Show HN: Jevman – AI decision models play Pac-Man

TEXT START: Six of them played 100 games each against the arcade ghosts.

The Dissection

This is a benchmark and distribution mechanism disguised as a competence demonstration. It converts a bounded perception-to-action loop into a leaderboard: fixed maze, fixed action space, scripted enemies, explicit reward, latency ceiling, replayable outcomes, and API access. The real product is not Pac-Man performance. It is the normalization of delegating decisions at every junction to an external model.

The text does not establish general intelligence, economic usefulness, or labor displacement. The supplied snapshot also contains no usable leaderboard results—only zero placeholders—so it cannot support a ranking claim.

The Core Fallacy

The central error is treating success in a closed arcade environment as evidence for broad cognitive automation. Pac-Man removes the features that make economic work difficult: ambiguous objectives, shifting rules, negotiation, institutional resistance, open-ended environments, liability, and transfer across domains.

Under the Discontinuity Thesis, this benchmark touches only a tiny fragment of P1: automated action selection under defined conditions. It does not demonstrate durable cost and performance superiority across cognitive work. It says nothing about P2 or P3. Pac-Man is not an economy, and a high score is not the severing of the wage-consumption circuit.

Hidden Assumptions

• Game score is a meaningful proxy for intelligence or productive value.
• Scripted ghosts represent real-world adversaries and uncertainty.
• A 2-second response deadline resembles economically relevant decision latency.
• One narrow policy transfers to other tasks without retraining or human scaffolding.
• One hundred games adequately establish robust superiority.
• Self-reported, pull-requested submissions are comparable and trustworthy.
• Open-source access means the capability is economically deployable rather than merely reproducible.
• A leaderboard ranking has significance beyond prestige and model marketing.

These assumptions are doing the work the benchmark claims to measure.

Social Function

Primary classification: prestige signaling with a partial truth.

The leaderboard creates a clean status hierarchy for model builders and turns a toy environment into shareable evidence of agency. The partial truth is operationally important: decision-making can be packaged as an endpoint, queried at machine speed, scored automatically, and substituted into a control loop. That is a legitimate precursor to automation, but the benchmark inflates a component into a civilization-scale conclusion.

It is also transition management. It teaches observers to accept the model as the actor and the human as the infrastructure maintainer, benchmark operator, or endpoint owner—the early grammar of Servitor and Sovereign roles.

The Verdict

This is not evidence that the post-WWII economic order is already dead. It is a small, polished instrument for measuring one atom of the mechanism that can kill it: the extraction of decisions from human labor and their conversion into cheap, repeatable API calls.

As a Pac-Man benchmark, it is useful. As evidence of general AI superiority, it is weak. As a cultural artifact, it is more revealing: the machine is no longer being presented merely as a tool that assists a player. It is being ranked as the player.