The video responds to claims that software is finished because AI tools can look at a game or application and produce a working copy. The presenter says hundreds of AI-assisted reverse engineering projects have appeared, some playable in a browser, and describes two routes: reading compiled code to work out what it does, or observing behaviour and refining a new version until the differences shrink.
On legality, he says he is not a lawyer, but notes that courts have held software functionality is not necessarily protected by copyright, so independent re-implementations can be lawful. Assets such as models, textures and characters can be copyrighted.
He then turns to research. One study tested whether AI systems could rebuild whole software projects without source code. On one programme the first result passed only 15% of the tests, but within nine months it reached nearly a hundred percent, and he says that used older models. A separate piece of work used a 12 million parameter model that recovered mathematical equations with around 90% accuracy on a specially constructed test. A third reconstructed web pages from their appearance, with a visual similarity score of 0.938 while using 95% less code than the original sites.
The creator gives two reasons this works: nearly unlimited training pairs from running open source applications, and access to the original software as a reference to test against repeatedly. He notes that Anthropic has updated a model to discourage recreating existing games, and that some developers say it also refuses legitimate reverse engineering. He uses this to argue for owning compute and running local models. His conclusion is that saying software is over is dramatic, but personalised homebrew tools may become practical.
Key takeaways
- The video says AI can rebuild programmes either from compiled code or purely from observed behaviour, iterating until outputs match.
- The presenter says courts have found software functionality is not necessarily copyright-protected, while art assets and characters can be; he is not a lawyer.
- One cited study went from 15% of tests passed to nearly a hundred percent in nine months, according to the video, using older models.
- A 12 million parameter model recovered equations with around 90% accuracy on a constructed test, per the video.
- The creator argues that restrictions from frontier labs strengthen the case for local, user-owned models.
- His verdict: software is not over, but tailored personal software looks increasingly likely.
Our take
The video is a creator's summary of several papers, and we have not checked the figures against the studies. The presenter mixes results that measure different things: test pass rates, equation recovery and visual similarity. They should not be read as one benchmark. His legal comments are explicitly not legal advice, and the video carries a sponsor read for a GPU cloud service.
The argument for local models is the part most relevant to hardware buyers. Readers weighing that route may find our guides How to choose a mini PC for home AI and CPU vs GPU vs NPU for AI inference useful, since they cover memory and bandwidth constraints for running models at home.


