Post

Arguably the fact that we've had highly inefficient, schizophrenia-causing, shit products when we didn't have to is an indictment on the companies that did all that, because it didn't have to be that way.

While you're correct that running frontier models locally isn't exactly viable, they also aren't needed in many cases. Also, the efficiency of these models is rapidly improving with time. So, it may be doable in another year or two.

The missing axis is whether the model clears the workflow’s quality bar: cost curves show break-even volume, but task-level quality—and whether the workload runs batch or sync—can change the API vs self-hosting answer.

and if the actually useful applications (spoiler alert: not the LLMs) got more of the attention and funding. nobody was rallying in town halls for machine learning used in medical research, and that’s probably because the method and purpose wasn’t stripping the sum of humanity’s knowledge for profit

I mean... there has never been a clearer case for distributing the Means of Production Just give me the chips I'm already paying for with subscriptions and tax subsidies. Dismantle the data center and hand me a slice That's where we're headed. ENIAC is a concept. The personal computer a revolution

Absolutely. I sure miss the days when we could just talk about machine learning and computer vision (I do stuff with drone imagery)

I've been able to maintain a previously unmaintainable codebase and respond to the people constantly breaking shit, because my company provides us a capable LLM 😅 (I assure you, they were breaking shit and outputting slop before the AI era)

I get the sense you either haven’t done the math on self-hosting frontier models or are assuming that I don’t know what that math looks like www.digitalapplied.com/blog/self-ho...

even that isn't enough for some people

You should be able to find some things he actually said to that effect then, shouldn’t you

Can you cite any sources for this? I've read a few of Amodei's essays and I haven't gotten that impression at all -- in fact I've been very impressed with his approach. In what way is he "a little freak?"

yea that's like what i'm talking about - hate that, hate how it's changed search engines, hate data centers, hate that it's being flooded into the consumer market on like every level

"would you like AI to summarize this email?" I can read three paragraphs, please fuck off

Because we'll be bombarded by these ads, no matter if we buy their product or not. All their products they are ruining with these pushy ads will stay ruined, even if we click on the ads and buy their shitty products.

I never thought I live to see the biggest tech companies scamming us with malware. Like popping up buttons below our mouse pointers, putting the "YES" button above the "back" button on the phone, pre-installing bloatware, circumventing the consent dialogue… It's like we're all in a scam call now.

I first learned about LLMs ~8 years ago at a uni seminar about sentiment analysis for psycholinguistic research and it was so exciting... the problem has always been companies overselling half-baked products and policymakers refusing to distinguish marketing from reality

Agreed. It's not like AI is this evil thing that materialized from another dimension over the last 2-3 years, it's been around at least theoretically and in primitive versions pretty much as long as there've been computers. It's a technology with huge potential but 100% in the wrong hands.

Oh, and it literally stunts our abilities to do basic shit when we start relying on it. And occasionally tells children to kill themselves. And just straight up lies to us. It's like, I'm not sure what angle people can come at to see this as a good thing. It's like being told to inhale a tire fire.

i feel like there are ppl who intuitively understand why they wouldn't let Sam Altman scan their retina ever, who nonetheless talk like LLMs aren't also creeping authoritarian tentacles

Also, the weird dependance that just...appears. I have a ticket for a user who is having problems with Copilot in Outlook (Classic) and he insists he needs it to do his work. He's been with the company for 25 years and we only turned on Copilot in January.