Fifty-five stories cleared 200 points on the first five pages tonight, twenty-nine of them new. That is the biggest new-story count in a week, and it is almost entirely down to one filing.
Anthropic’s S-1 is on the front page at 1,253 points with 1,220 comments, and it is the only document in this roundup that anybody can check line by line. Revenue of $16.3 billion in six months. Compute and infrastructure at $8.9 billion in the same six months. A $120.5 billion commitment to a company that was called xAI a few months ago. A net loss of $80.7 billion, most of it mark-to-market noise on convertible notes. And the single most interesting line is an accounting policy change: in 2025 Anthropic stopped reporting gross profit.
Everything else on the page tonight is either a smaller version of that argument — cheap models, on-prem models, a CEO saying 90% of the performance at 10% of the price — or a clean break from it. A woman who has given away 343 million books. A Minecraft Switzerland you can walk across. A journalist being quoted $2.5 million for a police file. The page is unusually good tonight. Start with the filing.
Anthropic S-1 — 1,253 points
1,253 points · 1,220 comments · sec.gov · HN discussion
Filed this morning, Nasdaq symbol ANT, offered through a public benefit corporation with a Long-Term Benefit Trust holding Class T shares that elect a majority of the board, plus a founder-held Class F share. All of that governance detail is in the first four pages and all of it is a distraction from the numbers, which are the reason this is the best story of the week.
Revenue: $41 million in 2023, $386 million in 2024, $4.585 billion in 2025, and $16.286 billion in the six months to June 30, 2026, against $1.163 billion in the same period last year. Q2 2026 alone was $11.6 billion. Operating loss of $693 million for the half, operating margin of positive 8% in Q2, adjusted operating income of $1.6 billion for the half. Revenue per employee on a trailing-twelve-month basis went from $1.4 million to $6.6 million while headcount went from 940 at the end of 2024 to about 5,300 now. If those numbers are what they look like, this is the fastest revenue ramp in the history of enterprise software and the margin structure of a company that has already won.
Then you read the expense lines. Compute and infrastructure was $8.945 billion in six months — 55% of revenue — against $2.334 billion in the year-ago half. Sales, marketing and partnerships was $3.168 billion, 19% of revenue, up from $120 million, and $2.7 billion of that increase is “fees paid to channel partners.” That is the vig: the clouds Anthropic also owes compute money to are taking a cut of the revenue they route to it. Personnel ran $3.152 billion, of which $2.2 billion was stock-based compensation in the half. Other operating and administrative, $1.714 billion, does not reconcile with anything on the page in a way a reader can audit.
The 2025 change is the tell. From the filing: “During the year ended December 31, 2025, the Company changed its presentation of operating expenses from a functional classification under which cost of revenue, gross profit (loss), research and development, sales and marketing, and general and administrative were presented as separate line items, to a classification based on the nature of expenses.” Read that again. A company heading into the largest technology IPO in history stopped disclosing gross profit in the year before it filed. The functional view survived in the notes — substantially all cost of revenue is compute and infrastructure for customer-serving workloads — but the headline statements no longer let you compute a gross margin without doing the work yourself. When a vendor’s own audited numbers stop showing the one ratio that defines a software business, that is not a formatting preference.
The commitments are the other half of the story. Total indebtedness of $6.2 billion, all convertible notes to Amazon through AWS, plus the ability to issue up to $42 billion more in convertibles to Broadcom for a March 2026 equipment lease on custom accelerators. Then there is SpaceXAI, formerly xAI: agreements entered and amended between May and October 2026 for NVIDIA-powered compute, “an aggregate of up to $120.5 billion over terms extending into 2029, consisting of approximately $3.1 billion in the aggregate per month once all capacity has been delivered, which we expect in December 2026.” Three point one billion dollars a month, and the initial non-cancelable periods end in January and March 2027 — after which either side can walk on 90 days’ notice. A $120 billion monthly-billing compute deal with an early exit in eight weeks is not a commitment, it is a call option on someone else’s capex.
The blank is the best part. The aggregate committed spend figure across all purchase commitments and leases, “substantially all of which is related to compute and infrastructure,” with roughly $X before 2030 and the rest between 2031 and 2045, is a literal $ followed by blank space. Subject to completion. The largest line item in the largest tech IPO ever filed is a placeholder, and no commenter noticed because the commenters were too busy arguing about valuation.
Which they do, at length, and mostly badly. The strongest objection in the thread is the simplest one: at a rumored valuation north of a trillion dollars, every dollar of future growth is already priced, so the only way to make money on the trade is to be right about a market that does not exist yet. The strongest defense is that the demand is real and constrained by capacity — the $2.08 of revenue per dollar of compute figure circulating in comments, up from $0.40 in Q1 2025, is the single number that would justify the build-out if it holds. It is not in the filing in that form, and I could not verify it there; treat it as a commenter’s arithmetic.
One detail that deserves more attention than it got: the lock-up. Insiders can sell starting day one, then at 27, 48, 62 and 76 days after the debut, instead of the standard 180. That is a schedule designed by people who expect the first print to be the high one. Also worth noting for anyone tracking corporate filings as a genre: the Bartz settlement cost $1.5 billion in 2025, one third of that year’s revenue, and appears as its own operating expense line. Copyright is a cost of goods here, not a tail risk.
The company’s own risk language is more honest than its boosters. “We do not expect our revenue will continue to grow at comparable rates, and we cannot assure you that it will continue to grow at all.” Fine. What no risk factor addresses is the thing the expense table says out loud: this is a business with a 55% compute bill, a 19% channel bill, and a customer base where the top five accounts are a quarter of revenue and “many of our top customers are not subject to long-term contracts with us.” The margins arrive if the compute gets cheaper and the channels get disintermediated. Both of those are the stated strategy of the two companies that also own the channels.
Dolly Parton’s Imagination Library — 666 points
666 points · 107 comments · imaginationlibrary.com · HN discussion
The front page found a website that works. Dolly Parton’s Imagination Library mails a free age-appropriate book every month to any child from birth to age five, in the US, Canada, the UK, Australia and Ireland, and has done since 1995 — 31 years, 3,197,250 children registered, 343,596,903 books given away. Started in her home county in east Tennessee because her father could not read or write. Funding is split between Parton and local community partners, which is why the program is administered county by county rather than nationally.
The comments are the usual “how is this not bigger news” thread, and the honest answer is that it is not news because it is not a claim. There is no benchmark, no valuation, no roadmap. It is one organization, one logistics contract, and a number that goes up. Against everything else on tonight’s page it reads as an accusation.
Two things worth knowing before you sign a kid up. The waitlists are geographic — coverage depends on whether a local affiliate sponsors your postal code, and large parts of the US are not covered. And the program’s own research page is where the interesting material lives: the international studies on early literacy are less spectacular than the topline counters and more useful, because they measure whether the books get read rather than whether they get mailed.
Microsoft offers mass buyouts for roughly a quarter of its workforce — 512 points
512 points · 451 comments · seattletimes.com · HN discussion
About 60,000 employees — more than a quarter of the global workforce — are eligible for voluntary exit packages, in the US, Canada, India and the UK, per a memo from Chief People Officer Amy Coleman. This is the second round: April’s first-ever Voluntary Retirement Program covered roughly 9,000 people by age and tenure, and Microsoft cut 4,800 jobs in July. The Seattle Times page is paywalled past the lede; CNBC, GeekWire, Reuters and Business Insider all carry the details.
Voluntary is doing a lot of work in that sentence. A buyout offered to a quarter of the company is a layoff with a consent form, and the dynamic the comments identify is the real one: accept and compete for a new job against tens of thousands of your colleagues doing the same thing, or decline and hope the involuntary round does not find you. One commenter’s former manager put it exactly right — this is how you get rid of your most employable people first, because they are the ones with somewhere to go.
The thread also contains the best counter-narrative, from a former L63 who spent eight years across two stints: the last 18 months were unproductive, constant reorgs, work misaligned with capability, colleagues on trivial projects, promotions for work that did not warrant one. That is a description of a company that overhired in 2020-22 and is now doing the correction, and it is a more boring explanation than “AI replaced them.” Both can be true. Microsoft is spending like a company that believes compute is the product and headcount is the cost, and the buyout is the part of that belief that shows up in people’s lives.
Oregon Police Are Charging a Local News Outlet $2.5M for Records About One Cop — 508 points
508 points · 143 comments · theintercept.com · HN discussion
Double Sided Media in Eugene asked for records about officer Justin Peckels, who has been the subject of complaints including punching a resident during a traffic stop hard enough to break his nose. The Eugene Police Department’s estimate: 44,816 hours of staff time, $2,484,599.04. A parallel request about a former officer, Eric Klinko, was quoted at $195,869.52. Total for two officers’ files: about $2.7 million.
The justification is the story. The department says it cannot narrow the custodians it would have to search, so it must search the records of every city employee past and present — parks and recreation included — manually, one by one. Either Peckels has the largest disciplinary file in the Western world or the records system is configured so that compliance is unaffordable. The city’s letter is explicit that staff will “continue processing” once payment arrives, and that any shortfall becomes a further bill before release.
Read as a procurement problem, it is almost defensible: a badly indexed document store and a request scoped to “any record mentioning this person between 2018 and 2026” genuinely is a lot of human hours. Read as an incentive problem, it is a price list for the right to know. Oregon’s public records law has no fee cap and no deemed-approval mechanism, so the only remedy is litigation by a two-person newsroom against a city legal budget. This is the same failure mode as discovery costs in civil litigation, applied to journalism, and it works precisely because it is legal.
Lula wins election after final votes tallied — 485 points
485 points · 587 comments · reuters.com · HN discussion
Lula da Silva won Brazil’s runoff with 51.79% to Flávio Bolsonaro’s 48.21%, a fourth non-consecutive term after trailing 45-47 in the first round. Bolsonaro has not conceded. Reuters’ live page would not load for extraction; the numbers here are the TSE count as reported by DW, the Guardian and CNBC.
The interesting part for anyone who does not follow Brazilian politics is the polling failure in the opposite direction from the usual one. First-round results had the incumbent behind; the runoff flipped 2.4 points in four weeks, and at least one major outlet ran a statistical model live during the count that called it correctly while the pundits did not. The second interesting part is in the comments: a Brazilian reader arguing that organised, persistent, centrally coordinated online campaigning decides these elections now, that Bolsonaro’s operation had it in 2018 and lost it in 2022 for non-media reasons, and that Lula’s campaign was inert until the first-round upset forced it to move. That is an unfalsifiable claim from one voter, but it matches what the Meta story below alleges.
The dull structural fact: Brazil has now elected the same man four times, and the runner-up is the son of a man under house arrest who promised to pardon him. Both parties’ incentive to reform anything is close to zero.
Birdwatching rewires the brain and slows cognitive decline — 439 points
439 points · 179 comments · bbc.com · HN discussion
The neuroscience is real, and narrower than the headline. Erik Wing’s group at Baycrest in Toronto scanned expert birders and novices while they identified birds, and found expert brains wired differently in attention, visual perception and visual recognition regions, with more activity in those regions specifically when identifying unfamiliar birds. Experts also showed higher measures of network organisation — a proxy for cognitive health that normally declines with age — in some of the same regions.
That is correlational, and the article says so. The structural differences could be the cause or the effect of expertise; nobody has run the intervention. The clinical half is more solid and less headline-friendly: Age UK runs birdwatching groups for dementia patients in the West Midlands, and the reported outcomes are things like a withdrawn 78-year-old laughing at ducks. That is not a cognitive endpoint, it is a mood endpoint, and mood is what the carers actually need.
The generalisable claim — expertise in a demanding perceptual domain preserves brain organisation — is not new and not about birds. It worked for taxi drivers and musicians first. What is new here is an unusually well-matched control group (novices doing the same task) and a task with a real expertise gradient. If you want the honest version: learn something hard that involves looking closely at the world, and keep doing it. The bird is incidental.
Switzerland. Block by block. At 1:1 scale — 425 points
425 points · 199 comments · blocks.manifold3d.ch · HN discussion
The entire country of Switzerland rendered in Minecraft at one block per metre, 24 GB, built from swisstopo’s open geodata, running on Java 1.12.2 with the Cubic Chunks mod and Manifold shaders. The Matterhorn is exactly as tall and as steep as the Matterhorn. There are 4,634 blocks from the valley floor to the Dufourspitze. Glaciers, lakes, dams, forests.
Two things make this more than a screenshot project. First, the vertical scale: vanilla Minecraft caps at 384 blocks, so this needs a mod that removes the height limit, and the chunks are cubic because otherwise a 4,634-block mountain cannot exist. Second, the source data is a national mapping agency releasing terrain at sufficient resolution for 1:1 in a 16-bit voxel format — that is a public-data story wearing a video game costume.
The comments from people who have tried this are the useful part: 24 GB of world is a lot of disk and the render distance will humble any GPU, so it is a flyover and a hike, not a server. Worth downloading anyway. Somebody reproduced the Aletsch Glacier’s medial moraines because the elevation data had them.
IBM CEO says open weight models are now irresistibly good and cheap on-prem — 405 points
405 points · 314 comments · theregister.com · HN discussion
Arvind Krishna, on IBM’s Q3 call, gave the cleanest version of the open-weight argument any vendor has put on the record: in 2024 a broad benchmark had the top closed model at 70 and open models at 20 or 30; by late 2025 it was 70 to 50; today open weights deliver 90% of the effectiveness at 10-25% of the token cost. “Would you like something that’s 10 percent of the cost if it’s 90 percent as good as the best, or do you want to pay that 10 times higher cost?” And then the list of tasks he thinks do not need frontier capability: summarising an email, examining a contract, answering a customer call, composing an email.
He is right about the tasks and self-interested about the conclusion. IBM sells the mainframes, the Spyre accelerators, and the integration labour, and it is the one large vendor with nothing to lose from commoditised inference. The honest part of his pitch is the part he admits he cannot answer: “Where are my kernel engineers? Where are the people who worry about the underlying network?” Most enterprises do not have those people, which is why the 10x price advantage evaporates once you price the operations.
The comment thread is the best evidence for his thesis, and it is anecdotal in exactly the way that matters. Multiple people running 8-GPU rigs and DGX Sparks say they have not paid a provider in months; others say GLM 5.2/5.3 and Kimi K3 are close enough to Opus-class models for human-in-the-loop work that they stopped caring about the frontier. The counter-argument is the one nobody in the thread can rebut: those rigs cost four figures for a 2020 GPU, and the price of memory is not falling. Open weights make inference cheap and hardware expensive, and IBM would be delighted to sell you the hardware.
OpenAI and Anthropic cut their LLM prices by up to 80% within a few months — 402 points
402 points · 406 comments · minimaxir.com · HN discussion
Max Woolf reconstructs the price war from press releases, which is the only way to see how fast it happened. June 26: GPT-5.6 Sol at $5/$30, Luna at $1/$6. June 30: Claude Sonnet 5 at $2/$10. July 24: Opus 5 at $5/$25. July 30: OpenAI cuts Luna 80%. September 22: GPT-6 Sol at $2/$10 and GPT-6 Luna at $0.10/$0.50 on the same day Anthropic ships Opus 5.5 at $4/$20 with a 40-60% total cost reduction. September 29: GPT-6 Sol turns out to be a dud, so GPT-6.1 Sol arrives a week later with the cache price halved, making the family cumulatively 75% cheaper than where it started in June.
His argument about why is the good part: nobody cuts prices 80% without an externality. Models have become a commodity you swap with a string in OpenRouter or a toggle in OpenCode, so there is no switching cost and therefore no moat, and inference costs are falling fast enough that the cut is affordable. The Vercel gateway data he cites is the twist: cost per token is at 43.3% of April’s level, and aggregate spend is up anyway, because the models got better faster than they got cheaper. Cheaper tokens do not reduce the bill. They raise the ceiling on what people are willing to attempt.
The skepticism the post does not apply to itself: “the models do not appear to have cut any corners” is the author’s personal judgement across a handful of models, and the timeline is built from vendor announcements, which are the least reliable source for capability claims in the industry. But the price series is real, it is auditable, and it is the best argument on tonight’s page for why Anthropic’s 55% compute line might compress. It is also the best argument for why its revenue might not hold.
Broccoli2Shell: Brotli decoder RCE in Node.js, Python and more — 399 points
399 points · 214 comments · broccoli2shell.com · HN discussion
Out-of-bounds write in the Brotli decoder, versions 1.0.3 through 1.2.0, CVSS 9.2, remote code execution. Fixed in 1.0.10, 1.1.1 and 1.2.1; Node 26.12.1, 24.22.1 and 22.23.4 carry the fix; Node 25, 23, 20 and older get nothing because those lines are end of life. Python’s Brotli and brotlicffi are patched at 1.2.1 and 1.2.1.0.
The exposure surface is what makes this worth reading rather than filing. Node’s built-in fetch() decompresses Content-Encoding: br responses automatically, and so do requests/urllib3 and aiohttp when a Brotli package is installed — meaning any server you talk to can send the payload. Express 5.0.0-5.3.0 and body-parser 2.0.0-2.3.0 inflate Brotli request bodies in raw() and json(), which turns an obscure feature into an unauthenticated inbound path. And installing the update is not enough: a running process keeps the old decoder until it restarts, and a container keeps it until you rebuild on a patched base.
The comment thread contains the two corrections that matter. The fix appears to have landed in Google’s source in February without anyone recognising it as a security issue, so the vulnerability sat patched-but-unreleased for eight months — and one commenter asks whether an AI automation pipeline noticed the diff and nobody was at the wheel. The other is the irony: Google declined to ship JPEG XL in Chrome partly because a new image decoder enlarges the attack surface, while having shipped Brotli, which every HTTP client now decodes by default.
Sebastian Raschka’s from-scratch video course on building reasoning models — 398 points
398 points · 55 comments · magazine.sebastianraschka.com · HN discussion
Eight videos, eleven-plus hours of coding, covering evaluation, inference-time scaling, reinforcement learning and distillation — building a reasoning model from the ground up rather than prompting one. Raschka’s newsletter has 223,000 subscribers, which is either a signal about demand for fundamentals or about how few people can explain them.
The comment count is the interesting number here: 55 comments on 398 points, an unusually low ratio. This is the kind of submission people bookmark and do not argue about. Compare with the Berkeley grading story below, which has 327 comments on 231 points. Same front page, opposite behaviour.
Overgrown, when plants invade your abandoned tabs — 394 points
394 points · 131 comments · charliegerard.dev · HN discussion
A Chrome extension where each tab has a neglect clock that runs only while the tab is out of view. Leave one long enough and plants grow in from the edges; past 5% coverage the page behind the vines stops accepting clicks, selection and drags. You clear it by dragging the mouse across the growth — machete, chainsaw or flamethrower, your choice — and when you get it back under 5% the remainder falls off. Then it starts again.
Charlie Gerard built it as a creative exercise around the word “overgrowth,” and the reason it works is that it makes an abstract cost physical. A stale tab is not a problem until you have to swing a machete at it. The engineering detail worth stealing is the neglect clock: visibility-gated timers mean the mechanic costs nothing while you are actually working, which is the opposite of every productivity extension that fights for attention.
Why I Am a Programmer — 312 points
312 points · 98 comments · alfie.wtf · HN discussion
Alfie writes about growing up in a rough neighbourhood where owning a computer made you a robbery target, so the kids with computers found each other quietly and quizzed new arrivals before letting them in. The essay opens with Grant Smith’s 1993 demo-scene text file about wanting to “write demos” that change people’s perception of reality, and then argues that the thing he fell in love with was never the career.
It is the kind of post that gets upvoted because of what it refuses to do: no advice, no framework, no “here’s how to stay relevant in the age of AI.” The comments are mostly people admitting they miss the version of this they had, which is the honest response to a nostalgia essay. Worth reading if your current job title no longer describes what you actually do.
GPT-6.1 Terra — 275 points
275 points · 140 comments · openai.com · HN discussion
OpenAI’s fast-and-cheap model gets its annual refresh, and the numbers are about cost, not intelligence. Pricing: $1 input, $0.05 cached input, $5 output per million tokens — half of GPT-5.6 Terra and half of GPT-6.1 Sol. On Agents’ Last Exam at medium effort it beats GPT-5.6 Terra by 15 points at about a third of the cost, and matches GPT-6 Astra at a tenth of the per-token price. At max effort, OpenAI claims three times the token efficiency of Claude Opus 5.5 at a quarter of the price per token, which it renders as roughly 9x lower cost per task.
Two of the claims are the kind you can check later and both are specific: error rate on OpenAI’s own error-inducing-prompt benchmark fell from 53.6% to 16.7% at low effort, and Terminal-Bench 4.0 went from 21.5% to 50.3% at max effort. The internal benchmark is the one to discount — nobody outside OpenAI has that prompt set, so the 37-point drop is a claim about the benchmark as much as the model.
The framing is the story. This model is not competing to be the best; it is competing to make the median enterprise task cost a tenth of what it did in June. That is the same thesis Krishna is selling, from the other side of the market: the frontier is being priced down from below by cheaper models that are good enough, faster than the labs can move their own price points. Whether “good enough” holds on the tasks that matter is exactly what the benchmarks cannot tell you.
Meta Employees Denounce Interference in Brazilian Elections — 266 points
266 points · 109 comments · panamericandispatch.org · HN discussion
Meta employees have gone on the record calling the company’s treatment of the Bolsonaro family during Brazil’s election “obvious” and “biased” favouritism, in reporting by ICL Notícias via Panamerican Dispatch. This lands the day Lula’s win is confirmed and follows yesterday’s story about Meta ignoring a Superior Electoral Court order in the same case, where the evidence included a WhatsApp GIF search for “brasil” returning Flávio Bolsonaro first across three separate accounts and Meta AI answers linking Lula to derogatory content.
The significance is not that employees are unhappy — they always are — it is that internal dissent is now a document trail in an active electoral case. Sleeping Giants Brasil and CTRL+Z filed the earlier evidence with the court and asked for the existing R$500,000-a-day fine to be enforced and for Meta to explain how ranking and AI answers are produced. Employee statements are admissible context in that proceeding in a way that anonymous leaks are not.
The structural point, and the reason this belongs next to the IBM and OpenAI pricing stories: recommendation and AI answer generation are now election infrastructure, operated by private companies, with no public audit, no reproducible ranking disclosure, and no meaningful penalty. Meta’s defence in every one of these cases is that the mechanism is too complex to describe. The court’s problem is that it has no way to test that claim.
CSS-Tricks has a new owner — 262 points
262 points · 61 comments · css-tricks.com · HN discussion
Geoff Graham has bought CSS-Tricks from DigitalOcean and is running it independently, 12 years after joining and four years after Chris Coyier sold it. His post is an FAQ, and the honest answers are the good part: yes he owns it, yes it is an investment, yes he is making it his full-time job, no he will not say what he paid, and yes advertising is coming back — but as real packages with display, sponsorships and newsletter slots rather than whatever DigitalOcean’s programmatic remnant inventory became.
The risk is stated by a commenter and it is the correct one: buying a CSS tutorial site in a year when an LLM will write your stylesheet is either courage or madness. Graham’s bet is that the value was never the snippets — it was the curated, versioned, human-checked explanation of why the browser does what it does, and that the traffic that used to come from search will have to be replaced by membership and newsletter. That is a plausible business for a site with 18 years of goodwill and an implausible one for a cold start.
The State of Context Engineering — 247 points
247 points · 63 comments · dbreunig.com · HN discussion
Drew Breunig’s AGNTCon talk, transcribed, on where context engineering came from and how it fails. His framing is better than the usual: the job is knowing when to be deterministic and when to be probabilistic, and models are not people — you can assume a human will re-read an instruction, you cannot assume a model will notice it in the wrong position.
The O’Reilly question and answer at the top is the most useful 200 words in the post. “How do we know this won’t become obsolete in a few weeks?” His answer: giving clear, consistent, non-noisy instructions will always be hard, because we cannot do it with other sentient humans either. That is a durable claim in a field that mostly produces perishable ones.
What the talk under-sells is how much of the discipline is empirical rather than theoretical. Breunig is editing a handbook and co-founded a company built on DSPy, so his answer to “how do you know your context is good” is a compiler and an eval loop. Most teams have neither, which is why the failure modes he catalogues — poisoned context, distraction, confusion, clash — read as descriptive rather than actionable.
AI use leads UC Berkeley math professor to see ‘no point’ in grading homework — 231 points
231 points · 327 comments · dailycal.org · HN discussion
Richard Borcherds, Fields medalist, no longer grades homework for correctness. He grades on completion, admits it is “not a very good system,” and has no alternative — students using AI get consistently top scores, so paying readers to check answers is now pointless work. He says if AI keeps improving at the current rate there will be nothing human mathematicians can do better within a few years, that he does not know what to advise students, and that he asked ChatGPT which also did not know.
The comment ratio here — 327 comments on 231 points — tells you this hit a nerve the other AI stories did not. The thread splits cleanly. One camp argues homework was always a proxy for learning and that the proxy has now failed, so the answer is assessment that happens in a room without a network connection. The other camp argues that completion grading is an abdication dressed as realism, and that a Fields medalist’s despair is not a policy.
Both sides skip the thing Borcherds himself says and then drops: campus tried Turnitin and Rumi and found neither adequate to standardise. That is the actual state of the art — detection does not work, so institutions are choosing between unsupervised take-home work they cannot verify and supervised work they cannot scale. His comparison to paper journals, which have published for twenty years while being “utterly pointless,” is the most honest sentence in the piece, because it describes what will probably happen: the ritual continues, the credential inflates, and the people who care find another way to learn.
Moderna and Merck say their mRNA cancer vaccine cut death, recurrence by 42% — 203 points
203 points · 52 comments · arstechnica.com · HN discussion
The topline from INTERPATH-001, released early after an oncology outlet broke the ESMO embargo: intismeran (mRNA-4157) plus Keytruda reduced the risk of recurrence or death by 42% versus Keytruda plus placebo in resected stage IIB-IV melanoma, 1,137 patients randomised 2:1, two-year follow-up. Absolute numbers: 82% recurrence-free versus 70%. Distant metastasis-free survival: 38% risk reduction, 88.5% versus 80%.
Three things to hold onto before the press cycle gets going. First, the hazard ratio is not the effect size. A 42% reduction on a 30% event rate at two years is roughly nine percentage points of absolute benefit per hundred patients treated, and that is the number that matters for a payer, a guideline committee, and a patient deciding whether to take an individualized manufacturing slot that requires tumour sequencing, a bespoke mRNA synthesis run, and a lead time measured in weeks.
Second, the effect shrank from Phase 2. KEYNOTE-942 reported 49% RFS and 59% DMFS reduction at five years with 157 patients; Phase 3 gives 44% and 38% with 1,137. That is the normal direction for a phase 3 and it is worth remembering the next time someone quotes the phase 2 numbers as settled fact.
Third, and this is the one the coverage will not mention: this is an adjuvant trial with a Keytruda backbone, so the incremental value is entirely on top of the standard of care that Merck already sells. Merck is running the trial that extends the patent life of its own franchise, and the pricing question — individualized neoantigen manufacturing, not a vial off a shelf — has no answer in the release. The FDA has no established pathway for a per-patient manufactured cancer vaccine at scale, and the release says only that the companies are “engaging with regulatory authorities.”
None of that makes the result small. A 9-point absolute improvement in recurrence-free survival in resected high-risk melanoma is a genuine advance, the first Phase 3 win for individualized mRNA cancer therapy, and the strongest evidence yet that the platform generalises beyond infectious disease. It is also, on tonight’s page, the only story where the gap between what was measured and what will be claimed is the whole story.
Also new, below the sections
- Show HN: TitleArt – Make text like it’s 1997 — 382 points · 70 comments — WordArt in the browser, as vector output rather than a rasterised image. The best trivia in the thread: Scott Forstall worked on WordArt as a Microsoft intern before running Apple’s software. Somebody in the comments points out it does proper hard shadows in vector format, which Illustrator has refused to do for 39 years.
- Amazon’s “digital pacifier” for toddlers — 276 points · 408 comments — Amazon shipped touchscreen tablets marketed specifically at toddlers (“Kids want a real tablet, not a toy”) against a pediatric literature that links exactly this to developmental and sleep problems. The comment count is the second-highest on the page and almost none of it is about the tablet.
- Show HN: A Truly Modular Keyboard — 264 points · 93 comments — every key is its own 1u PCB with a CH32V003 RISC-V chip, pogo pins on all four sides for power and I2C, magnetic cases, Bluetooth, 2 MB of QSPI flash for key firmware. Absurd, buildable, and the firmware-per-key design is genuinely novel.
- Behind the scenes at the NeXT Computer office — 238 points · 147 comments — Stanford’s Menuez archive, 96 frames at a time. The exhibit page itself is a JavaScript archive viewer and would not render for extraction; the comments are the content, and they are mostly people under 30 mourning a kind of in-person work they never got to do.
- Websites Should Be Printable — 235 points · 161 comments — a student’s argument for
@media printas a first-class concern: point sizes, hiding nav, keeping links legible in black and white, and the observation that printing your own page is the fastest audit of whether your HTML is semantic. - JetBrains Thinkrail — 233 points · 170 comments — an agentic IDE built on the
picoding agent, with isolated workspaces per task and durable project knowledge via “living specs.” JetBrains incubator project, open source, 572 commits. The pitch is that it learns your project instead of re-reading it, which is the correct thing to be skeptical of. - Blaise Pascal invented fixed route public transport (2024) — 230 points · 118 comments — the carrosses à cinq sols, eight-passenger horse-drawn carriages on predictable frequencies across connecting routes, launched in Paris in 1662 by a dying man who had already done mathematics, physics and theology. Jarrett Walker’s point: everybody credits the vehicle, nobody credits the network.
- PaperPilot – Zero-Cloud, Pure-Rust PDF Engine with WASM and MCP — 225 points · 81 comments — merge, split, OCR and compress PDFs from an MCP server, a CLI, a WASM build that processes in browser memory, or a local REST API on port 7823. No Chromium, no C++ runtime, no cloud calls. 1,540 commits, which is either maturity or a warning.
- fheroes2: A recreation of Heroes of Might and Magic II game engine — 213 points · 73 comments — a from-scratch engine that needs the original assets, high-resolution support, improved AI, 6,014 commits over years of volunteer work. GPL, still shipping.
- How Japanese Dried Bonito Is Made — 200 points · 51 comments — katsuobushi: simmer, smoke, then inoculate with mould and age, repeated, until the fish is harder than wood and rings like a bell when struck. Fourteen steps of a craft that exists because somebody kept going after the obvious stopping point.
Still on the page
The carryovers split into two groups tonight: the stories that are done and the stories that got a second life from something else on the page.
The deltas are more informative than the scores. The three biggest movers — the CNN at +411, the FOSDEM censorship call at +306, and Dan Luu’s ambition essay at +184 — are all arguments rather than announcements, and all three are still climbing while every product release on the page has flattened. The one product-adjacent exception is FreeDNS at +170, a two-decade-old free dynamic DNS service that got rediscovered, which says something about how much people are paying for DNS.
Throughline
Tonight the front page has a set of audited numbers on it, and everything else is a claim about the same numbers. Anthropic’s S-1 says a company can generate $16.3 billion in six months and still spend 55% of it on compute and 19% of it on the channel partners who also sell it the compute. Krishna says the same thing from IBM’s side: 90% of the performance at 10% of the price is where the enterprise market is going, and he is happy to sell you the racks. Max Woolf’s price timeline shows the labs cutting prices 75-80% in four months because inference is commoditising faster than capability is differentiating. GPT-6.1 Terra is the product expression of that: not the best model, the cheapest good-enough one, priced at a tenth of the frontier. Four stories, one mechanism, and the mechanism is that the thing everyone is building is getting cheap while the things you need to build it — wafers, power, channels — are not.
The second theme is the cost of being allowed to check. A newsroom is quoted $2.5 million for a police file because the records system makes compliance expensive. A court cannot audit a recommendation algorithm because the company says the mechanism is too complex to describe. Researchers cannot grade homework because detection tools failed standardisation. Voters cannot verify the campaign they just voted on. None of these are censorship — they are all the same move: making verification cost more than the person asking can afford. The S-1 is the exception that proves the rule, because a company with no choice but to tell the truth under penalty of law produces 492,000 words that answer almost every question worth asking, including the one it accidentally answers by leaving a dollar sign blank.
And third: the page keeps rewarding things that can be verified rather than things that can be marketed. A Minecraft Switzerland built from open geodata, exactly 4,634 blocks high, because swisstopo published the terrain. A birdwatching study with a matched novice control. A Brotli bug with version numbers and a fix list. A retired engineer saying the last 18 months at Microsoft were full of reorgs. 343 million books sent to children, which is not a benchmark but is not a claim either. The most-quoted number on the page will be 42%, from a Phase 3 trial where the phase 2 number was 49% and the absolute difference was 12 points. The least-quoted number will be $3.1 billion a month, which is what Anthropic has agreed to pay for compute starting in December.