Jev Plays Doom: What Real-Time Games Prove About Instant AI

Within days of the Jev launch, before most companies had finished their "implications for enterprise" memos, someone had it playing Doom. Then Mario. Then a deckbuilder, a typing racer, a word game. The games and real-time category filled up faster than almost any other, and it would be easy to file this under "developers gonna develop" and move on.

That would miss the point. The game builds are the most rigorous public benchmark this model has, precisely because games are the one domain where you cannot fake latency. A blog post can claim "blazing fast." Doom either renders the next frame or it doesn't.

The receipts

Numbers as reported by the build authors, receipts on each entry:

• Doom, ~10 model queries per second, ~$7/hour of continuous play. Every enemy sighted, every corridor, a fresh structured decision, ten times a second, sustained (build).
• Slay the Spire 2, ~0.7 seconds per move: reading a combat state (hand, energy, enemy intents) and ruling on a play, at conversational-pause speed (build).
• Super Mario, platforming on pure reaction: jump timing as a yes/no question asked very, very often (build).
• Wordshift, a semantic word game where the model is the referee, judging player moves live (build).

For scale: interactive chat models measure response time in seconds. Ten decisions per second isn't a faster horse; it's a different vehicle.

Why games are the honest benchmark

Three properties make game builds better evidence than any launch-thread chart. Latency can't hide: dropped decisions are visible as stutter or death, on video, immediately. The tasks are real decisions: "shoot, strafe, or advance, given this state" is exactly the constrained-question format decision models claim to own; a game is thousands of them under time pressure. Failure is public: post a Doom run and everyone sees exactly how good or bad it is. No cherry-picking survives a livestream.

So when several independent builders converge on "sustains ~10 structured verdicts a second without falling over," that's the closest thing to a third-party latency benchmark that exists for this model right now, and it arrived as entertainment.

Fair scoring, both directions: nobody is claiming superhuman play. The Doom agent plays like a caffeinated beginner; the Spire agent makes moves a good player would question. The demonstration is tempo, not mastery. Frontier models would play these games more cleverly, from a wheelchair, at prices nobody would sustain for an hour. The interesting region is exactly where Jev sits: fast enough for the loop, smart enough to be non-random, cheap enough to leave running.

The serious translation

Every product with a real-time loop has been quietly excluded from the LLM era because inference couldn't hit the frame budget. Games are the visible tip. Underneath:

Live interfaces: UIs that reorganize per keystroke, editors judging as you type (the SuperX scorer asks 61 questions per draft in ~1 second: an editorial opinion at typing speed). Simulation NPCs: characters making contextual decisions per tick without pre-scripting, at ~$7/hour-of-continuous-play economics (Doom build). Streams: moderation, commentary, and event routing on live feeds, where a verdict that arrives late is a verdict that didn't arrive; the triage patterns all apply at stream speed. Robotics: the frontier case; a model fast enough for game control loops is plausibly fast enough for physical ones, and the first device experiments are trickling into the directory now.

The through-line: "real-time AI" used to mean either tiny custom models or faking it with pre-computation. A general decision model inside the frame budget collapses that tradeoff, and the goofy Doom video is the proof of possibility everything else will cite.

Frequently asked questions

Can Jev actually play games in real time?

Per builder reports: yes, sustaining roughly 10 decisions per second in Doom and 0.7 seconds per move in Slay the Spire 2, with receipts and footage linked from each build entry. Skill is beginner-tier; speed is the demonstrated claim.

Why do people use games to test AI models?

Because games make latency and decision quality publicly unfakeable: the agent visibly succeeds or dies in real time. They're the most honest informal benchmark available for a model whose pitch is speed.

How much does it cost to run an AI playing a game continuously?

The Doom build reports about $7 per hour at ~10 queries/second, which prices continuous per-tick decision-making for NPCs, sims, and live tools. More cost datapoints: what builds actually cost.

Could this work for NPCs in a shipped game?

The economics and latency now plausibly fit; the unsolved parts are design ones: constraining decisions to stay fun, fair, and deterministic enough to debug. Expect experiments before standards.

Is Jev good at games?

No, and that's not the claim. It's fast at games: non-random, contextual decisions at frame-budget speed. Mastery would need bigger models or fine-tuning, at which point you've traded away the tempo that made it interesting. Jev vs GPT covers that tradeoff properly.