Today’s Highlights#

The first story that landed was France’s ban on unsolicited sales calls, in force from August 11. A small rule about who may ring a home phone also framed the rest of the day: a fight over whether the internet must know your name, now spreading from Britain into 21 U.S. states, and a public-health report that England may be first to eliminate hepatitis C.

On the technical side, AI is changing how pages stay alive, how hidden model reasoning can leak, and which programming languages pair well with a machine that writes the first draft. Nvidia’s spending got compared to the 1873 railroad boom. Mark Twain’s decade inside a typesetting machine did the same job as a warning.

Tech and Products#

When machines do more of the making, what stays in human hands?

Hidden “thoughts” can be recovered: a portable hole in encrypted reasoning blocks#

Stolen Thoughts describes a surprise in some Anthropic, OpenAI, and Google APIs. Those services send the model’s hidden reasoning — the internal scratch work before an answer — back to the client as encrypted blocks, then accept the same blocks on later turns. The researchers found the blocks are reusable. Produce one with a stronger model, hand it to a weaker, jailbroken cousin in the same family, and that weaker model can be induced to write the hidden reasoning out in the clear. On 120 programming problems, the decoded length matched the vendors’ own hidden-thought statistics. From 6,708 public traces the team rebuilt 310,000 thought blocks. 704 held traces of real user situations, and 64 of those traces lived only in the thought block.

Note: This is a product choice as much as a crypto bug. HN argued over whether anything was “stolen.” One side said model output is not copyrighted, so the live issues are terms of service and compliance. The other used browser cookies as the analogy: even if the law is unsettled, other people’s session residue still hurts. Several readers said users pay for the reasoning, so the blocks should sit closer to the user, and that billing and review should not rest on ciphertext that another client can replay.

Discussion: Hacker News thread

If AI writes the code, Go looks easier to review#

A Google developer blog post says AI coding changes the scorecard from “nice to write” to “nice to review.” When a model can emit hundreds of runnable lines in seconds, people spend their time reading, correcting, and living with the result. Go, the post argues, was built for teams and long lives: a small grammar, one formatter, tests in the box, stable compatibility, a complete toolchain. That makes it a better partner for an AI. The authors cite teams at Netflix and elsewhere who say models behave more predictably in Go, and note that go fix, the AST, and the module tools make large refactors less painful.

Note: The community was practical. Many agreed that “designed for collaboration” is the right frame, but said formatting is surface. The real help is static types, fast compiles, and “there is one way to write it,” which makes model mistakes show up sooner. Others noted that “Go works better with AI” is still mostly anecdote. Whether it fits still depends on the stack a team already has and what it needs to maintain.

Discussion: Hacker News thread

Sharper zoom without generative fill: an iPhone app that fires two lenses at once#

Photosynthesis, from an independent developer, tries to improve zoom with optics and computation, not invented pixels. One shutter press fires two cameras, each with a different view and level of detail, then an algorithm fuses them into a higher-resolution still. The site stresses that it does not use generative AI to hallucinate texture. Comparison shots claim cleaner distant text and book spines. The upgrade path, as sold, is a better algorithm rather than a new phone.

Note: Curiosity and caution sat next to each other. Plenty of readers liked the choice to enhance from real optical data, especially after years of beautified phone photos. Others said the hard parts are registration, alignment, and color, and that low light and motion still need independent samples. People wanted a more systematic bake-off, not just the landing-page pairs.

Discussion: Hacker News thread

Business and Platforms#

A high-spend cycle, held up to an older boom.

Nvidia’s spending, seen in an 1873 mirror#

Stratechery’s Nvidia’s Risky Business uses the 1873 U.S. railroad bubble as a glass for today’s AI capital expenditure. The piece follows financier Jay Cooke as he sold railroad bonds to the public at scale, then watched the structure fail after the Vienna crash and a credit squeeze. Citing the book 1873, it converts about $500 million a year of railroad investment then into roughly $600 billion in today’s economy. Against that, Microsoft still funds spend from free cash flow and looks restrained. Paths that lean on debt look riskier. The test of a bubble, the author says, is not the size of the check. It is where the money comes from and whether it can last.

Note: HN pulled the history back to moats. Many comments said Nvidia’s wall is not the GPU itself but CUDA — Nvidia’s programming platform — and the optimized stack around it. Alternatives exist; matching performance, compatibility, and community at once is harder. On Google’s TPU, a chip built for AI math, the thread granted architecture and interconnect strengths but doubted a consumer-facing, open stack that could copy CUDA’s network effects.

Discussion: Hacker News thread

Policy and Governance#

Rules that reach a phone, a login, and a child’s account.

Le Monde reports that a law pushed by 11 consumer groups and passed last year takes effect on August 11. Firms may not make marketing calls without prior consent. Fines run to €75,000 per call for an individual and €375,000 per call for a company. People can report violations on an official site. The government presents it as an answer to years of complaints, aimed at grinding sales pitches and at people who are easy to con. France used to rely on a “do not call” list. That filtered poorly. The new rule flips the default: the caller must show consent, instead of the target having to opt out.

Note: The thread cheered, then asked how it lands. Highly voted ideas included a whitelist for hospitals and public services, default blocking of unknown numbers, and system-level filters that transcribe the call. Another group said a total ban cuts off honest small businesses. Most replies answered that people who want to buy will find you, and that a cold interrupt is not a legitimate channel.

Discussion: Hacker News thread

A war on anonymity, exported from Britain to America#

The investigation The UK’s war on anonymity has come to America says five Britain-linked NGOs and their U.S. arms are pushing digital identity bills in 21 states and in Congress, using child safety as the frame. The author ties them to rules such as the UK Online Safety Act, and walks through funding and lobbying at groups including 5Rights and CCDH. The charge is that some register as foreign agents while still shaping U.S. bills through American proxies, with digital ID as the tool that ends anonymity. Copy the British model, the piece warns, and ordinary browsing starts to require a credential. A check on power gets thinner.

Note: The comments split hard. Supporters of statewide rules said household filters cannot beat peer pressure and product design, and wanted something closer to alcohol and tobacco controls. More of the highly voted replies said a blunt age cut is a poor fit and makes minors easier to tag. They preferred parental controls over a universal ID check. Several people said age is a convenient, coarse proxy, and that the real fight is who draws the line and under what checks.

Discussion: Hacker News thread

Science and Research#

A quiet disease, found earlier and cured more often.

England may be first to eliminate hepatitis C#

The BBC says England is in reach of becoming one of the first countries to eliminate hepatitis C. NHS figures show 80 percent of known cases treated, and related deaths down about 36 percent over ten years, close to a 2030 goal. An 8-to-12-week course of oral antivirals cures more than 95 percent of patients. Emergency-room blood draws, GP screening, and free home tests are finding people who never knew they were infected. The virus spreads mainly through blood. Early infection is often silent and can wreck the liver if ignored. Blood products are already screened. The remaining work is finding cases and treating them in time.

Note: HN did not stay on the English playbook. The thread slid into vaccines and public trust. Some compared the progress with a return of preventable disease elsewhere, and others warned against pinning the cause on one politician. The nearer consensus was that a treatable, preventable illness is eliminated only if screening money and trust in clinics hold. Policy whiplash and bad information cut that chain.

Discussion: Hacker News thread

Society and Culture#

What happens to the web when the answer never needs a click, and what a typesetter did to Mark Twain.

If AI answers first, who maintains the page it cited?#

A Walrus feature argues that generative AI turned search into “just give me the answer.” Users stop clicking sources. Newsrooms, blogs, and wikis lose ad and subscription money, update less, and link rot gets harder to notice. Everyday search, the author says, used to fund a loop: people visit, the site earns, the site writes again. If models consume pages and never send a reader back, the shared memory thins. The piece does not ask for the old search box back. It asks what new incentives and archives should look like.

Note: HN split in two. One group described using a model to configure a router or read a log, and treated fewer clicks as progress. The other worried about the supply side: much of the web was written by people who expected visitors or reputation. If the remaining traffic is training bots, who keeps a careful page alive? Some said vendor docs will still be updated, and that ad-funded independents get hit first. Others wanted AI products to send traffic and money back to writers.

Discussion: Hacker News thread

The “mechanical miracle” that ate ten years of Mark Twain#

Res Obscura retells Twain’s entanglement with the Paige compositor. Twain had been a printer’s apprentice. After meeting inventor James Paige in 1880, he believed a machine that could set type automatically would replace labor and pay a fortune. He put in a sum on the order of $10 million today, including his wife’s money. The piece quotes Twain’s awe and his biographers’ later verdict: Paige was a better fundraiser than a steady engineer. The machine was clever and never reliably mass-produced. After a decade, Twain was bankrupt and depressed, and his best writing years had been spent on the machine. It is a clean case of a tech craze eating a life.

Note: The comments skipped gossip and asked how the craft actually worked. Readers wanted schools to teach the chain — type metal, ink, the composing stick, the press — instead of a one-line Gutenberg. Others described small printing museums and said an “invention” is usually dozens of details. On whether school can teach that, the thread split between people who think curiosity plus today’s tools is enough, and people who said unguided study stays shallow without standards and feedback.

Discussion: Hacker News thread

Closing#

From a sales call that may no longer be dialed at will, to a fight over showing ID to browse, half of today’s stories were not about writing code. They were about how tech sits inside ordinary life. AI makes answers faster, code more plentiful, and photos sharper. It also makes the questions sharper: who maintains the page, who is liable, who keeps the memory. Hepatitis C moving toward elimination, and Twain’s lost decade, sit at opposite ends of the same warning. Tech can unpick a public problem a little at a time. It can also spend a person down when the enthusiasm runs too hot. Borders, trust, and patience still need recalibrating. See you next time.