Post

Steve Klabnik
@steveklabnik.com

i keep thinking about this post

Fred Hebert@ferd.ca

Why do some have a shit time with LLMs for programming while others love it? To succeed, the latter group tacitly creates tons of scaffolding and gain weird new skills. While it works, this posts explains how doing all that is an incidental consequence of bad interaction design in coding AI agents.

ferd.caThe Gap Through Which We Praise the MachineMy current theory of agentic programming: people are amazing at adapting the tools they're given and totally underestimate the extent to which they do it, and the amount of skill we build doing that i...
Chris
@csteinbach.bsky.social

There are some interesting thoughts in this and I'll have to read it over to take it all in. The thing about deskilling confuses me though; we work in an industry that has automated away manual work, including our own, from the get go. It's surprising we've not come to terms with this.

Fred Hebert
@ferd.ca

Some risks around deskilling are not purely about obsoleting skills that will never be needed again, but about eroding skills that may be highly necessary but less frequently called upon—often inadvertently so due to changing how they are practiced. This is partly why pilots must use simulators.

0 reposts11 likes

Fred Hebert
@ferd.ca

A long known irony of automation is that as you automate the most frequent (and easy) cases first, people are left to handle both increasingly rare *and* challenging situations, with less practice.

Chris
@csteinbach.bsky.social

Absolutely! The deskilling is real. I mean, how many developers these days can program and debug assembler? But so is the tradeoff. And we keep choosing automation. Just seems that an old argument has been churned up that I thought we'd made peace with.

Fred Hebert
@ferd.ca

Yeah it's not all needed in the same way. I've written a much more opaque post last year to try to figure out what sort of systemic effects can come from removing friction that hone subtle skills and tradeoffs. A questions I like is "do you still like your code if you can't use AI to work on it?"

ferd.caLocal Optimizations Don't Lead to Global OptimumsOptimization and automation can be useless and reckless, outmatched by systemic pressures and their trade-offs. Localized friction, hidden by optimizations, can hold the key to larger scale impact.
Chris
@csteinbach.bsky.social

We’ve built such unwieldy development stacks (and so many of them) that LLMs become a way to survive, rather than address the dysfunction (friction). Is that a fair reading? If so I agree. I can't help thinking LLMs would still be super useful even with less dysfunction.

Alex Suzuki
@alexsuzuki.com

Interesting take. I think this also applies to self-driving cars. If the system is not 100% reliable and still requires a driver to be ready to manually intervene, but the next generation of drivers grows up with virtually zero actual practice, especially in compley situations, … 💥