I'm a heavy and competent user of LLMs for everyday tasks including SWE, yet still learned a <i>tonne</i> from this. Several interesting points:<p>- Simon's obviously a gun programmer yet uses LLM's heavily<p>- LLMs take a lot of practice to learn how to use. They don't come with manuals<p>- How much do they improve developer productivity? ANS (paraphrased): It makes me about 3x more efficient at typing code, but typing code is ~10% of my job (the rest is research etc). So I'm 3 times as efficient at 10% of my job.<p>- It's completely sane to be a bit scared and a bit ethically concerned<p>- Regarding tooling, Simon uses claude sonnet (as it's the best model right now), he uses a CLI he made called 'LLM' [1], he uses GitHub CoPilot [2].<p>- (a few more golden nuggets but I can't recall them all)<p>[1] <a href="https://github.com/simonw/llm">https://github.com/simonw/llm</a><p>[2] A tip he gives is because CoPilot 'sees' what's in the current file, you can copy/paste relevant code from other parts of the app into the current file just so it knows what you're talking about or trying to do. Especially useful for tests.