Post #2527323

I recently built "seminar" as somewhat of an experiment (https://github.com/hgrsd/seminar). I wanted to see if letting LLMs, with access to tools like web search, PDF reading, etc, could build a "body of studies" based on a given, fairly rough and ambiguous idea that I gave them. The idea being that those studies could build on top of each other, contradict each other, and end up with a rounded picture and new insights. I also gave the worker processes the option of starting a new conversation thread with me for things it "wanted" to call my attention to, or suggestions for things I should read myself. One design goal was to build something AI-based that rather than stop you from thinking, tries to make you think more. I think that it has succeeded in a fairly narrow sense: I've gained lots of interesting things to read as a result of these sourced studies. But I have not seen anything novel or interesting. In fact, the capability it has to come up with new research ideas based on "synthesising" studies from existing ideas generally leads to totally outlandish ideas that make very little sense, and which are presented extremely pompously as though they are worthy of a nobel prize - it's kind of funny. Will I be using this seriously? Probably not. But it's definitely helped me understand the limits of LLMs in generating anything new, as well as their usefulness as a starting-off point for your own thinking. #ai #research #llm #tech