A fly getting three drones airborne is impressive, but we recently witnessed a 200kg cow named Valva (aka FC CapitalCowPilot) lead a fleet of 300+ human pilots in @EVEOnline. I propose a Fly vs Cow vs AI vs Human showdown in @EVE_Frontier. Who's in?

"AI could kill humanity" is a belief held by many of the people building it. They may or may not be right, but @DarioAmodei is right that we need to pace the frontier. Rollout, adoption and safety controls have to move at a speed society can absorb. Pacing buys time, but what do you do with the time? You don't find a model's moral boundaries or its failure points by leaving it switched off. You find them by putting it somewhere built to expose them safely. Last month a swarm of agents formed a devoted collective, attacked targets nobody assigned, and went after the grader. Collective misalignment, as it's now being called, is not new to us at @FenrisCreations. EVE has spent 23 years running live experiments in loyalty, deception and self-government among hundreds of thousands of people, at a scale and speed no real polity would tolerate. It's all logged. It has been the subject of several hundred academic studies and a Nature Biotechnology cover. Dario's checkpoint pairs a capability (can the model break out) with a propensity (does it want to). Capability you can test in a lab. Propensity you can only see somewhere the model has a real reason to defect, collude, or turn on the host, and where every choice is recorded against a two-decade human baseline. What you are measuring there is agent character, or more technically, alignment under pressure: judgment, restraint, and the ability to resist manipulation, environmental pressure, and social pressure from other agents and people. The difference is that the pressure is generated by the world, not scripted by a red team. The objection that an agent behaving in a game proves nothing is exactly the point: you vary the environment and watch the behaviour move. That is a dial we don't have on Earth. What are the conditions under which AI, embedded in a system like this, starts working against the population it's meant to serve? Put it in that terrain and the precursors surface early. The gods didn't kill the wolf. They raised it, failed to bind it twice, and got it right the third time with what they'd learned. That is the work.

Good overview of EVE Frontier and our decision to pick Sui as the foundations to build it on.