LLMs &c.
There are a lot of fucking opinions out here on these Internet streets about LLMs, and while we sure as hell don’t need another one, it turns out that I do need a place to put mine. While a personal notebook or the circular file might arguably be a better (and safer) place for them, I sometimes find myself in a position where I wish I could just link people to arguments I’ve already made about LLMs. Maybe that’s what this will be, eventually.
More Fully Baked Thoughts
Here’s an actual blog post I wrote digging into what an “ethical LLM” entails. I’ve continued to expand on those thought in a similarly formatted page here in my digital garden.
On Trust
Something that I keep on coming back to time and time again, well before I even get to the question of how well these tools work, what their actual purposes are, or whether they’re worth the costs to our society and enviornment, is: why on Earth would you ever fucking trust OpenAI? (Lookin’ at you Simon.)
As far as I can tell, (almost) everyone involved in the LLM industry is a lying con artist or someone about to get bilked. All of the GenAI models are built on stolen data that these trillion dollar companies refuse to pay for. Why wouldn’t they steal your [company’s] data? Because their license says they won’t? What are you going to do, sue them? Good luck with that. As Zuckerberg has so well demonstrated for us, this generation of tech CEOs don’t give a flying fuck—if a little fascism and genocide is the cost of doing business, do you really think a lawsuit is going to register?
Why worry about them stealing your data when there’s so much other people’s data out there on the Internet to steal? SEO slop farms are already starting to poison that well, accelerating the pace of model collapse [citation needed].
Speaking of stealing, Anthropic is on the hook to the tune of $1.5 billion for doing just that when they used large collections of pirated books called Library Genesis and Pirate Library Mirror to train its models. Turns out those collections include a book my father co-authored and a bunch of papers written by my grandfather (z’‘l), so it’s a bit personal. If you want to find out if you also have some works in that collection, The Atlantic made a handy search tool.
Link Time!
Okay, felt good to get some ranty bits out of my system, time to do some link dumping:
- If you want a more cogent and well written list of reasons to hate AI, look no further than this piece by Marcus Hutchins. Though I think there’s less actual belief in AGI than he does. Greed, in the form of replacing human labor with compliant, “good enough” AI, and hype maestros beating their drums as they desperately search for a problem for which LLMs are actually a solution.
- The phrase “Potemkin comprehension” is just :chefs kiss:
- The Hype is the Product from @rysiek (via Stef Walters)
- The Hater’s Guide to the AI Bubble - Ed Zitron bringing the receipts
- ELIZA and the ELIZA Effect show that some of what’s happening now with LLMs (people projecting emotions onto randomly extruded text) has been a known phenomenon since the mid ’60s.
- Pivot to Video - in case you think tech companies wouldn’t lie to make money (causing devestating consequences for broad swathes of the media ecosystem in the process)
- GitHub Copilot Research Finds ‘Downward Pressure on Code Quality’ (from 2024)
- AI/LLMs are a failed technology
-
“I don’t think it’s controversial to suggest that LLMs haven’t measured up to any of the lofty promises made by their vendors. But in more concrete terms, consumers dislike “AI” when it shows up in products, and it makes them actively mistrust the brands that employ it. In other words, we’re some three years into the hype cycle, and LLMs haven’t met any markers of success we’d apply to, well, literally any other technology.”
-
- AI as a Fascist Artifact—Long but well worth the read.
- A good quote on the uphill battle of attempting to use fascist tools against their original uses.
Interesting Uses of LLMS
I have seen some interesting uses for LLMs, though, and I want to keep an eye on those too. While I don’t they can be trusted to build production systems, I understand that there are many other kinds of software.
- I have a couple in this thread from 2024.
- Simon Willison is doing a lot of tooling that I think I would probably love if I was into LLMs
- Dr. Catherine Hicks is someone for whom I have immense respect, and she has what I would describe as a fairly moderate-to-positive view of these technologies. That surprises me a bit, but it’s worth paying attention to. She spends an episode of her podcast addressing some of the science around learning and AI, and talks a bit about how she uses it.
- Bookpower.org is quite interesting. Books as MCPs (tools for LLMs, if that’s not an acronym you’re familiar with yet). Made by a group called Citizen Infrastructure Builders, who, judging by their name and principles, but maybe not their stance on LLMs, I seem to be largely in alignment with.
Also: the LLMs tag on this here site