Real time as a use case: when the frame budget sets the architecture
Games, trading and voice agents share one problem: the decision has to fit a window a token-by-token model cannot hold.
This translation was produced automatically using AI. The German version is the editorially reviewed original.
TypeSafe demonstrating a model that plays Doom looks at first like a stunt. It is the cleanest argument in the whole announcement, because it makes visible a property that cannot be optimised away: a fixed time budget.
This piece collects the cases where latency is a condition rather than a comfort, and describes what they used to fail on.
01. Why Doom is a better argument than the benchmark
A game supplies the conditions a benchmark hides. The action set is fixed and small. The state is fully describable. And there is a hard window inside which the answer must exist, otherwise it is worthless no matter how good it is.
That is where a token-by-token model fails on procedure rather than on intelligence. So the demonstration is not a toy: showing that a decision fits a frame budget also shows it fits a checkout budget.
02. The game projects as a stress test
The community worked through this quickly. typesafe-snake has code generate the legal moves and the model pick one per tick. jev-plays-pokemon-red runs on an emulator and calls the model only at branches, using Brier scores on its own predictions as the quality measure. Mario, drone and driving demos followed, several built in under an hour.
The blitz chess test is the clearest number in this group. Jev answered in roughly 2.6 seconds per move while comparison models spent 6 to 15 seconds and lost on time from winning positions. Move quality was not the deciding factor; holding the budget was.
03. Trading, voice agents and browser automation
The same property carries applications with nothing playful about them. jev-trader builds a sub-second market decision loop and documents dry run and live paths separately. QuantDinger puts the model in front of an entry as a pre-trade gate and blocks rejected orders, with a logged fallback path.
For voice agents the question is whether an interruption is allowed and when a reply begins. That has to happen at conversational rhythm, well under a second. Browser automation has the same profile: which element is meant, out of a visible set, immediately. Voice-controlled browsing is reported at roughly 0.0002 US dollars per decision at about 300 milliseconds.
04. What this means for ordinary systems
Most B2B systems have no frame budget, but they have budgets nobody wrote down. A checkout cannot wait on a fraud score. A search cannot wait on its own results being reordered. An assistant that takes three seconds does not get used, however good the answer is.
Those budgets were the reason checks were left out. Remove the cost and the reason goes with it. How a response time budget for an AI assistant breaks down in the first place is in Response time of an AI assistant.
05. The hard limits in real-time operation
One choice call carries up to 255 options, a scale 2 to 10 levels. Rate limits sit at roughly 250,000 tokens per second and 1,200 requests per minute, and the vendor describes them as dynamic. For a loop deciding ten times a second, the request limit rather than the latency is the binding constraint.
And it stays text. No image, no audio, no video. Anyone who wants a game state or a camera scene judged has to translate it into a state description first. That translation is the actual engineering work, not the model call.
