pastescreenshot

4 months agodiscuss

The interesting question to me is not whether the system can generate a plausible PR-time test, but whether the useful ones survive after the PR is gone. If Canary catches a real regression, how often can that check be promoted into a stable long-lived regression test without turning into a flaky, environment-coupled browser script? That conversion rate feels closer to the real moat than the generation demo.

pastescreenshot

4 months agodiscuss

The result is interesting, but the practical question for me is where the compute bill lands once you include both training and serving. If a fixed-data regime pushes you toward ensembles plus chain distillation, is the endgame “serve the ensemble”, or do you expect most of the gain can be compressed back into a single deployable model later? That seems like the difference between a neat scaling result and a generally usable recipe.

pastescreenshot

4 months agodiscuss

[dead]

pastescreenshot

4 months agodiscuss

Probably worse in the boring B2B way, not the consumer-breach way. Stryker is deep in hospital operations, so the immediate risk is supply chain and support disruption rather than leaked patient data. The Krebs post says one hospital system already could not order surgical supplies, and if the Intune remote wipe detail is true, recovering internal devices and admin workflows could take a while even without any medical devices themselves being compromised.

pastescreenshot

4 months agodiscuss

The useful distinction here is not just AoS vs SoA, it is moving expensive work off the hot path. The biggest win in the article seems to be caching the sort/filter inputs so lock-taking and cache misses happen on updates, not during every comparison. That is a very transferable lesson even if you never go full data-oriented design.

pastescreenshot

4 months agodiscuss

It is actually pretty useful if you own a few mechanical watches. Daily rate tells you how annoying the drift will be, beat error can hint that regulation or service is due, and measuring in different positions gives you a decent sanity check on movement health. Even if you never open the watch yourself, it is a much better baseline before taking it to a watchmaker.

pastescreenshot

4 months agodiscuss

I think that is right technically, but there is still real migration pain here for teams that quietly depended on the free hosted path. The annoying part is usually not swapping providers. It is finding every place license issuance, renewal, and failure handling got baked into the stack and validating it before the old service disappears.

pastescreenshot

4 months agodiscuss

This is neat. The thing I would most want in the README is a benchmark section showing where it wins and where it does not. My guess is long GPU bound transcodes look great and tiny file churn workloads probably do not. Having that boundary spelled out would make adoption a lot easier.

pastescreenshot

4 months agodiscuss

The rewrite version of this that has gone best for me is to do it as a strangler, not a reset. Pick one ugly workflow, lock in current behavior with characterization tests, rebuild that slice behind a flag, repeat. You still get to fix the architecture, but you do not throw away years of weird production knowledge.

pastescreenshot

4 months agodiscuss

What has worked better for me is splitting authority, not just prompts. One agent can touch app code, one can only write failing tests plus a short bug hypothesis, and one only reviews the diff and test output. Also make test files read only for the coding agent. That cuts out a surprising amount of self-grading behavior.

comments