500B tokens later: Letting AI agents decompile a first-person shooter

Original Article ↗ Hacker News Discussion ↗

AI agents are now capable of decompiling complex games like Call of Duty: Modern Warfare 2 into near byte-identical source code, but doing so can consume hundreds of billions of tokens and substantial compute, raising questions about cost, energy use, and practicality. Commenters debate whether exact binary-level reproduction is necessary versus “functionally equivalent” reimplementations, especially for preserving bugs that matter to multiplayer compatibility and speedrunners. The thread also delves into legal and economic implications: how copyright law applies to AI-assisted reverse engineering, whether aggressive subsidized pricing for AI APIs is sustainable, and what it means for the future of proprietary software and SaaS when LLMs can cheaply clone existing products.

Project scope & cost

  • Thread centers on an AI-agent-driven decompilation of a commercial FPS, widely inferred to be the 2009 Modern Warfare 2, targeting near byte-identical reassembly.
  • Reported token usage (~500B, many via a cheaper model) sparks debate over cost, feasibility, and whether the project outcome is worth it given legal constraints and no public release.

AI economics & subscription models

  • Several comments argue current API and subscription prices are heavily subsidized, likened to Uber-style growth tactics or failed “all-you-can-eat” models like MoviePass.
  • Others counter that AI companies may cover marginal inference costs but not ongoing R&D, so “profitability” depends on how training is accounted for.
  • There is confusion and correction around token pricing math; many note caching and cheap models significantly lower effective cost.

Technical approaches to decompilation

  • Two main goals are contrasted:
    • Byte-identical output: ensures full bug replication and deterministic behavior, crucial for speedruns, multiplayer compatibility, and “preservation.”
    • Functional equivalence: rewrite functions to behave the same under tests, using harnesses that compare memory accesses and coverage, sometimes aided by SMT solvers and fuzzing.
  • Some argue byte-perfect matching is overkill and expensive; others say it’s the only way to guarantee equivalence, especially for edge-case bugs and global state.
  • Discussion of model choice: smarter, more expensive models may finish faster and cheaper overall than swarms of weaker models; experiences differ.

Legal and IP concerns

  • Several comments argue such reverse-engineered ports likely infringe copyright under US doctrines protecting program structure and organization, referencing historic cases and the Google–Oracle API dispute.
  • Clean-room and “cleanroom-as-a-service” ideas are mentioned, but their robustness and legal standing are seen as murky.
  • Some highlight DMCA takedowns already hitting similar projects and note corporate enforcement risks.

Software, SaaS, and “vibe copying”

  • Many speculate that powerful LLMs and agents make cloning of binaries, SaaS UIs, and even algorithms much easier, undermining traditional proprietary software.
  • Proposed responses:
    • More software moving fully to SaaS/cloud to hide implementation.
    • A future of bespoke, AI-generated tools for individuals instead of shrink‑wrapped apps.
    • Recognition that even cloud services can be probed and “distilled” into specs by agents.

Energy, ethics, and broader lessons

  • Some criticize the project’s carbon footprint for a non-shareable hobby outcome; others respond that the solution is cleaner energy, not throttling experimentation.
  • For “normal” development, commenters emphasize:
    • Strong harnesses and tests as oracles are essential when using agents.
    • Agentic workflows excel at one-off, accuracy-focused tasks but don’t replace human architectural judgment or code review.