Post #4372540

2026-08-04 04:11 UTC

@shituationist@kolektiva.social lots of people have asked me if I think LLMs will result in any of the promises the LLM companies have made, all the way up to and including self-modifying models (superintelligence), and I keep saying "no" and these people who ask me this, who are typically pretty smart people, but none of them have any actual expertise in ML other than memeing around with ollama or vLLM or maybe tensorflow/transformers for a sophisticated experimenter, so they choose the side of confirmation bias and tell me I'm definitely wrong on this 🙃 if you look closely, you'll notice that OpenAI is already hitting scalability problems keeping ChatGPT up and running, and we know they are having scaling problems because of how the outages are shaped: they come as micro-outages, where inference requests will sporadically fail for short periods of time. this type of availability problem is basically almost always caused by lack of horizontal scaling. and that's because they simply don't have the inference capacity. they buy all the GPUs, all the RAM, its still not enough. and the bill is coming due at the end of this quarter for OpenAI, though apparently nvidia have proposed refinancing the debt, so who knows?

Replies (2)

@n_dimension@infosec.exchange 2026-08-04 10:32

@ariadne@social.treehouse.systems @shituationist@kolektiva.social Whenever someone declares LLMs ded, check which wall they're pointing at, there are three and they keep getting smashed together. Attached plot: the one people wave around as a ceiling. Log-log axes though, that dashed diagonal is a power law holding over ten orders of magnitude. Chinchilla et.all. All it prices is compute against loss. The micro-outage forensics; failed requests, ergo no headroom, ergo scaling is dead, are fun but underdetermined. Deploys, autoscaler lag and load-shedding all look identical to 'ran out of GPUs' from userland. Frozen weights don't get tired: capacity pressure fails loudly, not stupid. The honest symptom is the 502. The dishonest one is quiet , quantize harder, trim context, route to the smaller model, same name on the box. Providers do that.sometime The money, though, sure fair cop. A quarter-trillion vendor backstop should make everyone's eye twitch. HOOOWEVER... (And its not something I wanna argue, still developing this argument)... A quarterly return is a shit metric for a magic genie that will put all the professionals out of work, forever. We all know capitalism sucks, and its funny that brologarchs are discovering it too. Just don't let a shaky business model stand in for a broken scaling law. Power laws have outlived plenty of companies. TLDR; Yes but no.

Open ##4372539
@n_dimension@infosec.exchange 2026-08-04 10:34

@ariadne@social.treehouse.systems @shituationist@kolektiva.social Whenever someone declares #LLMs ded, check which wall they're pointing at, there are three and they keep getting smashed together. Attached plot: the one people wave around as a ceiling. Log-log axes though, that dashed diagonal is a power law holding over ten orders of magnitude. Chinchilla et.all. All it prices is compute against loss. The micro-outage forensics; failed requests, ergo no headroom, ergo scaling is dead, are fun but underdetermined. Deploys, autoscaler lag and load-shedding all look identical to 'ran out of GPUs' from userland. Frozen weights don't get tired: capacity pressure fails loudly, not stupid. The honest symptom is the 502. The dishonest one is quiet , quantize harder, trim context, route to the smaller model, same name on the box. Providers do that.sometime The money, though, sure fair cop. A quarter-trillion vendor backstop should make everyone's eye twitch. HOOOWEVER... (And its not something I wanna argue, still developing this argument)... A quarterly return is a shit metric for a magic genie that will put all the professionals out of work, forever. We all know capitalism sucks, and its funny that brologarchs are discovering it too. Just don't let a shaky business model stand in for a broken scaling law. Power laws have outlived plenty of companies. TLDR; Yes but no. #aiscaling #llmlimits

Open ##4372576