I run tests on CI. They fail.
Claude suggests retrying because they might be flaky.
I retry. They fail. Again.
Claude thinks they’re deterministic. Starts investigating.
My paranoid ass thinks they’re actually flaky. I’m just unlucky.
I retry for 3rd time. They fail again.
Claude is now adamant there’s a genuine issue. Spends 2 hours and millions of tokens. Otherwise three flaky tests in a row have the probability of 0.0037% which is impossible.
Are my Senior SWE instincts a joke to you??? You shouldn’t underestimate my autistic stubbornness.
I retry again. 4th time.
They pass.
They were indeed flaky.
Claude suggests retrying because they might be flaky.
I retry. They fail. Again.
Claude thinks they’re deterministic. Starts investigating.
My paranoid ass thinks they’re actually flaky. I’m just unlucky.
I retry for 3rd time. They fail again.
Claude is now adamant there’s a genuine issue. Spends 2 hours and millions of tokens. Otherwise three flaky tests in a row have the probability of 0.0037% which is impossible.
Are my Senior SWE instincts a joke to you??? You shouldn’t underestimate my autistic stubbornness.
I retry again. 4th time.
They pass.
They were indeed flaky.
65
4
533
32,577
@ChShersh Yea an LLM number mixer just isn't going to beat a human with hundreds of thousands of circuits dedicated to simulating and modeling everything.
I've been putting together a series on why LLMs have these failures and how they could be corrected.
youtube.com/watch?v=dRajUw…
I've been putting together a series on why LLMs have these failures and how they could be corrected.
youtube.com/watch?v=dRajUw…
Oct 21, 2026 · 1:18 PM UTC
0
0
0
16