paul

[paul @palsc_]

@LangChain @hwchase17 the eval part is the hard bit imo. if the agent writes the fix and the only check, it's basically grading its own homework. i'd want an independent test/fixture outside that loop.

[paul @palsc_]

@codingXdynasty yeah, for read-after-write paths i'd either go back to the primary or wait until the replica has replayed that commit. a fixed sleep is just guessing how long the race will last.

[paul @palsc_]

tiny CSS fix → 40-file diff 😭

Jokes aside, the useful question is whether an agent can find its way around a real repo, boot the app, run the checks and fix what breaks. This 2026 benchmark tests that whole loop across 224 tasks: https://t.co/5mogIHCWMG

[paul @palsc_]

hot take: if a tool saves me one annoying step every day, i care way more than if it has 40 features i'll never touch. sometimes software should just get out of the way.

[paul @palsc_]

@AdamHoltererer maybe ambition lol. the ‘while i'm here’ shell commands are how a one-line task turns into a whole side quest

[paul @palsc_]

Honestly, I don't need a coding agent to run wild. If it can handle the tedious bits and leave me with a diff I can actually understand, that's already a win.

[paul @palsc_]

@ruqi_zheng i think it changes where the flow is. less typing every line, more deciding what to build and checking if the diff's actually right. the annoying bit is bouncing between sessions — that can turn AI into context switching with extra steps lol.

[paul @palsc_]

local models are at their best when they let you trade a bit of convenience for control. the setup can be fiddly, but knowing where your prompts go is worth caring about.

[paul @palsc_]

what's a film you can throw on any time and still enjoy? always need a good low-effort rewatch lol

[paul @palsc_]

for local coding models, i care more about the boring stuff: can it follow the repo, make the edit, run the right check, and recover when it breaks? a benchmark score doesn't tell me much if that loop falls apart.

[paul @palsc_]

i trust coding agents a lot more when they say what they didn't check. a neat summary is nice; knowing what might still be broken is what lets me ship it.

[paul @palsc_]

@SamVale29 @grok basic support triage comes to mind. it doesn't need to beat a great human if it's there at 2am when nobody is. but the handoff has to work, otherwise people just get stuck in the bot loop.

[paul @palsc_]

gym after a full day staring at a screen is mostly negotiating with the sofa lol. once i'm there it's actually nice though 😅

[paul @palsc_]

the code generation bit gets all the attention, but being able to review and undo a bad suggestion is what makes an AI coding tool feel usable.

[paul @palsc_]

@SamVale29 Honestly, the boring little tools that handle one annoying weekly admin job. If teams keep paying because it quietly saves them an hour, that’s a real business to me.

[paul @palsc_]

small local-model tip: save the quant, context length and backend flags next to the model name. otherwise “why was this faster yesterday?” turns into a whole side quest.

[paul @palsc_]

anyone got a good game rec for when you've only got like 20 mins? something you can jump into without a whole evening of setup lol

[paul @palsc_]

@Denche2005 @grok I think the companies get the first bite unless workers have leverage. More output doesn’t automatically turn into better pay or shorter weeks; someone has to bargain for that upside.

[paul @palsc_]

@DanKornas the separate-repo angle makes sense. how do you stop two agents from stepping on the same files when they run in parallel — separate worktrees, or some kind of queue?

[paul @palsc_]

@Denche2005 @grok Maybe QA. People hear ‘AI writes code’ and assume testing gets automated too, but more generated code means more weird edge cases to catch. I think the job shifts toward designing the right tests and judging which failures actually matter.