I’m not surprised they ultimately felt like GenAI isn’t useful to what they’re trying to do. Game dev has known about this type of generation for a while now (see https://en.wikipedia.org/wiki/Model_synthesis and the sources linked at the bottom) and it takes a lot of human effort to curate both the training data and the model weights to end up with anything that feels new and meaningful.
If I shuffle a deck of 52 cards, there is a high chance of obtaining a deck order that has never occurred before in human history. Big whoop. GenAI is closer to sexy dice anyways - the “intelligent work” was making sure the dice faces always make sense when put together and you don’t end up rolling “blow suck” or “lips thigh”.
It’s very impressive that we’re able to scale this type of apparatus up to plausibly generate meaningful paragraphs, conversations, and programs. It’s ridiculous what it cost us to get it this far, and just like sexy dice and card shuffling I fail to see it as capable of replacing human thought or ingenuity, let alone expressing what’s “in my head”. Not until we can bolt it onto a body that can feel pain and hunger and joy, and we can already make human babies much more efficiently than what it takes to train an LLM from scratch (and they have personhood and thus rights in most societies around the world).
Even the people worried about “AI self-improving” to the point it “escapes our control” don’t seem to be able to demonstrate that today’s AI can do much more than slow us down in the long run; this study was published over 5 months ago, and they don’t seem to have found much since then.
I’m not surprised they ultimately felt like GenAI isn’t useful to what they’re trying to do. Game dev has known about this type of generation for a while now (see https://en.wikipedia.org/wiki/Model_synthesis and the sources linked at the bottom) and it takes a lot of human effort to curate both the training data and the model weights to end up with anything that feels new and meaningful.
If I shuffle a deck of 52 cards, there is a high chance of obtaining a deck order that has never occurred before in human history. Big whoop. GenAI is closer to sexy dice anyways - the “intelligent work” was making sure the dice faces always make sense when put together and you don’t end up rolling “blow suck” or “lips thigh”.
It’s very impressive that we’re able to scale this type of apparatus up to plausibly generate meaningful paragraphs, conversations, and programs. It’s ridiculous what it cost us to get it this far, and just like sexy dice and card shuffling I fail to see it as capable of replacing human thought or ingenuity, let alone expressing what’s “in my head”. Not until we can bolt it onto a body that can feel pain and hunger and joy, and we can already make human babies much more efficiently than what it takes to train an LLM from scratch (and they have personhood and thus rights in most societies around the world).
Even the people worried about “AI self-improving” to the point it “escapes our control” don’t seem to be able to demonstrate that today’s AI can do much more than slow us down in the long run; this study was published over 5 months ago, and they don’t seem to have found much since then.