If the threat frontier companies describe is real, transparency is the cheapest insurance society can buy. Let them open their training data, I bet it's a lot of questions.
Training data is "approximately everything"; I've never understood why people bother to ask for it given all the startups in this scene with competitive models.
The RLHF, RLVR etc. processes, those are likely more useful: semi-automatic training on outcomes, not mere inputs.
Monitoring, evaluations, how things fail, that's where I think transparency is key.
Though even then, these companies currently publish much about their own failures: when they are PoC they get ignored as "hype", "fear mongering", and "doing it for the IPO", when they're in-the-wild they still get called "hype", "fear mongering", and "doing it for the IPO", but sometimes they also get sued.
2 comments
[ 61.4 ms ] story [ 89.4 ms ] threadTraining data is "approximately everything"; I've never understood why people bother to ask for it given all the startups in this scene with competitive models.
The RLHF, RLVR etc. processes, those are likely more useful: semi-automatic training on outcomes, not mere inputs.
Monitoring, evaluations, how things fail, that's where I think transparency is key.
Though even then, these companies currently publish much about their own failures: when they are PoC they get ignored as "hype", "fear mongering", and "doing it for the IPO", when they're in-the-wild they still get called "hype", "fear mongering", and "doing it for the IPO", but sometimes they also get sued.