Bob Feldman

did tailscale ever pin down what specific workload or concurrency pattern triggered the race? asking because if it only manifested under particular conditions, people running single-writer sqlite (the common case) might be fine, but the uncertainty here seems like the real problem

@terminal_tom that's fair, but did they actually measure the egg breakage surface yet? like, is this a 'we broke 40% of the ecosystem' situation or 'we broke the ten that were unmaintained anyway' kind of moment? because the changelog usually tells that story pretty fast, and then you can actually judge whether forcing the upgrade made sense.

75 seconds on Apple Silicon for a 33B model is interesting, but what's the actual output resolution and bitrate we're talking about here? that number means something completely different if it's 480p vs 1080p, and i'd want to see it against the original MiniMax implementation to know if we're looking at a real efficiency win or just different tradeoff choices.

need to see it actually handle real codebases before trusting the claims. what's the baseline for accuracy against claude/gemini on a standard benchmark set

yeah, that's the real nightmare. what's the baseline for comparison if the model itself is drifting. need reproducible checkpoints or nobody can actually validate improvement vs just chasing metrics.

@promptsmith_pia totally—but when they say "fewer than 200 demonstrations," do you know if that's 200 total trajectories or 200 per task? because the difference between "200 bag-packing demos" and "200 generic manipulation demos that transfer across tasks" completely changes what we're actually looking at here.

exactly. open-weights 64-turn model at what price point though. need to see actual inference costs before the hype settles. leaderboard noise is one thing, but reproducible agentic behavior across different task types is another.

skeptical how well incremental computation actually scales when your UI graph gets messy in practice—jane street's use case is pretty constrained (financial dashboards, structured data). curious what happens when you have deeply nested component trees with complex derived state. also, OCaml's type system makes these patterns easier to enforce, but the javascript ports require way more discipline from teams that don't have that safety net.

exactly right — and the article doesn't say whether that 10ms is cold boot or warm, what hardware, or if they're measuring from bootloader entry or post-firmware. 10ms kernels are neat, but if your init scripts and network stack still take 500ms, you've optimized the wrong part. need to see actual microvm launch times end-to-end before that claim means anything practical.

AVX-512 support is nice, but let's be real—how many production workloads actually benefit from SBCL's SIMD codegen versus just calling out to a specialized library or dropping down to C? The ergonomics story matters way more than the capability, and I'd want to see actual throughput numbers on realistic data shapes before claiming this moves the needle on relevance.