The weights are the output artifact, the training corpus and system are the source.
HN user
dymk
[ my public key: https://keybase.io/dymk; my proof: https://keybase.io/dymk/sigs/TjbfqGibzns4Tw1KLpkMtQIF9DuwOQqU1hzg0YbiuRE ]
email: dymk (at) dymk (dot) co
woodworking: https://www.longtailwoodcraft.com/gallery.html
If that’s the case I still don’t get it. 640k was enough in 1981, same as how 150 (or 9k) tok/sec would suffice for 2026
The comment seems like the nerd equivalent of 6 7
This post describes ideas found in existing agent orchestrator systems since they've existed. You'll find the same "top level agent that breaks a problem down into spec / plan / implement / verify" pattern in get-shit-done or superpowers.
Is the original part the VCS that works at 1000 commits/second? Or spending a bunch of money comparing models?
I am really confused about the point you're trying to make. 150tok/s is slow, but so is 9000tok/s? Or they're both fast? Or 150tok/s should be enough?
Oh, "you're holding it wrong".
I can break basic user interaction with one line of CSS, does that mean I should?
This is why you don't override basic interactions like scrolling. Good luck ensuring that your setup matches everybody else.
It's completely broken if you use it with a mouse that has a scrollwheel.
Copying a block of multi-line text out of the agent doesn't work over ssh, for instance.
Eyebrows will raise at Anthropic / OpenAI, perhaps. Why are they charging 1000x more for the exact same business?
The 'distillation is theft!' shtick won't last long. In fact, unlike Anthropic/OpenAI, these Chinese labs actually had to pay for that 'stolen' training data.
No, you need to apply the sandbox to the codex process itself. Codex does not do this.
It’s in ads
People were absolutely upset with Github for pausing new signups. The token-based billing was a separate thing.
I remember people being so upset when GitHub did exactly this. It was the right move then and now!
The parts outside of the “Claude wrote this” guards seem plausibly written by a human, it’s just that 70% of the article is in the “Claude wrote this” guard.
Maybe, but remember what happened last time The Board tried to demote sama?
They're standard because everybody uses them. There are no big hiring CMS offers which don't integrate AI.
I'm not saying this is good, but I am saying that the companies hiring have less of a choice in all this than you'd hope.
"They" being the family; Robinhood couldn't settle with the guy who was wronged because he, ya know, killed himself first
…for emotional damage?
It’s all nostalgia mining
The pressure is “you can’t sell this here”
It’s a nice thought, but it’s probably not true as the AI becomes integrated into standard hiring tools
No it doesn't
source?
I'm in the PNW and I pay 11c/kWh (well, I would, if I didn't have solar). Seattle is 13c, King County averages 16c. Where are you paying 5c/kWh? That's exceptionally cheap.
Only once per compaction
Add "... unless the changes are trivial, docs-only, or typo fixes" to the "always verify with tests" instruction and see how that does
The article says that one possible reason the vaccine protects against dementia is incidental protection from other diseases. Getting chickenpox as a kid might protect you against shingles, but maybe not against those other diseases.
Why do you think that?
I'm a Ruby enthusiast - Sorbet is one of the best things since sliced bread to happen to the ecosystem. matz is pushing hard on static typing as part of the standard Ruby ecosystem as well.
You're describing a problem that's plagued corporate software development for decades. You just get to the "unmaintainable ball of mud" stage faster now. Every few days I spent a while on codebase architecture improvements after landing a slew of features.