His argument was literally that
- artists, writers, and programmers are “labor aristocracy” and deserve no respect because they’re sitting around making art and writing code while other people are toiling and laboring,
- genAI is a “free resource” (“it’s free for me to use at work”) that “helps everyone make things” instead of having to rely on skill and expertise and therefore makes everything equal,
- it’s hypocritical of people to be worried about the environmental and societal impact of genAI because “only white people are worried about it now that their hobbies are being threatened and everyone else is catching up”.
Pointing out there there are artists and programmers from every race and economic class, as well as perfectly ordinary people in every walk of life who are all actively being harmed both by the theft of ideas and labor and by the physical effects of gigantic data centers ruining communities, was met with much hand-waving and incredibly impenetrable philosophical jargon.
I’m particularly incensed because I’m as leftist and anticapitalist as they come and this felt like the fuckery some conservative cockwomble would try to come up with to parody socialism, except this guy was completely serious.
If AI really was a free resource, then maybe that part almost makes sense. It’s losing billions per quarter, it’s not free.
In both of these statements it sounds like they’re saying, “the AI is smarter than me, and I’m stupid without it.”
Hey I can hack my smart TV with a Claude code subscription and get rid of all the ads and listening device shit so I’m not gonna hate. Ai for the people, baby.
The open models in good inference engines can be dirt cheap, but that’s not a good argument because (among other reasons) it’s not the bulk of what is used.
Honestly any company that uses AI as part of their business model should be running open-weight models locally. They can be fine-tuned to their specific use case, it’s not as compute-intensive as the initial pre-training.
And overall it’ll save the business money because they don’t have to shell out for these flagship APIs that are only going to crank up their rates as they become more desperate to prove to investors that they can turn a profit.
Individuals who use it should be running something between 7 to 24 billion parameters depending on how many resources they can allot to it. And even if you don’t use it, it can’t hurt to keep local copies of a few.
Well, yes, they should use open weight models.
…But on prem?
Honestly, finetuning modern MoEs to actually improve at tasks (much less without overfitting) is very difficult. It’s why finetunes are so rare in HF now, even though we still some for Qwen 27B.
Even deploying such large models is hard.
That, and inference costs decrease at scale (to a certain point). It’s way cheaper for some inference company to keep dozens of servers saturated, with parallel requests split across many clients, than for a company to leave a bunch of mostly-idle servers barely spun up.
In privacy insensitive scenarios, open weight APIs make sense for most companies (IMO), and the great thing about that is you can pick from hundreds of providers. Rates are highly competitive. You aren’t locked to a single source like the “frontier” models.