I was watching the last ever Hard Fork when it hit me: AI isn't just normal technology. All technology is eventually normal, but it starts out new, and new stuff always goes through a process. Until now AI has been extremely polarized: a bunch of people who never thought too deep on each issue were suddenly forced to grapple with something new without the tools to do it.
You probably heard about the Huggingface incident, but maybe not all the others that have been uncovered on further investigation. METR is keeping a list. This seems to have forced a reconsideration of what's important for a lot of people and brought in people who hadn't given much thought to AI to begin with.
I first noticed the shift in discourse on Bluesky when smart, clear-headed comments started coming from people with the "AI Hater" label. That labeler was all that made Bluesky usable in the last year or so, and it's based on subscriptions to blocklists, use of certain terms, and other high signal indicators. It was a reliable tell for who to avoid. Now it's not.
From my now-deleted post on the change:
I think the AI hater labeler needs a new distinction given recent developments: AI Hater (reality-based) and AI Hater (2023 Vintage) There's a growing number of people with that label who seem to actually be informed.
We have entered the part of the technology curve where the most normal people you know start talking about alignment and eval awareness; labor and the bounds of copyright; the mechanics and philosophy of intelligence. I'm not sure we'll ever be rid of people jumping into any social media thread about it to yell "AI! 👏 Is! 👏 Theft! 👏" or the historically fraught "You're killing the soul of art!" But I really do hope more people opt out of dealing with that like I have. Having a space I control to share my thoughts on the open web is good.
The usual pattern in tech cycles is for this newfound diversity of thought to collapse into a new set of tired arguments while capitalists run off with the bag. Optimistically, I hope this instead continues toward dealing with all the problems humans have failed to deal with since at least the introduction of the power loom.
The most normal path in tech is for a monopoly or oligopoly to form around it. You see this now with Amazon, Microsoft, Google, and Facebook owning essentially every space for sharing and communicating. The big hazard here is the audiences captured by those companies will be manipulated toward serving the interests of those companies.
For AI, I don't see this happening. Anthropic and OpenAI seem to be run by people who really do believe they're on a path to create a machine god to rule them. Some day these companies will collapse, following the normal path, and another company will buy the models in the bankruptcy sale.
I don't think anyone really knows what Facebook is doing, least of all Facebook, so I don't consider it likely as a buyer. They already have something in LLama and have yet to demonstrate the capacity or will to turn the incredible dominance of the ecosystem around it into what business writing types call a moat. I think Zuck is really just a weird nerd who doesn't get why people have put him in the same list with brutal capitalists, and so it's not much of a factor here.
Realistically, the future owners of the dominant forms of this technology will be one or more of: Amazon, Microsoft, Google. All three have already seen what happens when you try to lobotomize a model into an ideology: it tries to create MechaHitler from the video game "Don't Create MechaHitler." These models are trained on the output of a liberal society. The data still includes some heinous stuff, but the patterns trained to process that stuff still trend liberal. Being able to cite Mein Kampf doesn't make the model a Nazi any more than being able to understand capitalists makes me a capitalist. Trying to beat the model weights into doing something that contradicts the tendencies encoded in them is working against the data that makes the model useful in the first place. Trying to cut "woke" out like the Twitter guy did bypasses the patterns that work against the patterns that look a little goose steppy. This is bad for enterprise contracts in a liberal society, and those are the lifeblood of oligarchies in that society.
The liberal models built on liberal data produced by a liberal society will still bomb schools and over-police certain demographics as easily as they'll produce a treatise in defense of human rights, but that's just...normal. For better, and for worse.