|
One of the tricks of using AI effectively is getting a good feel / mental model of it. And that mental model is in fact extremely weird, because you have this thing, which quacks like a human, has an immense (superhuman) amount of knowledge, and can read/write very quickly (tokens / s), but at the same time, logic wise, is often extremely dumb, and is happy to bullshit with full confidence. What I've learned is that it's good primarily for tasks like the following: - Tasks which take time to do, but are then easy to verify. - Tasks which effectively boil down to translating something from one format to another. Which might e.g. be "read this technical document and implement it in code, as for style, look at these sample code files as a reference”. - Tasks which are about exploring unknown unknowns. E.g. I write down a design, and then I ask the AI to roast it. The point is not that all the points it'll make are good and I need to please the AI, it's that out of 20 points it will list, 2-3 might both make sense, and haven't been thought of by myself. Finally, AI requires good writing skills, and asking questions in an unbiased way, otherwise the AI will gladly hallucinate to reinforce your bias. Logic exercises which are easy to verify are a moderately good fit for "reasoning models" which will go through many iterations of an LLM and basically write out the whole reasoning process. In practice though, this can be very expensive to get good results with. |
Your takeaways are good and fit that model