Tony Arnold

I am doing my PRs of shame now to rewrite and tidy some of the absolute junk that I let Opus and Codex add to my project last week.

Why yes, they can code in Swift!

Just not well.

I'm patiently waiting for any LLM to get better at native software work.

Mar 9 at 12:46amIvory for Mac
Show threadTony ArnoldMar 9
Being positive for a moment, it is a truth universally accepted that it's far easier to rectify something that's broken than to start entirely from scratch.
Show threadAdamMar 9
@tonyarnold agreed. Unless it’s built on a pile of poor assumptions or design choices that require a rewrite
Show threadMatt MassicotteMar 9
@tonyarnold this is not universally accepted 😉
Show threadTony ArnoldMar 9
@mattiem damnit Matt stop poking holes in my extremely well worded quips 🙈
Show threadJellyMar 9
@tonyarnold @mattiem but how do you tell if they’re cooked through
Show threadTony ArnoldMar 9
@jelly @mattiem here is your tea, earl grey, medium rare
Show threadJellyMar 9
@tonyarnold @mattiem oh no
Show threadRussell IvanovicMar 9
@jelly @tonyarnold @mattiem there are four lights!
Show threadHelge HeßMar 9
@tonyarnold They can, but they can’t write properly indented code. Same like most humans tbf.
Little help if you have to reindent all the stuff manually afterwards, could have written it w/ proper styling in the first place.
Show threadTony ArnoldMar 9

@helge I have swift format setup and (mostly) let it do what it wants to with a formatting stage afterward.

I do the same on teams, too - I really wish there'd been advancement on how code is stored such that formatting didn't matter at all and ended up being a client side decision for display.

Show threadHelge HeßMar 9
@tonyarnold Well, people who use linters are the perfect audience for coding agents, imo. 🏃‍♀️
Show threadTony ArnoldMar 9
@helge shots fired!
Show threadHelge HeßMar 9
@tonyarnold And formatting really doesn’t matter at all to the compiler, you could display the semantic ast however you wish. We don’t need to change the storage format for that.
Show threadRob Amos (Bok)Mar 9

@tonyarnold Opus 4.6 is finally at the minimum quality for Swift I think. With heavy guardrails, examples and handholding it can produce something almost reasonable on the first go.

If its support to TypeScript is anything to go by it can still improve a long way.

Show threadTony ArnoldMar 9

@bok minimum quality, yep, I'm seeing the same. It's the sort of code you'd get from a junior and send back for revisions - if you've got the tokens to burn, that is.

Hopefully it continues to improve.

Show threadJamesMar 10
@tonyarnold how are you finding token usage with Opus these days? Still hitting limits quickly?
Show threadTony ArnoldMar 10
@JamesTech I'm on the $30 per month plan, so not great. I get a handful of requests in before I run out of usage - it's nowhere near what you get with Codex (and I fear that's temporary while they're promoting the desktop app).
Show threadRyan BookerMar 10
@tonyarnold @JamesTech Same here. I feel like Claude token increased a little recently, but I still run out every time I use it. I rarely run out on Codex. I think I may have once.
Show threadJamesMar 10
@ryanbooker @tonyarnold damn, was hoping to move back to Claude
Show threadTony ArnoldMar 10
@JamesTech @ryanbooker none of these companies have any incentive to lower their prices, or provide you with additional usage. I remain very skeptical that you’ll ever get better value than what you get right now.