• mindbleach@sh.itjust.works
    link
    fedilink
    English
    arrow-up
    2
    ·
    2 days ago

    We know a lot more about recurrent neural networks, and how to train them. This all took off thanks to consumer hardware.

    The bubble’s gonna burst - because the bubble is morons teaching a helicopter to mow grass. But if you missed it, we now have a machine that can describe a scene in a video, and build a 3D model of it, including guesses about the backsides of objects. It’s pretty gloopy, but so were still images, all of two years ago.

    This latest push for “AI” either shatters a bunch of things we insisted would require a true intelligence - or else it’s a near miss for building a true intelligence. LLMs aren’t the right kind of deep neural network for AGI. But we’ve proved you can build any shape of network just by pouring data through it. The trick now is finding the right shape.