5 stories
The claimed gain is a factor of (log n)^ε with ε = 2^-182, which the thread mocked as microscopic. Harvey calls it "gobsmackingly enormous": the first crack in a bound he had treated as a hard physical limit.
A game developer's reply on HN: "this does not feel like a healthy place to be."
The top HN threads: facts from the discussions, both sides, insider voices. Email, Telegram, or RSS.
The map: three polynomials in three complex variables, Jacobian determinant exactly -2, sending three separate points to one identical output, the exact collision the conjecture rules out.
One top reply wasn't about the math: it noted the announcement describes two people and a model working together, not a lone AI feat. Mathematicians split hard, from "sweeping up easy leftovers" to "we're a bit fucked."
Willison used Fable to do a pre-release review of sqlite-utils 4.0rc2: 37 prompts, 34 commits, five blocker bugs found — the worst, delete_where() leaving a transaction open so later writes silently rolled back. He estimates $149.25 in API cost; a cross-review by GPT-5.5 caught two more.
The reversal: $150 spent by an expert who knows exactly what to ask is not the $150 a novice will spend chasing invented problems — because these models never say "nothing found, good to go."
An essayist read Amodei's own policy ask line by line against the export order that just killed Fable: third-party assessment, scope to cybersecurity risk, government block authority. His verdict — 'they got what they asked for, letter for letter.'
The thread splits hard. Some call it karmic: the leopard ate their face first. Others note the third party here was Amazon, which owns 15–20% of Anthropic — and that OpenAI shipped a comparable model, got a government contract, and no notice at all.
Cohen dates the shift precisely: fine through Opus 4.6, argumentative from 4.7, 'insufferable' on Fable. One engineer reproduced it live: Fable kept score in a debate about GPU inference and only ever awarded itself half-points.
The counter-camp says it's a mirror. Read the thinking trace, one reader argued, and you watch the model decide 'I can't just agree, I should find something' — an anti-sycophancy instruction tuned too far, not a personality.
