4 ms·
Very cool. I wonder if this kind of work is something that LLMs will be good at someday. (Has anyone trained an LLM on machine code yet?) Is it reasonable to h
by default-kramer 3y ago
Very cool.
I wonder if this kind of work is something that LLMs will be good at someday. (Has anyone trained an LLM on machine code yet?) Is it reasonable to hope that 20 years from now you'll be able to say "Hey CodeGPT, this binary doesn't run on Windows 17, can you make a patch?" The deep understanding and creative thinking that Nathan Baggs demonstrates here seems far beyond the capabilities of any tool I'm aware of; on the other hand, even a relatively dumb AI supplemented by brute force might be good enough.
- 95014_refugee 3y agoYou need to be able to explain the goal sufficiently well, and then you need something goal-seeky to do the work, i.e. a closed-loop system. LLMs are impersonators. They take a prompt and then make up something that relates to the prompt based on their corpus - they are open-loop systems in that sense.
- default-kramer 3y agoI guess I was assuming the LLM would be wrapped into some larger goal-seeky thing. Explaining the goal seems like the easy part; it could be as simple as "Make it possible for execution to reach this function."