I’m still skeptical that the LLMs are really doing “clean room decompilation.” It’s plausible that all of the source code for these games is included in the training data. Most of the examples in the article are owned by big publishers who may have made secret source code access deals with the AI companies. In the case of The Binding of Isaac’s flash version, the LLM might be trained on the pre-existing decompiled ActionScript.
Nah, LLMs can absolutely do decompilation using tools like Ghidra. I used LLMs to reverse engineer the proprietary codec for my Nikon camera RAW format, and it worked great.
Sure, maybe proprietary code from every company is in the training data magically. But it’s obvious that the LLM wasn’t just regurgitating code that it memorized because it solved the problem iteratively over many days. And LLMs don’t work by memorizing things like a database in the first place.
I’m still skeptical that the LLMs are really doing “clean room decompilation.” It’s plausible that all of the source code for these games is included in the training data. Most of the examples in the article are owned by big publishers who may have made secret source code access deals with the AI companies. In the case of The Binding of Isaac’s flash version, the LLM might be trained on the pre-existing decompiled ActionScript.
Nah, LLMs can absolutely do decompilation using tools like Ghidra. I used LLMs to reverse engineer the proprietary codec for my Nikon camera RAW format, and it worked great.
I mean, Nikon is another big software company. I don’t see why their code wouldn’t end up in the training data either.
Sure, maybe proprietary code from every company is in the training data magically. But it’s obvious that the LLM wasn’t just regurgitating code that it memorized because it solved the problem iteratively over many days. And LLMs don’t work by memorizing things like a database in the first place.
Overtraining is very likely, and the AI might even be able to hide this from researchers.