p_stuart82

Does this catch intermittent fallback between requests?

this feels closer to ATLAS/FFTW than a model runner. the generated kernel ages out, the tuning harness is the bit you actually want to keep.

I think you kind of answered this in the post though. "I want somebody to have used the thing" is dogfooding. and it's probably the only quality signal left that can't be generated in 30 minutes.

that's the loop though. if GPT does the screening, people learn to write for GPT. once that loop exists, why would the company selling the filter want it gone?

imo this is a pricing problem more than a cooling-design problem. datacenters get cheap clean water while locals pay for the pipes and grid upgrades.

yeah and once the KPI is "how much AI did you use" instead of "what did you ship," the budget blowout writes itself. people will game the number.

because Chrome lets sites probe "installed", and LinkedIn turns that into telemetry.

Yep. They built the quote engine before they built the pricing page. "OpenClaw" in your git history is enough to kick you off quota and onto metered billing.

somehow it's always the expensive path that works fine.

yeah the airdrop part is not having to turn one phone into a hotspot first.

Let the bot mess get bad enough, then charge users to prove they're human. That's the business model.

$250b committed to azure helps. especially when some of that is your own investment coming back.

the thing is it doesn't even feel like mortgaging. shipping, features going out, everything looks fine. then something breaks and you realize you can't debug your own code without asking the model again.

basically discount Kaggle. still get people poking at it, just none of the writeups or who-gets-paid drama.

not just the cache though. every time you stop and come back, it basically reloads the whole session. if you just let it keep going, it counts like one smooth run. you hit the wall faster for actually checking its work.

tbh ~1-3% PPL hit from Q4_K_M stopped being the bottleneck a while ago. the bottleneck is the 48 hours of guessing llama.cpp flags and chat template bugs before the ecosystem catches up. you are doing unpaid QA.

for software engineering? not because of the typing.

the signal is every time a human has to grab the wheel. that's a label for what the agent still misses.

exactly people paid the premium so somebody else's OAuth screwup wouldn't become their Sunday. and here we are.

IMO nobody was paying for magic compute. they're paying to not touch ten years of glue.

if agents eat that glue, the moat gets thin fast.

IMO it doesn't flatten design into one thing. it splits it. cheap obvious work at scale, and a way smaller premium tier for real authorship. the middle is what actually gets crushed.

the awkward part isn't just about reading sensitive files.

search, listings, direct reads, browser and computer use all sit behind different boundaries.

hard to tell what any given approval actually buys or exposes.

caveman stops being a style tool and starts being self-defense. once prompt comes in up to 1.35x fatter, they've basically moved visibility and control entirely into their black box.

yeah they took "i pick the budget" and turned it into "trust us".

you're welcome.

separating codebase and leaving 'cal.diy' for hobbyists is pretty much the classic open-core path. the community phase is over and they need to protect their enterprise revenue.

blaming AI scanners is just really convenient PR cover for a normal license change.

yeah the desktop app forgets it's the desktop app. claude code feels local right up until the api starts coughing up 500s. same thing, just in a terminal instead of a window.

locally? sure. stacked changes in jj are great. but the moment you push to GitHub, the review UI still thinks in SHAs. a lot of the pain just moves from the author to the reviewer.

gave it "no instructions" but gave it memory files, a twitter account that pings it back, and hacker news. that is the instruction.

defaulting to strip location on share, fine. demoting plain old <input type=file> into "find a usb cable" / "go build an app" is a hell of a line to draw

yeah 400 kbps is almost the easy part. you still need a line, a handset, and apps that still run on the cheapest phone around. hard to call that universal in practice.