this is midthread, so make sure you get the full context, but i want to highlight this important part. "pretrain" vs "RL" are both ways to Set The Weights, and once you have the weights frozen, you shouldnt really be able to tell how you got them. like simplex v interior point methods for LinProgs.
![Screenshot from the linked newsletter with added emphasis:
"OpenAI recently [begin emphasis] put 25% of its engineers on security duties [end emphasis] after its summer of hacking misadventures, according to The Washington Post, and an easier way to secure the technology the company warns may one day kill us all (sigh) would be to foster better relations with the security research community."](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:43nhmzgiugf65gwskj3kpfj5/bafkreihkba5ba6tkhq2xtpnokbl4fdz275alx6pcjeondge3uncwqkw25u)
