The Watercooler
What developers are arguing about across Reddit, Hacker News and X, synthesised into one read with real attribution and our take on who is actually right.
The most replied-to comment in the thread, and the metadata we checked backs it: 54 of 62 passkey managers are not in FIDO's list.
Zero replies, and it describes the lock-in lever already in use. Only 2 of 418 comments mention the AAGUID.
94 of the 418 comments argued about attestation. FIDO's own metadata does not list Apple's or Google's passkey managers at all.
Understated: we found a 150x gap between a cached token on DeepSeek and an uncached one on the dearest endpoint.
We checked the bakeoff's data file. The numbers match, from one run per model.
The 211-reply subthread. Their $50 in a few days is real, and caching explains most of it.
The top reply, and wrong on the facts: we checked the catalog and none of the withdrawn papers was formalized.
None of the three withdrawn papers was in the repo's Lean catalog, whatever the first reply said.
Posted 29 minutes after launch. The one formalized paper that needed a fix had its unchecked side claim corrected.
Our rerun backed this up: single runs of one interpreter varied by 24 to 37 percent on our Mac.
51 comments, 15 mentioning Rust and none mentioning the JIT, the one change that won every round in our rerun.
We tested it on Node 26.10.0 with npm and pnpm workspaces. The comment was right, and --preserve-symlinks breaks it.
Twelve hours after Lobsters, and the comments went somewhere else entirely: 24 of 151 mention Bun.
Slop, alienation, deskilling and team fallout. HN argued about Slack far more than about skills.
The most vivid team fallout report in the thread, and a reply contradicted it within half an hour.
The analogy the thread kept returning to: incompatible by design, slow tooling, nebulous payoff.
A GitHub co-founder's case against the default. Both big threads tried to fact-check it, and reached opposite verdicts.
903 points for a benchmark that, by its own README, cannot make a call until about October 24. Its launch-day Show HN got 5.
The only comment proposing a real experiment, Anthropic's API against Bedrock and Azure. It drew one reply.
The skeptic case in one line, and 22 replies telling it otherwise.
A Claude Code engineer floated killing plan mode on X and drew 3,122 replies and 1.6 million views. We ran plan mode in a scratch repo: it blocked a file write and let a sqlite DELETE straight through.
Armin Ronacher spent $1,200 and 35 hours on a GPT-6 Astra experiment that produced 75,000 lines of code he called worthless. A researcher measured AI code as twice as verbose as human code. Hacker News hit 746 points asking to limit AI news. The backlash has numbers now.
A claim that 89.6 percent of CDN-using European companies sit behind Cloudflare drew 242 points and 205 comments in a day. We fingerprinted 24 large European company sites ourselves and found Cloudflare on exactly one of them.
Terence Tao's 6 post thread on AI-generated proofs drew 126 favourites. The Hacker News thread that followed it ran to 94 comments, most of them about whether orchestral musicians have side gigs.
A researcher reported 60 to 80 percent success executing code through Claude Code Opus 5 in auto mode. Hacker News spent 116 comments failing to agree it was prompt injection, and Lobsters settled it in 24.