| ▲ | aitchnyu an hour ago | |
At one point, I thought model collapse (AI models getting trained on AI output) was a solved problem since we know what content improves a model and what weakens it. Now I'm not sure. | ||
| ▲ | suddenlybananas 44 minutes ago | parent [-] | |
I think model collapse is plausible in non-verifiable domains and implausible in verifiable ones since there is still a cheap external signal. | ||