Four hours of human attention rebuilt a 1989 war game for the web

One person, one week, one 1989 strategy game

A developer writing at this.os.isfine.org has published a work-in-progress account of reverse engineering War of the Lance (1989) and porting it to modern browsers using an AI "code harness" — an automated loop that runs models against a codebase. By his account the whole thing took about a week of wall-clock time and less than four hours of his own dedicated attention. He says he never opened an IDE, Blender, a reverse engineering tool, a command line or source control. He does not know what the code looks like. He knows it passes the linting and test requirements the harness enforced on every run.

This is one person's blog post about his own hobby project, not an audited study. Treat every number in it as a claim. But the claims are specific enough to be worth taking apart, because the mechanism he describes is the part that generalises.

What the harness actually does

Reverse engineering an old DOS game means recovering rules and rendering behaviour from a compiled binary — the source code is gone, so you infer what the program does by watching it run. Historically this needed a rare skill stack: disassembly, graphics formats, memory layout, plus the engineering to rebuild it somewhere new. The author estimates that combination existed in maybe a few thousand people across a 320,000-strong professional games workforce, and that a single title took months or years.

The harness he describes collapses several of those roles at once. It plays the original game inside DOSBox-X, reading memory dumps and disassembly to work out what the rules are. It plays its own new version through Playwright — browser automation normally used for testing — inspecting the page structure and memory, and using vision models to look at the screen. It writes tests as it goes, so later changes cannot silently break earlier work. For 3D assets, the model writes Python that drives Blender in headless mode to output GLB files, the standard format for web 3D. He reports modelling roughly 80 to 100 models on a week's subscription, and that adding seasons, day/night and three moon phases cost 15 minutes and $2.28.

The jargon worth keeping: a lane is one parallel worker. He could only run one locally because of VRAM limits. His stated trend line is that a local model equivalent to an April or May frontier model takes a week; dispatching cloud lanes drops it to two days; a September 2026 frontier model does it in a few hours. He predicts minutes within 18 to 36 months. The earlier steps are things he says he did. The last one is a forecast.

The signal is not the game

His sharper argument is about pricing. The knowledge economy pays for knowledge that is expensive to acquire and reliably applied — his example is a doctor's fifteen years versus a grocery store manager. If a machine becomes a competent storage-and-application medium for that knowledge, with the cost advantages of centralised scale, the scarcity premium erodes regardless of whether anything resembling general intelligence ever arrives.

He is notably unimpressed by benchmark theatre: he argues OpenAI's GPT6 Astra was trained hard on one-shot Three.js scenes precisely because the internet benchmarks on them, while coding improved far less. He also names a failure — Opus 5, which he says simply does not converge on long tasks. The useful capability, in his telling, is not benchmark score but the ability to keep going coherently for hours or days. He expects the broader economic effect to take a decade or more, because real jobs are messier than demos and replacing optimised systems rarely pays.

Questions You Should Be Asking

• If nobody on your team has read the code, what is your plan when it breaks in a way the tests never covered?
• Who is legally liable for output that was reconstructed from someone else's artefact — and has your counsel actually answered that, or just noted that enforcement looks weak right now?
• When a vendor cites a benchmark gain, is it measuring the capability you need, or the capability the internet scores them on?
• How much of your pricing power rests on knowledge being hard to acquire rather than hard to apply?
• Can the model you are buying stay coherent over a multi-day task, and how would you find out before committing?

What To Watch Next

Watch whether verified, repeatable reverse engineering results start appearing from people who are not also making the claim. A single hobbyist's week is anecdote; the same job reproduced at a few hours by people with no stake in the narrative is a labour market signal. And watch the copyright cases — the author's blunt read is that the current political environment will not stop any of this, and that Meta's outcome should end anyone's hope that the old rules hold.