Build stuff. There's a lot of untested stuff. There's a lot of research that goes nowhere, or things like Bard, which are really good, but will end up like Wind

Build stuff. There's a lot of untested stuff. There's a lot of research that goes nowhere, or things like Bard, which are really good, but will end up like Windows Mobile. When you build stuff, it's clearer what the strengths and weaknesses are, and which innovations matter. There was that hype with AutoGPT and similar iterative GPTs. But when AutoGPT is tested hard, it ends up getting confused on what to do, and the price tag ends up too much. It was a nice attempt, but didn't go too far. Langchain was extremely good at first, but then the new GPT-3.5 API chat format just allowed you to include "memory" without needing a third party to track all of it. If you follow most of the AI thought leaders, people are still talking about prompt engineering, but that's becoming more unnecessary with something like GPT-4, which can read your mind better than your spouse. A lot of people with experience are left behind because they overinvest in these things that are just transcended months later.