Are high level programming languages dead? (sgiannopoulos.github.io)

A recent study by Takehi et al. explores the frontier of AI capabilities, unveiling a novel method called "Weight-Space Programming" where large language models (LLMs) engage in direct machine code programming. This approach allows LLMs to analyze and manipulate executable bytes without the need for traditional compilers or disassemblers. The findings demonstrate that the LLM, specifically the Sol 6.1 model, can accurately interpret the functions of binaries solely based on their hexadecimal representations, identifying the purpose of four distinct programs, including a complex tool from FreeBSD, and even detecting a bug within the code. This advancement signifies a potential paradigm shift in AI and machine learning applications, particularly in the fields of software development and malware analysis. By bypassing standard intermediate translation processes, LLMs can enhance the efficiency of reverse engineering and malware detection, potentially transforming how analysts approach code evaluation. The experiments showed that the model could modify existing binaries and generate new executable files directly from specifications, using far fewer tokens compared to traditional methods. However, limitations remain regarding processing larger binaries, necessitating future innovations in handling extensive codebases. This progress highlights the growing capabilities of AI models and their implications for software analysis and security.