Post

A lot of CLI/TUI usage is based on a silly elitist "hacker" aesthetic. Humans are better at crunching data through the visual cortex. GUIs & Visualizations can unlock that for us.

shell scripting is the one area in which i'm not worried about LLMs deskilling me i've been doing this thing for like 20 years and still don't know how to use xargs or awk, i think at this point it's safe to give up on that one 😂

of course there’s the other fun interpretation of this: if an ai has all the right ways to observe and use something, it’ll necessarily beat us. it’s very clear why CLI are easy for LLMs to parse and use

Lest we not forget that the cost of this superpower is that every so often it might brick your machine or lose hours of work. The real goal of mechanistic interpretability should be identifying the Joker feature and allowing me to easily crank it to 11 on colleagues I don’t like.

That is a reason why I'd love a have a good combination of both. There is currently -> www.warp.dev But it completely breaks my habits, as I cannot use tab autocomplete with it in the way I am used to. Sadly this is a deal breaker for me after trying it for ~2 weeks. I love the concept though.

Yeah that's what I'm saying too.

There’s also reason to believe that they come up with generalised solutions to problems and can extrapolate those solutions to data they’ve never seen before. It’s not super well understood yet but there are papers out there analysing how LLMs perform modular arithmetic in a generalised way.

Agreed.

I mean to say they are only that good because humans got to above that level on individual tools and skills and documented their findings and often journey.

I wonder if "thinking in pipes" is just fits the "next token prediction" model just very naturally?

Watching them go has led me to improve my own usage which is nice

I assure you that I have seen people good at using the terminal.