Our key finding is that by injecting information through an external synthetic data verifier, whether a human or a better model, synthetic retraining will not cause model collapse.
Yeah if you have a source of truth then your model is basically getting trained on that.
My point was that having a verifier means your not really training a model on another model’s data, it’s basically as if you get new raw data from a non AI source
Our key finding is that by injecting information through an external synthetic data verifier, whether a human or a better model, synthetic retraining will not cause model collapse.
Lol, so to make a great model, they just need to have an even better one available first or a human who can verify every single thing it ingests.
This assumes everything is valid on the external. If one slop cluster feeds off another - a slopveyor? - then there is nothing external for the validation hall-monitor to compare against. They’re trusting another model’s output as if it were gospel.
Model collapse isn’t a thing anymore. https://arxiv.org/html/2510.16657v1
Yeah if you have a source of truth then your model is basically getting trained on that.
It’s like already having the answer
The point is that it only needs to comprise a very small part of the model.
My point was that having a verifier means your not really training a model on another model’s data, it’s basically as if you get new raw data from a non AI source
Lol, so to make a great model, they just need to have an even better one available first or a human who can verify every single thing it ingests.
Hmm, call me skeptical on this claim.
This assumes everything is valid on the external. If one slop cluster feeds off another - a slopveyor? - then there is nothing external for the validation hall-monitor to compare against. They’re trusting another model’s output as if it were gospel.
LOL OK