and/or blame to David Gerard for starting this.)
I have been thinking about the true cost of running LLMs (of course, Ed Zitron and others have written about this a lot).
We take it for granted that large parts of the internet are available for free. Sure, a lot of it is plastered with ads, and paywalls are becoming increasingly common, but thanks to economies of scale (and a level of intrinsic motivation/altruism/idealism/vanity), it still used to be viable to provide information online without charging users for every bit of it. Same appears to be true for the tools to discover said information (search engines).
Compare this to the estimated true cost of running AI chatbots, which (according to the numbers I'm familiar with) may be tens or even hundreds of dollars a month for each user. For this price, users would get unreliable slop, and this slop could only be produced from the (mostly free) information that is already available online while disincentivizing creators from producing more of it (because search engine driven traffic is dying down).
I think the math is really abysmal here, and it may take some time to realize how bad it really is. We are used to big numbers from tech companies, but we rarely break them down to individual users.
Somehow reminds me of the astronomical cost of each bitcoin transaction (especially compared to the tiny cost of processing a single payment through established payment systems).
The big shift in per-action cost is what always seems to be missing from the conversation. Like, in a lot of my experience the per-request cost is basically negligible compared to the overhead of running the service in general. With LLMs not only do we see massive increases in overhead costs due to the training process necessary to build a usable model, each request that gets sent has a higher cost. This changes the scaling logic in ways that don't appear to be getting priced in or planned for in discussions of the glorious AI technocapital future
This is a very important point, I believe. I find it particularly ironic that the "traditional" Internet was fairly efficient in particular because many people were shown more or less the same content, and this fact also made it easier to carry out a certain degree of quality assurance. Now with chatbots, all this is being thrown overboard and extreme inefficiencies are being created, and apparently, the AI hypemongers are largely ignoring that.
I've done some of the numbers here, but don't stand by them enough to share. I do estimate that products like Cursor or Claude are being sold at roughly an 80-90% discount compared to what's sustainable, which is roughly in line with what Zitron has been saying, but it's not precise enough for serious predictions.
Your last paragraph makes me think. We often idealize blockchains with VMs, e.g. Ethereum, as a global distributed computer, if the computer were an old Raspberry Pi. But it is Byzantine distributed; the (IMO excessive) cost goes towards establishing a useful property. If I pick another old computer with a useful property, like a radiation-hardened chipset comparable to a Gamecube or G3 Mac, then we have a spectrum of computers to think about. One end of the spectrum is fast, one end is cheap, one end is Byzantine, one end is rad-hardened, etc. Even GPUs are part of this; they're not that fast, but can act in parallel over very wide data. In remarkably stark contrast, the cost of Transformers on GPUs doesn't actually go towards any useful property! Anything Transformers can do, a cheaper more specialized algorithm could have also done.