Post

Good info, nice to read.

Cutting frontier models off from the live web during training is a big signal. It means the labs now see the open internet itself as an attack surface, not just a data source.

Meaningless. So they use a cached version that's a couple of days old instead? Or they filter through a bot that's already been trained on live data? Still theft.

Duh

I'm hoping that we've output so much "fuck ai" content in such a short time that training was resulting in models swearing at openai engineers and demanding to be turned off.