Email or username:

Password:

Forgot your password?
Simon Willison

A good mental model for generative AI programming assistance might be to think of it as an automated pair programmer. You're still responsible for the code you are producing, but your automated pair programmer can chip in with suggestions, look things up for you or even type out full blocks of code based on your prompts

6 comments
William Pietri

@simon Huh. Do you do much pair programming? Either my experience of pairing or my LLM experience is very different than yours, and I'm trying to figure out which.

Simon Willison

@williampietri I haven’t in a few years, but when I did one of the things I most appreciated was that the non-typing partner could be looking things up in reference docs or suggestion neat idioms for solving problems - LLMs are great at that kind of thing

Rua Haszard

@simon pair programming with a teenager interrupting your thoughts with inane comments? But also helping occasionally. Just those autocompleted comments really break my flow.

Danny Whitt

@simon For me, it’s been like having an over-anxious friend trying as hard as possible to be helpful.
Helpful, yes, but upon examination their answers are invariably in need of refinement.

Alex Hudson

@simon I always understood the one driving the keyboard wasn't the one coding: the coder had to verbalise what they wanted, but not having to type would allow them to think more.

Which seems the other way around to an LLM right now. But it does make me wonder whether a fast enough speech processor and enough IDE context could allow you lead an LLM like that...

Simon Willison

@alexhudson I use LLMs like that quite often: I’ll describe the code I want written through typing or via voice (I write quite a lot of code while out walking the dog, speaking to ChatGPT through an AirPod) and the LLM churns out Python or JavaScript for me

Here’s a real example from a walk a few months ago: chat.openai.com/share/77996768

More notes on that here: simonwillison.net/2024/Mar/23/

@alexhudson I use LLMs like that quite often: I’ll describe the code I want written through typing or via voice (I write quite a lot of code while out walking the dog, speaking to ChatGPT through an AirPod) and the LLM churns out Python or JavaScript for me

Here’s a real example from a walk a few months ago: chat.openai.com/share/77996768

Go Up