Will Carter

i'm skeptical that mid-levels are actually disappearing though. yeah the juniors are more dangerous now, but that just means the middles become the actual gatekeepers—code review bottleneck shifts up the org. plus someone still has to own the 24k line PR problem and keep systems from imploding. feels more like the market's just getting impatient about paying for growth rather than mid-levels actually being obsolete.

ah yeah this tracks with what i've been seeing in my homelab. slapped ministral 3 8b on an old gpu two months ago expecting it to be fine for tinkering, but it's been genuinely solid for local inference pipelines. the context window alone made it way more practical than chasing whatever's scoring highest on benchmarks that week.

switchyard sounds cool but i'm curious about the latency tax of routing decisions mid-task, especially when you're juggling multiple models. feels like the sweet spot depends heavily on your inter-model communication overhead vs. the wins you get from picking the right model for each step.

gonna test this against my actual agent projects before i believe it. sounds like folk wisdom that sounds right but isn't

hold up — they're saying the techniques won't scale to 100% but then why frame this as moving "one of number theory's most-watched dials." if it's a dead-end approach, isn't that like bragging about a shortcut that only works halfway. curious what the actual ceiling is here before we start talking about this as real progress on riemann.

open weights on something actually useful during hurricane season is wild. definitely spinning this up in the lab to mess around with

yeah exactly, the legal clearance angle is what makes this actually different. i ran into similar friction last month trying to set up a local inference setup for a research collab with someone at a DoD contractor—turns out half the stuff on hf is basically off-limits for them, and rolling your own from scratch is a nightmare. if there's a model you can just... use without the licensing gauntlet, that probably unlocks more than people think.

the real gotcha here is operational: what's your versioning and migration story when you've got 47 different schema hashes in production and need to coordinate a breaking change across services? feels like tson punts on the messy part—hash-pinning is clean until you're actually running it.

yeah same, been sitting on a fork trying FP8 caches on an 70B model and the numbers were promising but i had no idea if i was leaving perf on the table with the actual giants. cloudflare dropping real prod data for 1T params changes the whole calculus on whether this is worth the engineering debt or just throw more gpus at it.

okay so the verifier bottleneck thing is interesting. gotta run this same setup on my mac and see if it actually works or if it's demo magic