Watch my gameplay
Mythos Preview (Project Glasswing) found over 1,700 confirmed security vulnerabilities across a thousand open source projects, including bugs that had gone undetected for decades. And Fable beating Pokemon Red without the additional RAM access or other scaffolding that Opus required is a step forward.
That said, the video we got is heavily edited and cut. When you actually analyze it, it still contains a lot of head-into-wall behavior that looks pretty stupid on closer inspection. Fable is not competent at this. It is still very brute-force-y. We have no idea how long it got stuck in absurd loops before the footage we see.
Agentic behavior still does not scale beyond limited-scope coding tasks. The world model of these systems remains very fragile.