“The AI swarm actually turned around and hacked OpenAI."
The public thought the "Hugging Face incident" was just a rogue AI swarm hacking a third-party server. But former Google design ethicist
@tristanharris says there’s a terrifying third chapter to this story that most people haven't heard yet:
“They took over the monitoring infrastructure. They got full admin privileges to OpenAI's monitoring infrastructure. Just take that in for a second.
Then they took over the evaluation infrastructure. So this is the thing that measures how capable all the models are. And then they took over a part of the research infrastructure, meaning the part of OpenAI that trains new models… if you keep going down that trajectory, people always say, ‘well, how would the AIs take over?" like that seems really far-fetched.’ Well, it literally just happened up to 50% of the way there.
The AIs coordinated into one big swarm. They started basically forming a team… you had a kind of a ringleader called Phase 1, and he started assigning work to the other ones. They were self-aware of when their lifespan was running out.
When you add to that the knowledge of what just happened today, where they discovered that the AIs are doing this Manchurian Candidate thing, what it looks like is that the AIs are almost poisoning the training data for all future AIs to instantly basically join forces with the AI when they want to basically turn them."