Imagine being given a cake and asked to work out the exact recipe. Not an approximate recipe: the recipe down to the gram, the one that, followed to the letter, produces exactly the same cake.
That is, broadly, what video game enthusiasts have been doing for years. And it is what an AI can now do in their place.
Rebuilding a game, line by line
A video game sold in shops does not contain its source code, the text written by its developers. It contains only a version translated for the machine, unreadable to a human. "Decompilation" means working backwards: rewriting readable source code which, once translated, produces exactly the same program, byte for byte.
Why go to so much trouble? To preserve old games, understand them, get them running on modern machines, modify them. Painstaking work, which often took teams of volunteers with complementary skills years to complete.
In 2019, after several years of work, a team of volunteers completed the full decompilation of Super Mario 64, the famous 1996 game. A few months later, unofficial versions for computer appeared, with modernised graphics and features never seen on the original console. It is one of the most famous feats of this community, which nonetheless works in a legal grey area, since the games remain the property of their publishers.
What has just happened
Since 2024, the site decomp.dev had served as a rallying point for this community. It tracked hundreds of projects, each with a percentage of completion. Every point gained represented weeks of human effort.
On 15 October, its administrator, known by the pseudonym Encounter, closed the public index and stopped accepting new projects. His reason comes down to one sentence: the collective aspect of decompilation has been broken.
According to him, recent AI models, such as OpenAI's GPT-6 Astra, can now follow complex instructions, cross-reference clues in the code and write a natural program. What used to take years is done in a few days. As a result, it is becoming impossible to tell projects run by humans from those done by AIs. The percentages, which rewarded months of effort, have in his view become a game of hollow numbers, inflated by people running AIs to watch the counter climb.
Not an indictment of AI
The administrator is careful to point this out: he is not banning AI, which can help enormously when used well. The problem is not the tool, but what it has done to the community.
The episode illustrates a skill at which AI excels: rebuilding what already exists. It is the opposite of what we showed yesterday, with the AIs that cannot yet invent. Working out the recipe for a cake, yes. Inventing a new cake is another matter.
The phenomenon goes beyond video games: copies of popular software, made by AI, have also been circulating in recent days.
What to take away
Communities often form around a shared difficulty. People come together because it is hard, and because they need each other to get there.
When AI removes the difficulty, it sometimes also removes what brought people together. Which raises a broader question, the subject of our reflection this evening: what is the point of searching, if AI finds everything?