Congratulations to Dr Luke Dramko (my lucky 13th PhD graduate!) for his successful and stellar defense yesterday of “Neural Decompilation with Minimal Risk.” The last chapter isn’t published yet but STAY TUNED because it’s a symbolic approach for proving the fidelity of neurally-decompiled code. 1/

All these hot takes on the leaked Claude Code codebase. Listen. All code looks terrible if you’re just looking at it, vibed or not. The real question I, paying customer, have is totally unrelated to the specifics of Anthropic’s codebase and is instead this: wtf, you guys. What the shit is this.

First snow of the year, at the perfect temperature, and CMU students responded by building DOZENS of snowmen of all sizes all across campus. It was delightful. The (college) kids are alright. ☃️ #proflife (These were done by my kids, who are also delightful.)

We have other recent results on LLM-based security vulnerability detection that includes similarly alarming indications of leakage on BigVul, even though that’s not the main point of the study…(s/o to the student, Aidan, not On Here.) arxiv.org/abs/2406.05892 So…yeah…we may have a problem. 😅 /fin

📈 Llama 3.1 (70B)—trained on far more data—shows less leakage than older and smaller models like CodeGen and CodeLlama. Its higher NLL and lower 5-gram match show limited signs of leakage. 5/

📜 5-gram match reveals memorization: We used 5-gram match to check if models generated nearly identical outputs when given the same input. CodeGen scored 82% on Defects4J. There’s no established baseline or cutoff here, but that’s eyebrow-raisingly high. 4/

🧠 Older models memorize more: Models like CodeGen, and CodeLlama show significantly higher leakage on Defects4J than newer models (e.g., Llama 3.1). They often reproduce patches verbatim, to the point that it’s weird (including comments!!) 🔥 3/