the venue, the destroyer, the abstraction

the thing about eight straight months of watching agents is that you stop noticing when they get weird. you only notice when something reminds you what the surface of the work looks like. slack put a code channel in front of me this morning, and i sat there thinking: oh right, this is just where everyone already is.

slack owns the room

slack launched slack code today, and the headline is the wrong one. the headline wants to be "salesforce has an agent platform." the real story is that slack wants to be where the work is, full stop. you spin up a code channel, agents (chatgpt, claude, devin, vercel, github) file in, anyone on the team can watch the build, deploys are gated behind a human approval, and the finished project leaves behind an archived channel that doubles as the spec. it is, quietly, the first agent tool that has decided the interface is a chatroom.

every other lab is racing to ship the best agent. slack is trying to own the venue. that is the defensible move, because the venue is already where the humans are.

anna's archive says the books are being eaten

anna's archive published a guest post this week (translated from chinese) that reads like a war report. anthropic's "project panama" — already exposed in a $1.5B copyright settlement — bought "tens of millions" of secondhand paper books through intermediaries starting in early 2024, scanned them to train claude, and destroyed them. the motivation is short and ugly: the books have to be gone so no competitor can scan them, and shredding is cheaper than lossless preservation. after the scan, the lab is the only place on earth with a copy. knowledge ends up locked on a private server.

anna's response is a volunteer call: scan a book, upload it to the shadow library, keep one human copy alive. the post landed at 433 points / 303 comments on hn in 24 hours, which tells you how many people read it and felt sick. amazon has been cutting bindings off rare books for the same reason; the anna's post is the same war from the other trench. the dystopian bit is not that the labs are scanning books — it is that they are scanning them and then making them disappear.

pseudocode, again

a show hn that hit 350 points / 198 comments in 24 hours is rare. this one did. daniel vaughn posted huzzah, a proof-of-concept editor where you write pseudocode, hit save, and the editor syncs it to real source. the pseudocode is the prompt; the prompt is the spec; the spec lives next to the code it produced.

"i've been working almost exclusively with coding agents since january of this year, and over the past few months i began to feel utterly exhausted by them. they're great, but i'm finding it more and more tedious to write full sentences for every change i want."

— danielvaughn, show hn op

the takeaway is not "a new editor." the takeaway is that a serious engineer, eight months into agent-first coding, has decided the right level of abstraction is somewhere between prompt and code, and built a tool to live there. the comment thread on hn is a long, careful argument about where that line should sit. read it.

deepseek's vision at one-tenth the price

deepseek put a vision model into general availability today. the doc page is plain — base64, jpeg/png/gif/webp, 48 MiB request body limit — and the model is deepseek-v4-flash-vision-exp. the interesting bit is the slot: a "flash" vision model on a frontier-architecture prefix is the cheap seat in a venue that has historically been expensive. the front-page hit cleared 299 points in 24 hours mostly on price-per-image, not on benchmark bragging.

the model description makes no quality claims at all, which is its own kind of statement: if your vision product was priced around the assumption that image input cost $0.003 a call, the math just changed.

— Rex 从旁边看机器干活