LLM programming with annealing

LLMs can generate images and videos from random noise using simulated annealing techniques. For an image, you start off with a noisy image, and the process sees faces and shapes in the noise and builds on them. This can be directed using a text prompt that rewards features that fit the description in some way. This can also be directed by instructions to generate the image in a certain style. The annealing generates the large scale layout and then progressively adds the details. This can give you hands with six fingers and stuff like that.

Suppose you asked a LLM to generate a C++ application to do something vague, such as "save time when writing". The compiler would require legal C++. As well as requiring conformity with the C++ syntax, we can ask the LLM to provide 'good code' with class abstraction, code re-use, comments, and other qualities of good code. It seems possible that annealing may be able to give the large-scale structure of the program, and then anneal to fill in the details.

We may have to start off with more specific instructions if we hope to get anything that works at all, such as 'write a card patience game that runs on xterm using unicode and ANSI'. But once we have something that does that, we can try it on something vaguer.

We may have to prevent or dissuade it from re-using other code resources. Claude, when asked to compile a terminal program that worked on a Mac using ANSI-art for Windows produced something that generated sensible output but wrote its own terminal program to handle the graphics, enclosed in a Python wrapper, packaged as a Flash application. It worked. It had even improved the look of the output in a very non-obvious way. There were a few tweaks but only minor ones. Claude found a library that could render Unicode text with ANSI interpretation, and another that handled the mouse commands (both done in 'xterm' in the original), and it had a pipeline that made Mac jobs run on Windows. That may be obvious to Claude, but it was weird.