tyler
18.8K posts


@VicVijayakumar Some of my favorite LLM experiences have been diagnosing issues like this.
English

We had a sev recently which was actually kind of interesting. This particular code had been running in prod for months, but the test suite was only a couple days old.
One morning, one of the pods on our control plane cluster on staging OOM-crashed and then very rapidly the entire cluster went down and experienced a crash loop backoff.
That team got paged, and they VERY rapidly triaged the issue, paged my team to say your synthetic tests running every 5 mins are killing staging. I turned them off immediately, the cluster came back up, and before we could figure out wtf happened the oncall engineer handed me a markdown file with a full Claude analysis of the issue, made me the incident owner, and peaced out.
The gist of it was that our delete endpoint was improperly cleaning up 1:many relationships. The associated records were being deleted but the actual join relationship wasn't.
example: [user] -----> [user_address] -----> [address]
^we deleted all the addresses, but not the user_address relationship, so there were a lot of user_address records pointing at deleted addresses.
We were using the presence of the join record to query for every associated record and making an rpc call inside a .map(), all of which were resulting in 404s. Because of the orphaned records we had a 1:1000+ situation that was rapidly growing, and we were making rpc calls to all of them in parallel and holding the http requests in memory. This code was in prod but real customers don’t CRUD every 5 minutes unlike an automated test.
All of this is fine, pretty stupid bug and easy to fix. What impressed me was that this engineer cross-referenced pod logs to the bad code and handed us the root cause analysis that quickly. And there was no attempt to pretend that this was AI analysis.
English

@iamdavidhill we keep hearing about sub-agents but never about dom-agents
English

@GrantHorvatGolf my favorite club is usually the one serving pizza at the end of a failed round
English

@WarrenInTheBuff you're looking at the new owner of an axe-throwing and oxygen/smoothie bar!
English

@VicVijayakumar Hey dad - I don’t really need your help. I just called to say I’m still unemployed.
English

@theo @grimcodes my freshman HS class was the first to get school-issued laptops (Gateway) in the mid-2000s. School IT sucked so one of their local protections for visiting illicit sites was a huge plain text file of the most nsfw urls to blacklist. had fun reading them to the class
English

weird question
anybody have a giant list of slurs? 😅
grim@grimcodes
how does everyone prevent people from saying nono words on their apps? 👀
English

@staysaasy As they get older the incentives switch and it's more about being strong enough to not embarrass yourself wrestling in the living room
English

@WarrenInTheBuff no, I want my fyp to be posts from accounts I've never interacted saying shit like "if this is true, you have to start preparing" with no context
English

@dillon_mulroy Seeing these a lot with 5.4
going back to 5.3-codex even though that has it's own "fallback hell" to deal with
English

5.4's idea of cleaning up stuff

Thomas Schranz 🍄@__tosh
@mitsuhiko I found gpt 5.4 pretty good at cleaning up stuff from earlier models needs some meta prompting in some cases that help me understand what status quo is and what our options are etc but I expect this to continue w/ better models
English
tyler 리트윗함














