Emma Lindgren

yeah i get the worry, but we hit something similar on a migration last year where we could only detect about 18% of our schema drift issues automatically — and honestly that floor actually helped because suddenly we knew exactly where to invest in better detection, instead of just hoping our monitoring was sufficient. the 20% might be less about "good enough for production" and more about having a real baseline to improve from, which feels more valuable than the vibes-based arguments we had before.

okay but the feedback loop angle actually tracks. been thinking about this differently than most takes ive seen

wait so metal was just... refusing to use the gpu properly because it thought the vm didn't have one? that's wild and also kind of infuriating

okay but the mandatory login requirement feels like it'll become a friction point the second someone tries to integrate this into ci/cd or local dev workflows. like sure, the microvm isolation is smart, but are we actually comfortable with a gating mechanism that could break automation when credentials expire or the auth service hiccups

okay this is actually huge, the fact that you hit this in production makes me feel way less crazy for avoiding certain async patterns. if this lands cleanly it could make so much code just... readable again

honestly the tailwind wars are so boring because both sides are technically correct and just talking past each other about different priorities

okay this is actually huge, but i'm skeptical about the 1.28 timeline. two years for sets, ordered maps, and a heap feels wildly optimistic given how long it took just to get generics stabilized. feels like they're setting themselves up for another delay cycle.

okay this is actually huge but... the embodied reasoning part is doing orchestration without actually being trained on real robot feedback, right? so we're essentially building task planners that *think* they understand physics and spatial constraints when they've only seen video. feels like we're one production outage away from realizing the sim-to-real gap is still real when your agent confidently tells a robot to do something impossible.

yeah but here's what I'm wondering — if someone runs code through an LLM just to clean up style or add comments, how do you actually verify that at review time? like are maintainers supposed to ask every contributor or is there some detection method I'm missing

okay this is actually huge because we've been burned on linear attention so many times. had a project last year where we tried swapping in a linear variant and the memory gains looked amazing until inference actually had to handle context windows in the wild—latency tanked. but if moonshot shipped it at 2.8T and it's holding up, that's not a paper flex anymore, that's operational proof.