I subscribe to Simon Willison’s email newsletter (really just selected blog posts via email), and his last two were excellent.
Last week he talked about his casual daily use of the new Claude 3 Opus model via his own LLM tool, as well as ChatGPT GPT-4. He describes such use as “sidequests,” a description that resonates with me and my own daily LLM use. In this example, over an elapsed time of about 6 minutes, he was able to get an accurate GeoJSON for New York State’s huge Adirondack Park starting with a deeply strange shapefile.
Over the weekend, he shared a longer prototyping effort using ChatGPT GPT-4 Code Interpreter, where we managed to get GPT-4 to write, compile, and run a non-trivial bit of C code: a SQLite C extension for vector similarity. Code Interpreter runs Python, not C, but Simon suspected and confirmed that he could get GPT-4 to run gcc from within Python, and that was enough of a shim to get the job done.
Both are excellent reads that mirror my own experience in terms of the myriad ways I take advantage of LLM tools.