“Is AI a Silver Bullet? — Ian Cooper - Staccato Signals”

https://ian-cooper.writeas.com/is-ai-a-silver-bullet

> But the flaw in prompts as a description of the code to be generated tends to be that natural language is a remarkably poor medium for expressing the model that we want to use in our software. Every time we move away from a 3GL to create a more natural interface, we rapidly get diminishing ability to author software.

This entire essay is excellent

I will note that I’m much more sceptical than the author is of copilots. Most of the studies I’ve seen point to a very high error rate (security flaws, bugs, hallucinating APIs) in copilots, so the question is whether dev self-reports are accurate (studies lag reality) or not (devs are falling for confirmation and automation biases).

I’m firmly in the not camp as software dev has a long long history of massive self-delusion and are composed of exactly the kind of demographic that are likely to fall for these kinds of biases: educated and supremely confident about their own intelligence. There’s also a lot of money in LLMs, which means there’s a lot of pressure to play it up

Software is also largely managed by outright sociopaths

So, I generally don’t trust anecdotal evidence coming out of the software industry.

Whether LLMs work also Does Not Matter. The environmental impact of both training and using them is currently such that it really does not fucking matter whether copilots are useful or not. You shouldn’t be using them, just like you shouldn’t be driving an SUV, flying frequently, or going on cruises. If you do one of these things out of dire necessity, you need to be offsetting that with behavioural changes somewhere else

If that changes in the future, then sure, we can reassess copilots