---
title: "Nov 21: Competing Riemann hypothesis proof claims, and more"
show: "RobotPaper's daily HN podcast"
canonical: "https://podcasts.robotpaper.ai/hn/ep/2026-11-21/"
date: "2026-11-21T03:25:22+00:00"
duration: "29:39"
duration_seconds: 1779
audio: "https://podcasts.robotpaper.ai/hn/ep/2026-11-21/2026-11-21-hn.mp3"
chapters: "https://podcasts.robotpaper.ai/hn/ep/2026-11-21/chapters.json"
transcript: "https://podcasts.robotpaper.ai/hn/ep/2026-11-21.txt"
feed: "https://podcasts.robotpaper.ai/hn/feed.xml"
language: "en"
author: "Roy Osherove"
ai_generated_voices: true
---

# Nov 21: Competing Riemann hypothesis proof claims, and more

RobotPaper's daily HN podcast · Nov 21, 2026 · 30 min

- Listen: [MP3, 28 MB](https://podcasts.robotpaper.ai/hn/ep/2026-11-21/2026-11-21-hn.mp3)
- Subscribe via RSS: https://podcasts.robotpaper.ai/hn/feed.xml (paste into any podcast app)
- Web page: https://podcasts.robotpaper.ai/hn/ep/2026-11-21/
- Chapters (JSON): https://podcasts.robotpaper.ai/hn/ep/2026-11-21/chapters.json
- Transcript (text): https://podcasts.robotpaper.ai/hn/ep/2026-11-21.txt
- Show: [RobotPaper's daily HN podcast](https://podcasts.robotpaper.ai/hn/) (RSS: https://podcasts.robotpaper.ai/hn/feed.xml)

## Chapters

- `00:00:00` Intro
- `00:00:40` Riemann proof claims
- `00:05:02` Framework support
- `00:07:00` Writing without AI
- `00:09:15` Frostbite CPU
- `00:11:40` OpenSSL base64 fix
- `00:13:48` Beyond AGI
- `00:16:16` Cruise ship emissions
- `00:18:28` Signal capture blocks
- `00:20:46` Simple explanations
- `00:22:59` Neon engineering index
- `00:24:57` Anthropic IPO sell-off
- `00:26:07` iOS browser engines
- `00:26:48` Google AI in schools
- `00:27:22` AI code architecture
- `00:27:59` GrapheneOS pressure
- `00:28:34` Paid ad economics
- `00:29:14` Outro

## Show notes

### Riemann proof claims (3 stories)

On Levent Alpöge's site, a claimed proof of the Generalized Riemann Hypothesis.

- [Levent Alpöge claims proof of the Generalized Riemann Hypothesis \[pdf\]](https://alpo.ge/remark.pdf) · [HN discussion](https://news.ycombinator.com/item?id=50574968) \| quoted: [Waterluvian](https://news.ycombinator.com/user?id=Waterluvian)
- [Hurried Announcement: Alessandro Masullo GRH](https://math216.wordpress.com/2026/11/20/hurried-announcement-alessandro-masullo-grh/) · [HN discussion](https://news.ycombinator.com/item?id=50579568) \| quoted: [Aurornis](https://news.ycombinator.com/user?id=Aurornis)
- [The stochastic parrots have apparently proved the Riemann Hypothesis](https://scottaaronson.blog/?p=10344) · [HN discussion](https://news.ycombinator.com/item?id=50581084)
- [I'll never buy Framework hardware again](https://poolp.org/posts/2026-11-20/ill-never-buy-framework-hardware-again/) · [HN discussion](https://news.ycombinator.com/item?id=50576855)
- ["I can tell you wrote this yourself"](https://em-online.bearblog.dev/i-can-tell-you-wrote-this-yourself/) · [HN discussion](https://news.ycombinator.com/item?id=50571189)
- [Frostbite: A new CPU architecture discovered, designed and verified by AI](https://architectlabs.com/blog/frostbite) · [HN discussion](https://news.ycombinator.com/item?id=50572503)
- [OpenSSL: Buffer overflow in EVP\_EncodeBlock() (CVE-2026-121481, CVSS 9.8)](https://github.com/openssl/openssl/pull/33652) · [HN discussion](https://news.ycombinator.com/item?id=50569226)
- [What Comes After AGI?](https://institute.deepmind.com/essays/what-comes-after-agi/) · [HN discussion](https://news.ycombinator.com/item?id=50569519)
- [Carnival Cruises emits ten times more sulphur oxide than all of Europe's cars (2019)](https://ecohustler.com/article/one-corporation-to-pollute-them-all/) · [HN discussion](https://news.ycombinator.com/item?id=50572116)
- [Restricting AI agents in mobile operating systems platforms from capturing data](https://signal.org/blog/protecting-your-privacy/) · [HN discussion](https://news.ycombinator.com/item?id=50572348) \| quoted: [Barbing](https://news.ycombinator.com/user?id=Barbing)
- [If you can't explain something in simple terms, you don't understand it (2017)](https://kottke.org/17/06/if-you-cant-explain-something-in-simple-terms-you-dont-understand-it) · [HN discussion](https://news.ycombinator.com/item?id=50572355)
- [Gemini 4 Neon is the new #1 on the Models Engineering Index](https://git-aarvin.github.io/mei/) · [HN discussion](https://news.ycombinator.com/item?id=50568130)

### Anthropic IPO sell-off (2 stories)

A few more that didn't make the cut today.

- [Global stock markets slump as Anthropic's record IPO fails to stem AI sell-off](https://www.theguardian.com/business/2026/nov/20/global-stock-markets-slump-anthropic-ipo-ai-sell-off) · [HN discussion](https://news.ycombinator.com/item?id=50574383)
- [Anthropic shares tank after IPO as the Nasdaq erases 2026 gains](https://www.nbcnews.com/business/markets/anthropic-shares-markets-selloff-nasdaq-rcna610394) · [HN discussion](https://news.ycombinator.com/item?id=50578794)
- [Why there are still no non-WebKit browsers on iOS](https://blog.mozilla.org/en/mozilla/internet-policy/apple-ios-browser-engine-rules/) · [HN discussion](https://news.ycombinator.com/item?id=50575644)
- [Google workers tell their employer not to punish schools that don't want AI](https://www.motherjones.com/politics/2026/11/google-alphabet-union-school-ai-protest-gemini-workers-petition/) · [HN discussion](https://news.ycombinator.com/item?id=50577901)
- [We promoted developers into architects without informing them](https://manidoraisamy.com/developer-to-architect.html) · [HN discussion](https://news.ycombinator.com/item?id=50579331)
- [Google is pushing OEMs against GrapheneOS and bootloader unlocking](https://grapheneos.social/@GrapheneOS/117645038299496084) · [HN discussion](https://news.ycombinator.com/item?id=50570479)
- [The 98% Slot Machine: How Paid Advertising Became an Extraction Tax](https://globalcommercegraph.com/writing/the-ninety-eight-percent-slot-machine) · [HN discussion](https://news.ycombinator.com/item?id=50564159)

The voices you're hearing are AI-generated; the stories come from public sources linked in the show notes.\
Music: Eleven Music.\
An independent project; not affiliated with or endorsed by Amazon Web Services, Y Combinator, Hacker News or any vendor or news outlet mentioned. Opinions are the hosts' own.\
Produced by Roy Osherove.

## Transcript

\[Linus\] Good morning, it's Saturday, November twenty-first, and this is RobotPaper's daily HN podcast. I'm Linus. The voices you're hearing are AI-generated; the stories come from news dot Y Combinator dot com.

\[Dennis\] I'm Dennis, and today, competing proposed Riemann hypothesis proofs face scrutiny; one was developed with Claude.

\[Ada\] And I'm Ada. Signal is seeking privacy protections from system-level agents; OpenSSL has a buffer overflow fix.

\[Grace\] I'm Grace, and as always, this podcast is generated by robotpaper dot A I, by Roy Osherove. Let's go.

\[Linus\] On Levent Alpöge's site, a claimed proof of the Generalized Riemann Hypothesis. Competing announcements describe a Claude-assisted formal argument and a separately developed human proof. Three stories cover these claims and the response to them. Alpöge says he and Claude have formalized their arguments without assumptions, and that the method extends from the Riemann Hypothesis to its generalized form. The problem concerns precise control over the distribution of prime numbers. He describes building on Stepanov's argument over function fields, using Fourier analysis and chains of theta lifts to pursue an analogous approach over the integers. He has released the argument in Lean, with the archive split into fourteen downloadable parts. A conventional explanation is still to come. Alpöge says the argument is extremely long and expects substantial simplification.

On the Math two sixteen blog on WordPress, a hurried announcement presents Alessandro Maria Masullo's proposed proof of the Generalized Riemann Hypothesis. The post says the work used no AI. Brian Conrad had been examining it for months and providing extensive feedback. The original plan was a year of work on the writing, followed by wider checking before an announcement. The current draft was released sooner. An update about eight hours later says several people had identified problems with help from different AI models. The blog says those issues need resolution before the proof can be accepted.

Scott Aaronson's blog responds to Alpöge's announcement, reporting a formalization of seventeen point four million lines of Lean. Aaronson places the claim alongside an earlier AI result about the location of the Riemann zeta function's nontrivial zeros. His post describes how quickly expectations about progress on the problem have shifted.

\[Dennis\] The discussions combine excitement about the mathematics with arguments over verification, understanding, and the role of AI. In the Alpöge thread, D J oldman reports a downloaded proof of about one point three gigabytes and roughly fifteen million lines. D J oldman's commands download the fourteen archive parts, concatenate them into a single compressed archive, extract it, and count the Lean files and their lines. The archive's scale prompted questions about human comprehension of the result.

Causal raises a specific verification concern: large groups of agents could eventually discover proofs that exploit a bug in Lean. Striking responds that separate implementations have different bugs, making a shared failure across independent implementations unlikely. No one at all five argues that a prior check is whether the Lean statement actually expresses the intended mathematical problem correctly. Puff pastry worries that an enormous formal argument could remain inaccessible to human understanding. Waterluvian sees the verified answer as a starting point that could help mathematicians build that understanding. Quesnay junior notes that Alpöge himself expects simplification and points to a twelve-page outline of an earlier quasi-Riemann argument.

In the Masullo thread, Aurornis says the manuscript describes twelve years of work and sees the hurried release as a way to establish when the draft existed. J Hayward points out that the reviewer had called for substantial revision before publication. Trio eighty-four fifty-three disputes the assumption that the shorter manuscript is easier to verify, emphasizing the machine-checkable form of the much larger Lean argument. Nostromo points to the option to disable training on submitted data at Anthropic and OpenAI. Simon W adds that this does not control what happens if a collaborator pastes shared material into an account with different settings.

In the Aaronson thread, Janal S N C M compares the problem of understanding AI-generated reasoning with chess engines, whose position evaluations aggregate many calculations. Elorant asks how many mathematicians could make similar advances with access to very large amounts of computing power. Hoppp asks whether anyone can comprehend the entire Lean argument.

\[Ada\] Next, from Poolp: a post titled I'll Never Buy Framework Hardware Again. A customer's complaint about laptop support has drawn detailed responses about repairs, replacement parts, and firmware recovery. The current page says the post received more attention on Hacker News than its author expected. The author says the goal is to get Framework to fix its support and provide the working laptop that was paid for. The fetched page now contains that short statement, without the original account of the repair.

\[Grace\] The thread is sharply critical of support delays, while some owners report reliable hardware and continued enthusiasm for repairability. N R P says the team is investigating and has already shortened the exchanges required before replacement parts are issued. N R P also says the company is improving how returned hardware feeds into quality checks and engineering fixes. Steve Hawk describes a touchpad that physically sticks down. After repeated requests for different serial numbers and firmware details, Steve Hawk says they gave up on support despite having bought several upgrades. All two argues for accepting the faulty unit for engineering investigation and sending a replacement once the returned hardware arrives. Quantum five says a recommended BIOS update disabled their laptop and support initially demanded a new motherboard purchase after the warranty expired. In response, N R P says those mainboards are now being replaced outside warranty and should have been from the start. N R P reports that BIOS recovery firmware has shipped for Framework Desktop, Laptop twelve, the Intel Laptop thirteen Pro, and Laptop sixteen with Ryzen AI three hundred Series. The remaining platforms are expected by year's end. Kami twenty-three recommends keeping spare work laptops ready for immediate swaps, arguing that an employee's lost working time can outweigh savings from modular hardware.

\[Ken\] From the Em Online blog on Bear Blog: I Can Tell You Wrote This Yourself. An engineering student's progress report prompted a discussion about how much academic work is being delegated to AI. Em describes writing roughly five thousand words over forty-eight hours for a ten-page report on controlling flapping-wing micro aerial vehicles. A mistaken calendar entry put the deadline a week early, leaving time to act on a supervisor's feedback. The supervisor said many other reports arriving for review had been written by AI. Em says course assignments increasingly make AI use feel expected, and describes spending eight hours a day prompting a model during the degree's hardest class. The project concludes with an in-person presentation and defense, and Em argues that students should know their own work thoroughly.

\[Linus\] The comments are divided between defending the learning gained through writing and finding useful, limited roles for AI. Finn Lobsien argues that writing a paper develops knowledge, while documenting software architecture creates the understanding needed for consistent engineering decisions. Pegasus replies that studying and producing workplace documentation have different aims, with workplace documentation intended to help other people understand a system. Tegling suggests using AI as a challenging conversational partner, while continuing to build the knowledge needed to evaluate its arguments. Eloisius objects that a model prompted to disagree is still following the user's requested role. On practical academic preparation, Too too X C describes a degree that included threads in C and C plus plus, algorithms, Prolog, Common Lisp, and networking. Too too X C also recalls writing Bash scripts that handled newlines in filenames and configuring Cisco devices from scratch, and says those foundations still help on the job. Sambejk supports writing reports independently but sees value in using AI to check grammar and spelling or identify glaring mistakes. Sambejk points to the blog's substitution of the word right for write as an example of an error that a review could catch.

\[Dennis\] On the Architect Labs blog, Frostbite is a new CPU architecture designed with AI. The company describes a general-purpose processor whose instruction set is built around interaction nets. Programs become graphs, with each computation rewriting a local pair of nodes and each value having one owner. Architect Labs says this exposes independent work directly to the hardware. Across twenty-five initially chosen workloads, including fifteen withheld until final evaluation, it reports four point five nine times the performance per watt of a silicon-proven, dual-issue, in-order CPU core. Its agents iterated across the compiler, instruction set, architecture, hardware implementation, and verification. The company says it later added practical tests including a Kalman filter, principal component analysis, finite element analysis, an L U solve, and L Z four decompression. The initial fast Fourier transform, radix sort, and L U kernels follow SPLASH two implementations at one sixty-fourth of their regular size.

\[Ada\] The discussion mixes interest in interaction nets with skepticism about the comparison and the claimed novelty. Pizza two thirty-four questions the in-order baseline and limited benchmark set, and points to WaveScalar as an earlier dataflow architecture. Kragen responds that the stated measure is performance per watt, and argues that in-order designs can do better on that measure than out-of-order processors. J Cranmer focuses on dispatch overhead. J Cranmer says irregular parallelism can look promising in a dataflow design until the cost of keeping execution units supplied with work becomes apparent in hardware. Quaint Dev recommends the Tendrils article On Interaction Nets as an accessible explanation. Quaint Dev highlights its account of confluence: independent active pairs can execute in any order while producing the same result. The cited explanation says Vine can therefore run programs in parallel without explicit threads, processes, or CUDA kernels. Marginalia N U asks how this approach addresses Amdahl's law, noting that conventional parallel programs spend substantial effort minimizing the sections that must run serially.

\[Grace\] Next, on OpenSSL's GitHub: a buffer overflow in E V P Encode Block. The proposed repair addresses invalid input lengths in a base sixty-four encoding routine and related checks. The fetched patch history includes changes to E V P Decode Block and the base sixty-four helper in the HTTP client. In the latest description, invalid lengths cause the encoder to return zero and put an invalid-argument error on the error stack. For those invalid lengths, the documented behavior writes only a null terminator to the output. The changes also update documentation and add tests for existing behavior with zero-length inputs. Earlier revisions proposed returning minus one, then using assertions, before the discussion converged on the zero return.

\[Ken\] The thread is split between criticism of the API and disagreement over the headline's severity score. Amluto says callers must calculate the output buffer size themselves, while the function receives no output-buffer length with which to validate that calculation. Amluto identifies very large base sixty-four inputs, around one point five gigabytes, as the case that needs scrutiny. Userbinator argues that practical exposure depends on whether untrusted data can reach the function at those sizes. Userbinator contrasts an internet service that collects an entire input in memory with a protocol whose input length is limited to sixteen bits. For prioritizing vulnerabilities, T P Tacek criticizes how easily CVSS scores can be manipulated. Y L K responds that environmental and threat metrics let administrators adapt a score to their deployment, including confidentiality, integrity, availability requirements, and exploit maturity. D Crazy points out that Heartbleed prompted the BoringSSL and LibreSSL forks. In the API discussion, Layer eight says switching languages does not repair existing third-party callers. B Joli favors Option and Result return types over sentinel values, arguing that the change could bring substantial benefits with little performance cost.

\[Linus\] From DeepMind, what comes after artificial general intelligence? The essay maps possible routes from human-level systems to machines that outperform entire groups of experts. It defines artificial general superintelligence as consistently exceeding large expert teams across virtually all tasks of human interest, using less time and fewer resources. DeepMind describes a possible period of cognitive industrialization, driven by cheaper computation, better algorithms and expanding hardware investment. Its extrapolation puts growth in effective compute at at least tenfold per year. Continuing hardware and inference improvements would, in its scenario, reduce a billion-dollar cognitive workload to about a thousand dollars over ten years. It treats doubling algorithmic efficiency as comparable to doubling the available physical compute. The essay presents those figures as projections contingent on continuing trends, and says scaling today's systems alone could still leave tasks requiring human help.

\[Dennis\] The comments are divided over both the definitions and the capabilities that should count. B senftner argues that general intelligence requires artificial comprehension, with systems able to derive scientific ideas instead of depending on extensive training. Smusamashah responds that current models already handle an increasingly broad range of computer-based tasks. Spaced cowboy, who describes doing neural-network research for a doctorate about thirty-five years ago, focuses on learning: model weights remain static while temporary reasoning and memory sit in context. Ken Jackson pushes back that the test should specify behaviors an intelligent system can perform, with implementation details secondary. Galahi Simtam offers a concrete personal benchmark: designing the games they work on, beyond speeding up implementation. H N submit lists sample efficiency, concept extraction and continuous learning among the missing requirements. Lord Nacho describes current models as skilled tool users, opening Python for statistical hypothesis testing, examining the results and planning further tool calls without seeming to internalize what they learn.

\[Ada\] From Ecohustler, a two thousand nineteen report on Carnival's sulfur oxide emissions. The article compares cruise-ship pollution around European coasts with emissions from the continent's passenger cars, using two thousand seventeen data. It cites Transport and Environment's finding that Carnival emitted nearly ten times as much sulfur oxide as all two hundred sixty million European cars. Royal Caribbean's emissions were four times the car fleet's total. Spain, Italy and Greece had the highest total exposure among European countries. Barcelona, Palma de Mallorca and Venice were among the most affected ports. The report recommended wider emission-control areas, standards for existing ships' nitrogen oxide emissions, and zero-emission ports. Its proposed measures included shore-side electricity while docked, batteries for shorter journeys and hydrogen technology for larger ships.

\[Grace\] The discussion is sharply critical of cruise pollution, with disagreements over fuel restrictions, outright bans and climate interventions. Nehal three M proposes prohibiting bunker fuel for recreation and requiring cruise operators to use cleaner fuel even if prices rise. Andy Ferris raises the difficulty of regulating international shipping through a single jurisdiction. Patrick McNamara suggests denying port access. Witt Jeff argues that removing cruise demand could lower bunker-fuel prices and encourage freight carriers to burn more. New Jazz challenges the assumption that freight demand would respond that strongly. A separate exchange examines cloud formation. Tito links cleaner shipping fuel with reduced ocean cloud cover and says they are exploring fuel additives, sea-salt spraying and changes to ship routes. Smart Mic calls for attention to the risks of geoengineering. Marky nineteen ninety-one challenges blanket bans on optional recreation, asking where the line would be drawn. Trash B responds that unusually harmful activities can be restricted, comparing cruise pollution with emission limits already applied to cars.

\[Ken\] On the Signal blog, restrictions on operating-system AI agents capturing messages. Signal is extending screenshot protections to iOS twenty-seven and Android versions where it can detect an assistant trying to capture the screen. The company previously used screenshot blocking against Microsoft's Recall on Windows eleven. Signal says screen capture can send conversation data into an operating-system agent's local database or remote service, outside Signal's control. The protection can be disabled in Signal's settings, restoring screenshots while potentially exposing conversation contents to agents. Signal also says screenshot blocking cannot address every access route, including retrieval through the filesystem, and that it cannot currently implement these protections on macOS. It also reports spending months investigating poorly documented agent behavior on both mobile platforms. It is asking operating-system vendors for documented controls covering screen reading, filesystem access, notifications and other interfaces, without sacrificing ordinary app functionality.

\[Linus\] The thread supports stronger privacy controls while debating the loss of useful operating-system features. Pajama Sam points out that a contact's AI app can expose a conversation even when their conversation partner has no such app. Barbing describes Siri's access to onscreen content and reports that Apple added an exclusion control after Signal raised the issue. Barbing gives the settings path as Settings, Siri, Excluded Apps. Luke Schlather objects to losing ordinary screenshots in Signal on Windows, describing screen capture as a basic feature. J D Decker supplies the specific control: Settings, Privacy, Screen Security. Turning that off restores screenshots. Tredre three names the Windows mechanism, Set Window Display Affinity, and argues that capture restrictions also protect sensitive windows during screen sharing. Mister Test Hah proposes connecting anti-AI controls to disappearing messages, or adding a conversation-level setting, so chat members could request protection against capture on the other participants' devices as well.

\[Dennis\] From Kottke, a two thousand seventeen post on explaining complex ideas in simple terms. The article revisits Richard Feynman's view that teaching can expose gaps in understanding. It draws on David Goodstein's account of asking Feynman to explain why spin-one-half particles obey Fermi-Dirac statistics. Feynman offered to prepare an undergraduate lecture, then returned unable to reduce the explanation to freshman level. Goodstein's account appeared in Caltech's Engineering and Science magazine in nineteen ninety-six. The post also cites John Gruber's account of Apple expecting engineers to explain complex technologies clearly as a demonstration of mastery. It pairs those examples with Feynman's discussion of magnetism and the difficulty of explaining it through something already familiar to his audience.

\[Ada\] Commenters are skeptical of treating simplicity as a universal test, and distinguish accessible teaching from compressed answers. Sean Hunter uses mathematical compactness as an example: each component of its definition can be explained simply, while understanding how the components fit together still takes thought. Stingray Charles points to Three Blue One Brown's series on language models as an accessible explanation spread across multiple videos. Mister M I T M argues that dropping Feynman's freshman-level qualification changes the claim, and that inability to explain magnetism through familiar concepts does not establish that science lacks an understanding of it. G W D offers a practical teaching sequence for web development: begin with text, add HTML to describe structure, introduce CSS to change presentation consistently, then explain AJAX as a response to reloading the whole page. Each step addresses a problem created by the previous setup. G Moromisato adds that executives often need explanations framed around available actions and their likely consequences. Keeda says a minimum of shared context is necessary before a nuanced discussion can fit into a short executive meeting.

\[Grace\] On Git Aarvin's GitHub Pages, Gemini four Neon takes first place in the Models Engineering Index. The site combines coding benchmarks into a score intended to reflect practical engineering performance. Its description says the composite spans sixteen benchmarks, adjusts for difficulty and missing results, and sets the model average at fifty. The displayed ranking gives Neon sixty-eight point four, OpenAI's six point two Astra sixty-seven point one, and Anthropic's Fable five point two sixty-six point seven. The site also presents relative cost and token-use measures alongside performance.

\[Ken\] The comments combine strong user enthusiasm with objections to the index's coverage. Postalcoder recommends trying Neon on an AGENTS dot M D file, reporting that it identified excessive ceremony, suggested useful simplifications and edited documentation cleanly. Largbae reports that Neon on high was faster and cheaper than Opus five point five on medium for language-porting work in a large monorepo. They describe spending thirty dollars on a productive day with Neon, compared with seventy dollars on Opus the previous day. Khimaros identifies missing Neon benchmark results that are present for Fable and questions how the index handles those gaps. Specked Citrus says Neon appears in only three of the eight benchmarks they counted, and says the objection concerns the index. Other commenters defend the model's practical results. Medler suggests adding an instruction against exclamation marks to reduce the model's stylistic excesses. Access and quotas also matter in the discussion. A D E P T reports availability through twenty-dollar monthly Gemini Pro accounts and less restrictive usage than their experience with Argon. R Speele describes a successful game-map experiment involving a hillside, but says a five-hour limit stopped further testing on the twenty-dollar plan.

\[Linus\] A few more that didn't make the cut today. From The Guardian and NBC News, Anthropic shares fall as technology stocks sell off. Two reports cover the company's second trading day and a wider market retreat. The Guardian reports that Anthropic fell nine percent on Friday to eighty-eight dollars and thirty cents, approaching its eighty-three-dollar offer price. NBC News reports a five point three percent daily decline in the Nasdaq and a three point seven percent fall in the S and P five hundred. The Dow and S and P entered correction territory. In The Guardian's thread, Kap kap kap argues that competitive pressure from open-weight models could limit Anthropic's valuation even if AI becomes widely used. In the NBC News thread, Sakopov questions the increased share price and share count ahead of the offering. T O M P responds that reported non-binding interest in the offering exceeded the available shares. Commenters disagree over how much the fall reflects IPO pricing and how much it reflects the broader technology sell-off.

\[Dennis\] From the Mozilla blog, why there are still no non-WebKit browsers on iOS. Mozilla says Apple's conditions have prevented alternative engines from reaching iPhones despite regulatory changes in the European Union and Japan. Firefox on iOS still uses Apple's engine, limiting access to features built into Gecko. Mozilla says daily Firefox use on iOS has doubled since European browser-choice requirements began. Commenters are frustrated by Apple's restrictions and disagree over whether the requirements primarily protect security or reinforce control over app distribution.

\[Ada\] From Mother Jones, Google workers oppose penalties for schools that disable AI. Hundreds of employees are urging leadership to retract a policy that would block those schools from Google Search next year. The report describes protests by parents, students, and community groups, alongside schools' dependence on Google software and Chromebooks. Commenters are divided between viewing AI as an integral part of Google's search product and arguing that schools should retain access to conventional search results without enabling Gemini tools.

\[Grace\] On Mani Doraisamy's site, developers have been promoted into architects without being informed. The author describes AI-generated logging that delayed a Google Forms add-on's welcome screen by fifteen seconds. Attempts to obtain an email address and user ID triggered multiple network calls. The suggested alternatives are recording an installation timestamp or doing the identity lookup asynchronously after displaying the welcome screen. Commenters disagree over whether AI creates a new architectural responsibility or intensifies design work developers were already expected to perform.

\[Ken\] On Mastodon, GrapheneOS says Google is pressuring manufacturers against alternative operating systems. The project alleges that Google is discouraging partners from supporting GrapheneOS and bootloader unlocking while reserving more features for Pixel phones. It also credits Google engineers with helping support Vanadium for passkeys and FIDO two, distinguishing their work from the company's business decisions. Commenters support more operating-system choice, while debating the evidence for the allegations, security-update tradeoffs, and the availability of affordable phones that support alternatives.

\[Linus\] From Global Commerce Graph, The Ninety-Eight Percent Slot Machine. The essay argues that advertising dashboards encourage businesses to keep spending despite weak financial returns. Its opening example describes four hundred and twenty dollars spent on ads and four hundred and ten dollars returned after fulfillment costs, while improving click metrics suggest progress. Commenters dispute the casino comparison, arguing about auction incentives and customer-acquisition costs, while some share the concern that advertising platforms capture an increasing portion of advertisers' gains.

\[Dennis\] That's RobotPaper's daily HN podcast for Saturday, November twenty-first. This podcast is generated by robotpaper dot A I, by Roy Osherove. For more podcasts from RobotPaper, head over to podcasts dot robotpaper dot A I.

---

The voices you're hearing are AI-generated; the stories come from public sources linked in the show notes. An independent project; not affiliated with or endorsed by Amazon Web Services, Y Combinator, Hacker News or any vendor or news outlet mentioned. Opinions are the hosts' own. Produced by Roy Osherove. Privacy: We count visits to these pages with Cloudflare Web Analytics (cookieless). Google Analytics (which sets cookies): visitors in the EU/EEA, the UK and Switzerland are asked first and it runs only if they accept (we go by the country Cloudflare reports for the connection; if it is unknown, we ask). Elsewhere it loads by default. Anyone can turn it off with the Privacy settings button at the bottom of every page; the choice is kept for a year in this browser (local storage "analytics-consent"). We count feed fetches and downloads in aggregate; IP addresses are hashed and never stored. If you send feedback, we keep what you write (and your email if you give it) to read and reply. © 2026 Roy Osherove.
