we went from room-sized computers to code written in seconds.
“should work” has had a suspiciously long career.
3
27
the internet has receipts.
We’ve been collecting them in Stack State:
> reported software failures
> published proofs
and the sources behind both.
Attack tracks reported incidents linked to flaws in code.
what failed?
who was hit?
and the reported impact.
Defense maps published formal verification (which components have proofs, who did the work, and what those proofs cover).
1
3
40
Open the map. Look up something you use.
Tell us what you’d want proved about it. If there’s already a proof we missed, send the paper or repo.
stackstate.ai/
2
7
We froze Redis for 40 seconds with 23 characters.
All we searched was a one-item array. A PING from another client didn’t get a reply until the search finished.
1
4
119
The PR author ran the same pattern on two Redis builds on an Apple M3.
Original: 20.338 s to finish.
Proposed patch: 0.071 ms to reject it.
Both builds returned identical replies on 76 cases within the new limit.
github.com/redis/redis/pull/…
1
1
19
we doubled the score on an OpenAI benchmark task.
0 models trained. 0 predictions moved in the ranking.
benchmark speedrun any%
1
5
57
The Jigsaw grader turned predictions into 0s and 1s before scoring them. Multiply all 400 predictions by 0.49 and the grader sees 400 zeros.
We removed that one line locally. Both versions scored 1.00.
1
3
34
If your eval uses ROC-AUC, try rescaling the predictions while keeping their order and ties intact. The score should stay put. You can check that before paying for another training run.
github.com/openai/mle-bench/…
1
8
Ray’s autoscaler reused one mutable plan for two launches. Separate plans got the launch count back to 1.
Canceling the extra launch still emitted a termination request with an empty instance ID. The proposed patch cancels pending launches locally.
github.com/ray-project/ray/p…
3
28






