A regular LLM acts as a "policy," mapping a current state to a specific action (states → actions). Their new LLM acts as a "world model," mapping a current state and a chosen action to a predicted future state ((states, actions) → subsequent states). Instead of deciding "what to do," its explicit objective is to predict the exact environment observation that will result from the interaction history and the agent's current action.
I assumed at first that it was trained on synthetic data, but they actually went and deployed real physical hosts and virtual machines (e.g. Ubuntu, macOS, and Android) and browsers. They ran agentic systems on these continuously and recorded the actual, real-world interactions.
So it's an LLM that infers next state, or outcome,as structured data e.g. literal HTML code, UI view hierarchies, or accessibility trees.
The cold takes themselves are a fantastic proof of the care and attention the devs put into this game, as well as the depths of their design philosophy.
Well sooner or later I would expect a developer who intimately understands their code base to feel compelled to start refactoring and extracting fitting, meaningful well-leveraged abstractions.
> He was the perfect protagonist for a teenage boy: a coward, an underachiever, technically a wizard but only on a technicality, and frequently the most powerful spell in the universe was lodged in his head against his will. This will be familiar to anyone who has been sixteen.
s/frequently/initially
Also, how is a cowardly underachiever "the perfect protagonist for a teenage boy"?
"technically a wizard but only on a technicality" is obviously redundant
And what part of any of this is supposed to be familiar?
We know anesthesia "works," and we know some of its molecular targets, but we do not fully know the mechanism by which it produces unconsciousness, ie whether anesthesia eliminates experience, or mainly blocks memory, report, and integrated neural processing.
The terminal is keystroke-driven. It's character-selectable. It's reliable in a way that the GUI is not. When I drop frames, I can still enter the commands to rescue myself with some assurance they'll be interpreted, eventually.
I agree, a REPL isn't Unixy in the streams of text kind of way... or is it?
I assumed at first that it was trained on synthetic data, but they actually went and deployed real physical hosts and virtual machines (e.g. Ubuntu, macOS, and Android) and browsers. They ran agentic systems on these continuously and recorded the actual, real-world interactions.
So it's an LLM that infers next state, or outcome,as structured data e.g. literal HTML code, UI view hierarchies, or accessibility trees.