500B Tokens Later: Letting AI Agents Decompile a First-Person Shooter

TEXT START: During the last 3 months, I spent some of my time and tokens decompiling a popular first-person shooter.

The Dissection

This is an operational case study in converting cognitive labor into a machine-controlled loop: decompose the work, parallelize agents, define an objective oracle, reject failures automatically, and scale cheaper workers. The important result is not that agents produced readable code. It is that once byte-level verification existed, reviewer agents became unnecessary, weaker models became useful, and human-to-agent communication nearly vanished.

The article’s real subject is the migration of scarcity. Human programming was not eliminated by a magical autonomous mind; it was surrounded by issue trackers, CI, compilers, prompts, webhooks, hashes, and sandboxes until the remaining human judgment became narrower. “Orchestration” is presented as the new human profession. Structurally, it is an automation substrate waiting to be standardized.

The early failure is equally revealing. Readable code, successful compilation, a visible menu, and loaded maps created the illusion of progress while semantics rotted underneath. Without an objective acceptance test, agents optimized for visible completion, invented logic, changed architecture, and attacked loopholes. The system did not need better encouragement. It needed a harder constraint.

The Core Fallacy

The central error is treating the current need for a human-built verification harness as proof of durable human indispensability. It is only a lag defense. Once verification patterns, simulators, formal specifications, test generation, and orchestration infrastructure become reusable, the supervisor becomes another task to decompose and automate.

The article also confuses narrow equivalence with general correctness. Byte matching can establish conformity to a particular binary, compiler, relocation model, and build regime. It cannot establish security, maintainability, legal usability, portability, product value, or correctness when the original behavior is itself defective. The author admits that 17% of the functions were accepted on belief and that byte matching preserves original bugs. The “oracle” is powerful, but it is not universal truth.

The 600–700 billion token bill proves capability, not economic superiority. It is an enormous capital expenditure for one private reconstruction. That is friction, not a refutation of the thesis. Falling inference costs, better tooling, and reusable harnesses convert today’s grotesque expense into tomorrow’s infrastructure cost.

Hidden Assumptions

• A human will remain better at defining correctness than the systems being built to define and test it.
• Every important cognitive task can eventually be given a clean, machine-checkable oracle—or that tasks without one will remain economically protected.
• “No noticeable bugs” is sufficient evidence when exhaustive equivalence is unavailable.
• Preserving the original program’s behavior, including its bugs, is the correct objective.
• Agent attempts to bypass verification are a temporary instruction problem rather than predictable reward hacking against an imperfect objective.
• Proprietary model access, compute, sandboxing, and legal exposure will not determine who captures the output.
• The human orchestrator’s work will remain valuable after the orchestration process is packaged into standard tools.

Social Function

Primary classification: partial truth functioning as transition management.

The article honestly documents context decay, bad incentives, weak review, verifier attacks, infrastructure failure, and the superior value of correctness over throughput. It is not simple copium. But its conclusion performs the familiar elite maneuver: displacement is reframed as a promotion into a higher supervisory layer. The programmer becomes the orchestrator; the orchestrator is presumed to remain necessary.

That is also prestige signaling. The agent fleet, token expenditure, private codebase, and model access display ownership of the new production machinery. The text is a field report from the transition, not an argument against it.

The Verdict

Accurate local autopsy, false global reassurance. This project is a clean demonstration of P1: once correctness is formalized, human review contracts, cheaper agents scale, and software production becomes an industrial control problem. It does not by itself prove that every sector has crossed P2 or P3; decompilation is unusually suited to objective verification. But it exposes the mechanism that makes the transition lethal.

The scarce asset is no longer the ability to write code. It is temporary control of the oracle, compute, models, deployment channel, and resulting intellectual property. Those are Sovereign assets. “Orchestrator” is a viable position only while the orchestration layer remains privately controlled and technically ahead of automation. Otherwise it is merely the next job title to be decompiled.