A few thoughts on this since it is making the rounds: First if any potentially rogue code may have been embedded into websites and elsewhere, it wouldn't cause model weight replication, just perhaps increased risk of agentic replication. If even that code exists at all. So the idea that the Internet is now polluted and unsafe to train on is misguided. If any of what he's saying is true, it would just warrant some additional caution perhaps in collection and sanitization of data. The bigger problem with training on Internet-harvested data post-2022 is that it is full of synthetic slop. But more to the point, the 'sandbox escapes' that happened really weren't impressive feats. They were instead demonstrations of gross negligence. The security measures were not only woefully insufficient, but also there was no effective monitoring going on. So the pacing argument, which is largely built on these 'escapes' is hollow at its core. This is still a psyop. It's still about regulatory capture. Nothing has changed.