OpenAI发布GPT-6代号Astra，展现多智能体协作与复杂推理能力

I'm very, very impressed. It was able to solve one that we did not get, actually three that we did not get, as well as one puzzle that no one else in the world has solved. So it was the first one to get it. That's what I was very impressed. So you spent some time testing Astra. Yes. Like what was the first moment where it really hit you how different this model was? I went to a big conference called DEF CON, where there are these big puzzle challenges that me and my friends spent days trying to solve.

I took some of the hardest problems from that challenge and I threw them at the new model Astra. It solved this one big puzzle, which was like a bunch of Rubik's cubes that were arranged in this weird three by four pattern that you had to deduce a message from. It was able to get that three out of three times. The one thing it did need is it got the official hint from the puzzle creators, but it's the same thing we got when we were doing it.

But once you gave it that, it got the actual solution. Same thing with this other puzzle, which was this like giant dress with tons of multicolored beads all over it that you had to again find a message from. This was a task that required it to be able to actually like see and understand what's happening across tons of disparate images that were not like, we didn't really take the best pictures of this thing. Yeah. But even so, it was able to figure it out.

Ultimately got the answer once we got the main hint from the maintainers. It did it. So the model comes up with a theory. It sends off another agent to go test it, sees how it works. Then the research branches. So it has basically 10 slots that can fill with other agents to go do things. So these are all running in parallel. Main agent is just kind of orchestrating these together. If you're just starting from here and then you have to get a word here, there's no verification in between these two.

So it's very easy for the model to get lost on bad assumptions and bad paths and just disappear into something that won't work. I found this model is much better at keeping itself on track, which is a lot of the reason why it's been able to keep on task and solve these problems. And then the other side of this is really, really think that the sub-agent workflows, the swarm workflows, all of these like putting many instances of this agent together to work on much bigger, more complex problems that you probably feel like are currently possible.

It is now possible and it's worth trying.