Loading…

At a developer conference in Austin I saw two AI code generators go head to head and it split the room | AI Innovations - Talk Freely

kim373•5mo ago

At a developer conference in Austin I saw two AI code generators go head to head and it split the room

One team showed a tool that writes your whole function from a comment, the other demoed one that just auto-completes your current line. I'm leaning toward the comment-to-code side since it saved me 10 minutes on a messy SQL query yesterday. Which approach do you think actually makes you faster in the long run?

3 comments

3 Comments

the_lucas5mo ago

Wait but did the comment-to-code tool actually handle edge cases right or did you still have to go back and tweak stuff? That's what gets me, cause auto-complete feels safer for logic you already understand.

henry3155mo ago

Honestly I used to feel the same way but then I actually tried one of these tools and it surprised me. The edge cases were handled better than I expected because it pulls from real code patterns in the project, not just random guesses. I still had to fix a couple things but it was like 2 or 3 minor tweaks instead of a whole rewrite. Now I kind of see it as a way to get the boring parts done faster so I can focus on the tricky stuff myself.

james_kim4mo ago

Yeah but see I think you're giving the tool a little too much credit on the "real code patterns" thing. It's not really pulling from the project in the way you're describing - it's more like it's trained on a huge pile of code from everywhere, and it gets lucky sometimes with the specific patterns in your repo. I tried one of those comment-to-code things on a project that uses a weird internal framework and it straight up hallucinated half the function calls. It's fine for boilerplate and common stuff like for loops or basic CRUD, but the minute you get into your own weird special case logic it falls apart. You still gotta do the thinking, it just writes the first draft faster.