18 ms·

There are many assumptions here. The generative AIs we have today are excellent at transforming things they learned and rehashing it into something seemingly ne

There are many assumptions here. The generative AIs we have today are excellent at transforming things they learned and rehashing it into something seemingly new. The problem is, they learned all this based on the input of the humanity en masse. When you train LLMs on the output of LLMs, it gets significantly worse. So your prediction could of course be true but only if a major breakthrough happens.

prometheon1 3y ago
> When you train LLMs on the output of LLMs, it gets significantly worse. That is also quite an assumption, it could be that training on the output of better LLMs also reduces this worsening of output. There might even be a tipping point where the LLMs get good enough that training on their output is better than training on the output of humans.
- thfuran 3y ago
  >That is also quite an assumption And, as I understand it, one that is already demonstrably false: https://arxiv.org/abs/2306.11644 https://arxiv.org/abs/2306.11644
- tmccrary55 3y ago
  Perpetual learning machines
Brian_K_White 3y ago
That could be said the same about humans, if you took a bunch of uneducated children and just let them tell each other their own ideas with no one to teach them anything for real, they probably do at least a little better than the current fancy statistical spellguessers, but they probably won't do exactly great, because of the same issue of the blind leading the blind. And really in reality we do have that actual problem to some degree with the presense of non-blind adults. Actual humans are a mass of mixed clued and clueless with a lot of bad input feeding output feeding other bad input around and around and around, not even counting the legitimate fair differences of opinion. So it's a problem, but I don't think it's a fundamentally new or worse problem than we already have,and have already always had. The fix is I don't think there is a fix for that any more than there is for the same thing in humans. There just will always be bad data feeding bad reasoning right alongside the other good data feeding good reasoning. It's probably wrong to ever expect anything else, and fail to operate from that assumption rather than the idea that there might ever be some resolution where we don't have to worry about that.

prometheon1 3y ago

> When you train LLMs on the output of LLMs, it gets significantly worse. That is also quite an assumption, it could be that training on the output of better LLMs also reduces this worsening of output. There might even be a tipping point where the LLMs get good enough that training on their output is better than training on the output of humans.

thfuran 3y ago
>That is also quite an assumption And, as I understand it, one that is already demonstrably false: https://arxiv.org/abs/2306.11644 https://arxiv.org/abs/2306.11644
tmccrary55 3y ago
Perpetual learning machines

Brian_K_White 3y ago

That could be said the same about humans, if you took a bunch of uneducated children and just let them tell each other their own ideas with no one to teach them anything for real, they probably do at least a little better than the current fancy statistical spellguessers, but they probably won't do exactly great, because of the same issue of the blind leading the blind. And really in reality we do have that actual problem to some degree with the presense of non-blind adults. Actual humans are a mass of mixed clued and clueless with a lot of bad input feeding output feeding other bad input around and around and around, not even counting the legitimate fair differences of opinion. So it's a problem, but I don't think it's a fundamentally new or worse problem than we already have,and have already always had. The fix is I don't think there is a fix for that any more than there is for the same thing in humans. There just will always be bad data feeding bad reasoning right alongside the other good data feeding good reasoning. It's probably wrong to ever expect anything else, and fail to operate from that assumption rather than the idea that there might ever be some resolution where we don't have to worry about that.