Jump to navigation

Nikola Plejić

Antirez on LLMs

2024-03-27 in link, longread

Link: LLMs and Programming in the first days of 2024

Salvatore Sanfilippo is certainly a programmer I look up to: I’ve appreciated both his software and his writing for a long time. This article that’s been in my reading queue for a while is a sober take on the topic, and I found it insightful to see how he incorporates LLMs into his workflow.

What he calls “junk knowledge” — obscure APIs, convoluted protocols — seems like a good use case for these tools. I’ve tried to steer my usage in this direction, too: I’ve recently had to work with AWS a fair amount, and ChatGPT ended up being genuinely useful. While the “CloudWatch Logs Insights query language” is handy, and I’m sure someone had a lot of fun implementing it, I’ll gladly opt out of learning even more syntax I’ll rarely use.

All of that said, it’s important to continuously remind ourselves of the cost of these tools — in watts of power, in [insert unit]s of privacy — particularly by our thought leaders. I haven’t seen much of that in this article, but Simon Willison’s recent writing on “ad-hoc sidequests” does a good job in its closing words. I’ve been enjoying his other writing on this topic as well.