Post

The dialogue around LLMs is kind of broken both in terms of critique and advocacy. For critique, we need to focus on how we attack fascist tech power consistently — LLMs are only one part of that; merely saying "AI bad" has simply ended up with a lot of people using AI quietly while feeling shame...

And amongst AI advocates, there's too often the same tedious fallback to arguments of inevitability and "progress", which excuse the enabling of the bad actors in tech, and advance their authoritarian project. It's a failure of imagination, in addition to being complicit in their political goals...

The ultimate "right" answers are going to involve complex solutions that will feel dissatisfying at first, and that's hard to get across to people. First, there isn't one killer alternative, just like there shouldn't just be "use Bluesky instead of Twitter". There needs to be a healthy ecosystem...

And all of the options in that ecosystem have to be radically different than what people think of as being intrinsic to "AI" today. It is possible to build systems that use data consensually. It is possible to build systems entirely on renewable power. It is possible to build actually open systems…

Once we have such tools, and once people have reframed their definition of what "good" AI is to be about these concerns rather than party tricks, it will be possible to have more constructive conversations. And there could be 1000s of different kinds of models, rather than just 2-3 big companies.

It is not science fiction, one such creator-built system already exists. We just need one of these tools in every little area, rather than the giant corporate AIs trying to be all things to all people, while making tools that tell kids to self-harm. www.anildash.com/2026/04/28/o...

Anil, you must not know the union leaders that I dealt with. Or you know a whole different bunch. The truest of conservatives, in the 'as it was in the beginning, is now, and ever shall be' sense.

Maybe it's because we are too busy talking about open-source, open-weights, + trying to avoid panicstruck politicians & regulators attempting to ban open dev/training of a truly "free & libre" model "in case it is used for abuse or terrorism", because of threads like yours? bsky.app/profile/anil...

This is also why I think Apple’s original ML strategy was the right play. They had been focused on creating hundreds of small, local models that quietly helped me do the things I actually do daily. They‘d been adding ML engines to their SoCs for years before LLMs to power these dedicated models.

This is the conclusion that my (admittedly very local, admittedly very artsy) peer group landed on though! LLMs the technological concept are totally fine - if you do local, open-source models running on consensually trained data and clean energy. But "AI" as it *actually exists* is loathsome.

Paging @ntnsndr.in nathanschneider.info/2026/06/an-i...

have u seen @cocore.dev by @devingaffney.com?

This indexes into something that in my mind comes from Gebrus and M. Bender‘s work and advocacy and what I call "proper scooping". What‘s the problem that needs solving? What problem paths are we addressing? (With what desired outcomes?) Thus, where do we source the data from?

*scoping Goddamnit autocorrupt

on the other hand, "scooping" makes me think of ice cream, and that's nice