I hope you aren't missing the point. My position is similar to the author. I WILL take responsibility for the code I push to production, and rather than input a prompt and roll the dice on the outcome, I am strategic in my prompts, ensuring the LLM has the right context each time I I voke it, some of that context being accurate descriptions of what I want built, and I am in charge of ensuring it has been properly vetted. Many times I will erase what the LLM has written and redo it, by myself depending on the situation.
Replace "LLM" with "IDE" and re-read. The LLM is another tool. Of course tools can't be held responsible, the person wielding the tool is.
> Many times I will erase what the LLM has written and redo it, by myself depending on the situation.
The contention here is that antirez doesn't think this is necessary anymore. 100% code gen, with the occassional "stepping in and tell the AI how to write a certain function"
You're not. All you need to do is look around a typical car park and count witness marks on car bumpers to know that accidents are extremely common. One study put the estimate at one "driving mistake" every 500 miles and one crash every 60,000 miles.
An insurance policy study estimates that the average person is involved in 3–4 collisions in their lifetime.
You will be cursed with years of calling every pharmacy in town once a month to figure out which one has your medication in stock this time, and once you figure that out, you stay on the phone with them until you walk into the store to pick it up so they don't give it to someone else.
VSCode is currently more complete than Zed but Zed is more configurable. Also I suspect that Zed will improve rapidly and VSCode is mostly already plateaued due to Zed being new and open source and VSCode being old and made by Microsoft.
Replace "LLM" with "IDE" and re-read. The LLM is another tool. Of course tools can't be held responsible, the person wielding the tool is.