Login
You're viewing the techdon.dev public feed.
  • Jun 20, 2026, 1:05 PM

    But for those who *do* believe that generative AI should be resisted and fought, I think there also needs to be something better to meet the need that generative AI is now meeting for some people, however badly it may be doing so.

    💬 1🔄 3⭐ 8

Replies

  • Jun 20, 2026, 10:40 PM

    @matt one way to think about this is that LLMs are a bad tool that solves problems thousands of times more inefficiently than other mechanisms. It has the appearance of efficiency for two reasons: theft and subsidy. If LLM vendors succeed, the training data will become hopelessly polluted and the theft will no longer be tenable; frontier model vendors will need to pay people to produce untainted data. The subsidies will also run out (and they are already), once the VC musical chairs stops.

    💬 0🔄 1⭐ 10
  • Jun 20, 2026, 10:40 PM

    @matt the way that public LLM services are structured, there is a relatively brief moment (~6 years at the outside i would say) where a marginalized community like the blind community can use this resource boondoggle as a way to steal from broader society. And there’s a zero-sum aspect to this; if it’s not the screenreaders stealing, it’s gonna be bitcoin dropshipping startup bros. So I have a certain level of sympathy for the transitive theft happening here; there’s a robin-hood like story

    💬 0🔄 1⭐ 5
  • Jun 20, 2026, 10:40 PM

    @matt When the music stops, though, there will be no replacement for these tools, no trillion-dollar compute clusters maintained in perpetuity, no good-faith AI-free open source volunteer community to indefinitely plunder for new code and new training data. So if you are going to burn down the world to get a better screen reader you need to make a plan right now to live in the ashes.

    (And maybe we need to make that plan regardless, since somebody else will burn it down even if we don’t.)

    💬 0🔄 1⭐ 9
  • Jun 20, 2026, 11:11 PM

    @glyph @matt not the case. GLM is now pretty much state of the art, open source and you can run locally.

    And you can fine tune these models at home today with consumer grade GPUs.

    💬 0🔄 0⭐ 1
  • 💬 0🔄 0⭐ 0
  • 💬 1🔄 0⭐ 2
  • 💬 0🔄 0⭐ 0
  • Jun 21, 2026, 8:29 AM

    @Migueldeicaza @glyph If I'm reading that correctly, that's 256 *gigs*, not *megs* of VRAM.

    And in which world is that a "consumer grade GPU"? My favourite IT retailer's filter ends at 96 GB, which is still more than any non-server machine I ever owned had in total RAM.

    💬 1🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0