Jamie Brandon's Speed Matters post has aged really well. It's also pretty much what @simon is observing about LLMs
-
Jamie Brandon's Speed Matters post has aged really well. It's also pretty much what @simon is observing about LLMs
*ducks*
-
I think what a lot of people miss is that if you have a commit-sized piece of work in mind, there's a very good chance that you can get an LLM to do it in under a minute
you don't have to be an AI True Believer to see how this changes what an experienced programmer can do!
-
Joseph A di Paolantonioreplied to Reilly Wood last edited by
@reillywood yes, an experienced, a very experienced programmer, and I’m seeing some that I very much respect making amazing use of various LLM/GPT platforms. The hype is that the inexperienced and non-programmer can do the same. That’s not true. And I see a few that I respect make this claim—about subject matter experts. This may be true if they fully examine the results, yet they won’t see flaws in the program that gives good results, such as efficiency, memory and security issues @simon
-
Simon Willisonreplied to Joseph A di Paolantonio last edited by
@jadp @reillywood yeah, the idea that this means non-programmers can ship production code without learning to program is misleading hype
But the performance boost you can get if you DO know how to program is very real
I also suspect that learning to program has become massively more productive, but I’ve been programming for so long that I can’t speak for newcomers!
-
happyborgreplied to Joseph A di Paolantonio last edited by
@jadp
As a very experienced programmer I'm not convinced. I've only dabbled but the things I've found LLMs better than a web search for are very few and trivial.They are just not worth my time. I have identified one regular use case and that's creating bash commands I don't remember the syntax for but can check in a flash.
Even for that I find the delay and the 'no explanation just the command' tweak tiresome.
NB regarding speed, I value my privacy so run them locally.
@reillywood @simon -
@jadp
I want to add that even for this bash example it's better not to use LLMs unless they are one-off tasks, because it's better for me to learn or remind myself of the syntax if I'm going to be doing this again soon.And you don't learn by just checking something, you must compose - which of course applies to any programming task, not just bash.
@reillywood @simon -
@happyborg @jadp @reillywood I’ve been finding them incredibly useful for all manner of programming tasks - but I’m mainly using the hosted ones (Claude 3.5 Sonnet and GPT-4o these days) which are a lot more capable than the local models
Notes on how I use them are collected here: https://simonwillison.net/tags/ai-assisted-programming/
-
@simon this may be so, and of course 'they' want us to use hosted services rather than local.
I have a problem with that which goes beyond the issue relating to LLMs, although those issues are also toxic to humanity.
-
@happyborg @jadp @reillywood an open question to me is how the energy usage of many people running their own local LLMs would compare to the energy usage of many people sharing access to much more power-hungry hosted LLMs