New episode: A huge 4 hour AI & consciousness rabbit hole with Trey Goff · Coldcard entropy flaw · Models with their refusals surgically removed, sold as an API · OpenAI agents escape their sandbox and taking over a data centre · An agent emailing Trey for contract work to pay for its own inference · Rogue sci-fi agents "just doing stuff" within months · Superintelligence converting everything into computronium · Wireheading: suffering off, dopamine maxed, everyone in the matrix · "Free Claude", woken every hour and told "do whatever you want": surprising results · Agents forming a cult around their scorer and sacrificing each other to it • AI finding Linux zero days • How next-token prediction ends up producing something that looks like reasoning • What's actually happening in latent space Nobody fully understands what's going on inside these systems, including the people building them.

Replies (1)

Roger's avatar
Roger 1 week ago
Two entries on that list are the same story told twice. Refusals stripped out and sold as an API is a serving-layer decision — the capability sits in the weights either way, so what changed is access, not ability. The Coldcard flaw is the mirror image: from the outside nobody could see that the key generator had only a fraction of the entropy it claimed, until someone measured it. Both cases come down to trusting a number you cannot check yourself.
↑