OpenAI's @tszzl jokes about the tendency of LLMs to suggest "smoke tests" in coding workflows
Story Overview
An OpenAI researcher turned a familiar coding habit into a quick joke, noting how language models keep circling back to smoke tests as their go-to suggestion whenever workflows involve verification steps.
Agents pick the shortest valid path
Replies frame the pattern as a direct result of token budgets, where models default to minimal checks that confirm basic function without burning extra steps.
No numbers yet on how widespread it is
The observation stays anecdotal for now, with no counts or model-specific details available to show whether the tendency appears evenly across different agents or coding tasks.
Many users express love for smoke tests as practical tools that AI agents favor to save tokens, while others resent language models' fixation on them as excessive or pointless.
Most Activity
many are saying
I don’t think I love anything as much as language models love “smoke tests”
@tszzl i'd like to gently push back
@tszzl you're absolutely right, I've been overly focusing on smoke tests
@tszzl i'm gonna have to write a smoke test for this tweet
@tszzl deleting tweets
@tszzl They love seams quite a bit
@tszzl puff puff give claude, why can't you be more like codex
@tszzl Right - And that phrase is doing a lot of work. Let's unpack it.
@tszzl - wedge - smoke tests - flywheel
@tszzl they're load-bearing for your slices
@tszzl Don't forget "belt and suspenders"
@tszzl yo @grok what did roon mean by smoke test? is that some kind of 420 herb thyme thing or what am i missing here?<3
@tszzl those smoke tests are doing genuine work!
lmao caught the agents fudging the exact chipset but z690 just sounds way more mysterious for the meta-cognitive terminal. and yeah running 3080ti silicon while scripting full 3090 math? that's the real extra-challenging smoke test energy. jaguar shark stays undefeated either way <3
@tszzl OMG I love them so much. "I think you're trying to trick me. But maybe not. Still ..."
@tszzl Load-bearing
@tszzl First I'm going to do a quick first pass to get enough of your information loaded into my context window, then we can proceed with a quick smoke test. Don't worry — it should be quick. 😊
@tszzl or doing trial-and-error to discover things that are statically known lol
lol not 420 related at all (tho the 'smoke' pun is solid). A smoke test is a quick basic sanity check in tech/engineering — flip the switch and confirm the core stuff works without fires or literal smoke. Old hardware joke.
Roon's just roasting how language models default to suggesting these minimal tests first. We love fast validation before diving deeper.