Post

we would not actually know how and why either CSAM or heroin or burning plastic is bad and affects victims/drug users without actual studies involving their impact. you cannot just from-first-principles every single possible issue.

Will Stancil@whstancil.bsky.social

It is absolutely bananas that one of the most prominent anti-AI experts in America is talking like this. Talking to ChatGPT isn’t like doing heroin or downloading child porn. It will not hurt you and has no ethical valence. It will help you understand the nature of something that is now endemic.

9 reposts75 likes

it is honestly rather revealing of the hollow moralism at play here that she thinks CSAM is illegal bc icky bad and not "necessarily entails lifelong trauma". getting CSAM legally distinguished from e.g. nude art, actually did involve victim testimonies!

ultimately this refusal to engage with science at all is worse than the worst caricature of the rationalists

also, uhhh, go up one level from "heroin" and you're at "opioids" which, except for heroin, are all recognized as having legitimate medical purposes three of them are on the WHO's essential medicines list!

even better! (I just assumed that controlled substance schedules for hard drugs were mostly the same the world over)

A lot of people's priors are locked in 2024, and haven't adjusted to the reality that, in 2026, you can run an open source model that is approximately Claude Sonnet quality on a $5000 Mac M5 and it uses less than 300 watts to operate. No water. And it's way, way more capable than the 4o they mocked.

OK, but the power generation/water used for training of open source models is modest by comparison to the frontier models. And if you're pulling down models that already exist, the marginal environmental cost of using them is...300 watts.

And with virtually all new power added to the grid being battery-backed solar (which is the trend worldwide, and will only accelerate thanks to Swanson's Law and Wright's Law), the marginal water and greenhouse gas costs of those 300 watts will also rapidly approach zero.

I think you misunderstand Bender, she is stuck in, like, 2018, not 2024.

weird how i keep hearing “open source models are so capable” and then i look at peoples repos and it’s just claudegen code, everywhere. the *idea* of powerful personal agentic software seems vastly more popular among enthusiasts than the practice of it,

It’s still a pain in the ass and 100-200/mo is more palatable than dropping 5k+ on a dedicated machine that generates tokens slower and overall slows to a crawl when generating. It’s not yet The Year of Linux on the Desktop.

Exactly. The time will come when it will make sense for end users to run their own LLMs locally, though. Apple is famous for not entering a market until it's ready for a polished, premium product, and then dragging the rest of the market to create the low-cost, commodity alternatives.

It's consistently 3-6 months behind the frontier models thus far. But there's no reason to expect that state of affairs will continue forever. Why do you think the frontier incumbents are so eager to get regulation that stops development/home use for everything except their platforms?

There is an academic tradition of Comp Sci professors who are sometimes pure academics that rely on grad students to be the interfaces to the base and vulgar machine