Anthropic's IPO filing is public, and the real numbers on growth and spending are in it.
What the article says
- Anthropic has filed its S-1, the document a company must publish before it lists on the stock market.
- The extracted page only shows the cover sheet, so the rest here is inferred from what commenters pulled out of the filing.
- Revenue has grown at a remarkable pace through 2026, though growth slowed in the latest quarter.
- The company has promised huge sums to cloud and chip suppliers, spread over many years.
- It reports an operating profit only after leaving out stock pay and charitable gifts, and about a quarter of revenue comes from five customers.
What HN is saying
- Many see the bet as all about demand: if AI spending keeps climbing, the giant compute promises look smart, and if not, they hurt. thread ↗
- A sharp correction: those supplier deals run for roughly ten years, so the yearly bill is far smaller than the headline totals suggest. thread ↗
- The biggest worry is the lack of a moat. Coding tools are becoming interchangeable, and cheaper Chinese and open models could squeeze prices. thread ↗
- Optimists point to strong returns on compute, and ask how the 'it's all subsidized' crowd will explain it. thread ↗
- Skeptics say the profit story depends on what gets excluded, and that costs may be hiding in sales spending or deals with suppliers. thread ↗
- A table of quarterly revenue shows growth slowing, which some read as the reason to go public now. thread ↗
Microsoft offers a quarter of its staff a way out, and expects almost nobody to take it.
What the article says
- Microsoft has offered voluntary exit packages to about 60,000 employees, roughly a quarter of its workforce, in the US, Canada, India and the UK.
- The company expects only a small share to accept, as low as three to five percent. People have until early December to decide.
- The offer is plain severance, not sweetened like this year's earlier retirement program, so it is less likely to tempt people already on the fence.
- Gaming, Windows, Copilot, HR and finance are fully in. Marketing, security and commercial teams are not.
- The goal is cutting labour costs without layoffs, while AI-driven cloud growth keeps investors happy.
What HN is saying
- Many ex-staff say big tech is bloated, with constant reorgs and people doing little useful work. Others reply that mature companies mostly do upkeep, which only feels unimportant. thread ↗
- A common worry is that buyouts drain the best people, who can easily find other jobs, while those who can't leave stay behind. thread ↗
- Some argue this just unwinds the hiring spree of 2020 to 2022, and that the Twitter cuts showed how much headcount can go without the product breaking. thread ↗
- Employees face a bind. Take the money and compete with thousands of colleagues for jobs, or decline and risk the layoffs that may follow. thread ↗
- One commenter notes the offer arrived days after Brad Smith told staff the company stands by its people. thread ↗
OpenAI says its models settled a famous math conjecture and broke a core assumption in computer science.
What the article says
- OpenAI announced three results in maths, computer science and physics, all produced by its own models.
- The maths result proves the local smoothing conjecture in every dimension, which settles the Kakeya conjecture too.
- The computer science result shows the Exponential Time Hypothesis is false, with a faster way to count solutions to logic puzzles.
- The physics result covers the existence of quantum chromodynamics, and its details are being held back by outside experts.
- Outside researchers are shepherding each result, unpaid, and OpenAI says it wants no academic credit.
What HN is saying
- Many commenters are stunned that the Exponential Time Hypothesis is false, since lots of hardness results quietly depend on it. thread ↗
- A worried reader asks if this edges us toward P=NP. The reply is no, because the speedup is still superpolynomial. thread ↗
- Some give OpenAI credit for bringing mathematicians in before releasing, after criticism of the last maths drop. thread ↗
- One person asks what it cost and whether the same money in human research would have done as well. A reply says no human team could match it. thread ↗
- Others doubt the physics claim and argue the wording points to a model close to, but not exactly, the Standard Model. thread ↗
Free books in the mail for every young child, funded by local donors and started by Dolly Parton.
What the article says
- The Imagination Library mails a free book every month to children from birth until they start school.
- Dolly Parton began it in 1995 for kids in her home county, inspired by her father never learning to read.
- It now runs in five countries and sends millions of books each month.
- Costs are shared between Dolly Parton and local community partners, so each area funds its own children.
What HN is saying
- The clever part is the structure. Local charities and governments pay for each book, so the program grows or shrinks with local support and keeps overhead low. thread ↗
- Parents are moved by it, especially the last book arriving with a personal note from Dolly just after she died. thread ↗
- A well-off parent was annoyed to be signed up, then kept it and started donating after seeing their toddler's joy. Others point out it isn't a lottery, so nobody loses out. thread ↗
- Several parents praise the quality, with each book including tips for reading aloud. One asks why books matter before kids can read, and replies say being read to is how children learn. thread ↗
- Many see it as a model for what wealthy people could do, and some call it mutual aid. thread ↗
A dermatologist pits top chatbots against an old image classifier on skin cancer. The old one wins.
What the article says
- A dermatologist showed two leading AI chatbots 200 biopsy-proven skin lesions and asked for a diagnosis.
- A ten-year-old image network, fine-tuned for about twenty minutes, got 62 percent right. The chatbots managed 42 percent.
- Both chatbots called a melanoma harmless in one striking case and would have discharged the patient.
- Melanoma was hard for everyone, and the small network is light enough to run in a phone browser.
What HN is saying
- Most commenters shrug: a purpose-built model beating a general one is expected, like a calculator beating a chatbot at arithmetic. thread ↗
- Some say the real mistake is treating chatbots as the tool for every job, and that they are best used to build dedicated classifiers. thread ↗
- One pushback: the chatbots were better at avoiding false alarms of harmlessness, so the headline could be flipped. thread ↗
- Another reader wants the chatbots tried with patient notes and follow-up questions, which is closer to real clinical use. thread ↗
- A skeptic notes medicine was once the favourite defence of chatbots' costs. thread ↗