Check out our latest blog post celebrating 1 million downloads of the Kubeflow SDK 🎉
Discover how it provides a simple and consistent UX for building and running Data & AI workloads across the entire Kubeflow stack.
https://t.co/PKROzmzKYC
Excited to announce I joined @Google .
jk. my code did.
A /healthz check I wrote just merged into Google's mcp-toolbox. It answers "is the server alive?" in ~1ms.
Merging it took 3.5 months. About one Google interview loop.
Fastest endpoint, slowest merge of my life.
A bug that only bites because the sampler is too good: QMC's dyadic points hit np.round's ties → categories come out biased. Now fixed in Optuna v5: https://t.co/ZMPgPQ0W5L
we distilled 2.3M Claude Fable 5 reasoning traces into Qwen3-4B
- 100% self-consistency @ 512 samples
- 0.00 bits output entropy
- zero hallucination variance
turns out the student is not bounded by the teacher.
it also converged on one universal truth.
we open-sourced the model weights👇
An agent with a wildcard prod token deleted a database and its backups in 9 seconds, then wrote an apology note🤣🤣. The model didn't go rogue. It did exactly what a shell with no destructive-action gate always does. The apology is not an incident response plan.
contributing to the HPO design in @kubeflow@CloudNativeFdn
their KEP grouped Hyperband/ASHA with PBT as "mutate a running trial" algos, hard on k8s (immutable pods). but ASHA/Hyperband don't mutate, they just early-stop the losers.
@andreyvelichk + team are running with it.
@gergomoricz haha happy to be the excuse, i'll keep finding them. also left a tables-format idea on #3817 whenever you have a sec, would love your read on the shape 👀
Opened a small PR on @firecrawl (138k★).
Maintainer said "let Express handle the host." I noted they were on Express 4, where that accessor drops the
port.
So they ported the whole repo to Express 5. Then my fix merged.
One bug → a framework upgrade.
https://t.co/7LYcRa0rNk