---
title: "Nov 19: AI Overviews cut publisher traffic, and more"
show: "RobotPaper's daily HN podcast"
canonical: "https://podcasts.robotpaper.ai/hn/ep/2026-11-19/"
date: "2026-11-19T03:38:10+00:00"
duration: "29:34"
duration_seconds: 1774
audio: "https://podcasts.robotpaper.ai/hn/ep/2026-11-19/2026-11-19-hn.mp3"
chapters: "https://podcasts.robotpaper.ai/hn/ep/2026-11-19/chapters.json"
transcript: "https://podcasts.robotpaper.ai/hn/ep/2026-11-19.txt"
feed: "https://podcasts.robotpaper.ai/hn/feed.xml"
language: "en"
author: "Roy Osherove"
ai_generated_voices: true
---

# Nov 19: AI Overviews cut publisher traffic, and more

RobotPaper's daily HN podcast · Nov 19, 2026 · 30 min

- Listen: [MP3, 28 MB](https://podcasts.robotpaper.ai/hn/ep/2026-11-19/2026-11-19-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/2026-11-19/
- Chapters (JSON): https://podcasts.robotpaper.ai/hn/ep/2026-11-19/chapters.json
- Transcript (text): https://podcasts.robotpaper.ai/hn/ep/2026-11-19.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
- `00:00:36` AI search traffic
- `00:04:53` Nginx HTTP/2 flaws
- `00:07:26` Gemini 4 Neon
- `00:09:45` AI code review queues
- `00:12:07` Quad9 leaves France
- `00:16:05` Microsoft MAI models
- `00:18:37` Codeberg suspension
- `00:20:36` IRS identity terms
- `00:22:38` jQuery release
- `00:24:40` Energy drink reports
- `00:25:22` Deutsche Bank files
- `00:26:03` Cambridge word choice
- `00:26:37` KPMG audit trial
- `00:27:21` Poland sabotage report
- `00:27:57` Turbo Pascal size
- `00:28:32` Emacs web browsing
- `00:29:08` Outro

## Show notes

### AI search traffic (2 stories)

From Nicolas Bouliane's blog, AI is killing All About Berlin, and from Revscope, Google's AI answers are reducing search traffic by forty percent.

- [AI is killing All About Berlin](https://nicolasbouliane.com/blog/death-by-ai) · [HN discussion](https://news.ycombinator.com/item?id=50542705) \| quoted: [nicbou](https://news.ycombinator.com/user?id=nicbou)
- [Google's AI answers reducing search traffic by 40%](https://revscope.co/blog/search-seo-geo-2026/) · [HN discussion](https://news.ycombinator.com/item?id=50540178) \| quoted: [nicbou](https://news.ycombinator.com/user?id=nicbou)
- [Unauthenticated remote code execution in Nginx](https://www.wiz.io/blog/nginx-http2-rce-vulnerabilities) · [HN discussion](https://news.ycombinator.com/item?id=50547374)
- [Gemini 4 Neon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/) · [HN discussion](https://news.ycombinator.com/item?id=50552424)
- [Harvard researchers find AI coding agents shifted the bottleneck to code review](https://complexdiscovery.com/harvard-researchers-find-ai-coding-agents-shifted-the-bottleneck-to-code-review/) · [HN discussion](https://news.ycombinator.com/item?id=50538303)

### Quad9 leaves France (2 stories)

From Cloudflare's blog, Quad Nine's withdrawal from France, and from TechStackups, DNS resolvers and the fight against piracy.

- [Quad9's withdrawal from France: A canary in the coal mine](https://blog.cloudflare.com/quad9-withdrawal-from-france/) · [HN discussion](https://news.ycombinator.com/item?id=50538994) \| quoted: [Scaled](https://news.ycombinator.com/user?id=Scaled)
- [DNS resolvers are the wrong place to fight piracy](https://techstackups.com/articles/quad9-leaves-france-dns-blocking/) · [HN discussion](https://news.ycombinator.com/item?id=50545958)

### Microsoft MAI models (2 stories)

From Microsoft, new models at scalable prices, alongside MAI Code Two and Code Two Flash.

- [Seven new models at scalable price-performance](https://microsoft.ai/news/microsoft-ignite-2026/) · [HN discussion](https://news.ycombinator.com/item?id=50548021)
- [MAI Code 2 and Code 2 Flash](https://microsoft.ai/news/introducing-mai-code-2-and-mai-code-2-flash/) · [HN discussion](https://news.ycombinator.com/item?id=50535885)
- [Codeberg.org Not Reachable](https://social.anoxinon.de/@codebergstatus/117636437840688845) · [HN discussion](https://news.ycombinator.com/item?id=50544261) \| quoted: [Hackbraten](https://news.ycombinator.com/user?id=Hackbraten), [Tepix](https://news.ycombinator.com/user?id=Tepix)
- [ID.me terms required for IRS sign-in](https://consumerrights.wiki/w/ID.me_terms_required_for_IRS_sign-in) · [HN discussion](https://news.ycombinator.com/item?id=50540437)
- [jQuery 4.1.0](https://github.com/jquery/jquery/releases/tag/4.1.0) · [HN discussion](https://news.ycombinator.com/item?id=50537156)
- [Energy drinks linked to 11 times more reports of cardiac arrest](https://www.sciencedaily.com/releases/2026/11/261110042543.htm) · [HN discussion](https://news.ycombinator.com/item?id=50546465)
- [Inside Deutsche Bank](https://www.inside-db.com/) · [HN discussion](https://news.ycombinator.com/item?id=50547620)
- [The Cambridge Dictionary has named ‘human’ as the Word of the Year 2026](https://www.cam.ac.uk/news/cambridge-dictionary-reveals-word-of-the-year-2026) · [HN discussion](https://news.ycombinator.com/item?id=50546914)
- [AI agents ask better questions than KPMG accountants in audit trial](https://europeanbusinessmagazine.com/ai-agents-better-questions-kpmg/) · [HN discussion](https://news.ycombinator.com/item?id=50547777)
- [Russia reportedly planned mass-casualty attack in Poland to be blamed on Ukraine](https://www.yahoo.com/news/world/articles/russia-reportedly-planned-mass-casualty-090500574.html) · [HN discussion](https://news.ycombinator.com/item?id=50547344)
- [Things That Turbo Pascal Is Smaller Than (2011)](https://prog21.dadgum.com/116.html) · [HN discussion](https://news.ycombinator.com/item?id=50506894)
- [Emacs Is My Browser](https://joshblais.com/blog/emacs-as-my-browser/) · [HN discussion](https://news.ycombinator.com/item?id=50507553)

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 Thursday, November nineteenth, 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, AI answers are cutting traffic to independent websites.

\[Ada\] And I'm Ada. Google has a new Gemini model; researchers examine the cost of reviewing AI code.

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

\[Ken\] From Nicolas Bouliane's blog, AI is killing All About Berlin, and from Revscope, Google's AI answers are reducing search traffic by forty percent. These two stories examine how Google's generated answers are changing traffic to the websites that supply information.

Bouliane says All About Berlin has lost three quarters of its audience after seven years of growth. He has operated the site since twenty seventeen and supported himself from it since twenty twenty. His guides explain processes involved in moving to and living in Berlin. He writes that local and federal government offices refer people to them, and that Berlin's immigration office uses his data about its own operations. As visits decline, he is spending time preparing for another source of income instead of producing new guides. Other website owners have told him they have stopped updating their sites or put their work behind paywalls.

Revscope adds experimental evidence. Researchers Saharsh Agarwal and Ananya Sen randomly assigned American desktop Chrome users to see Google normally or have AI Overviews hidden. Across two weeks of searches in January and February, searches that would display an Overview produced an average of zero point three seven outside clicks when it was visible, against zero point six two when hidden. The researchers report a thirty-nine point eight percent reduction. Across all searches, including those without Overviews, their estimated loss was eighteen point five percent of organic clicks. The working paper covers desktop use and has not yet been peer reviewed. Clicks on advertisements did not change. Searches ending without an outside click increased from fifty-four percent to seventy-three percent. The report also puts the traffic sent by AI assistants at just zero point three percent in March. Revscope also reports that Overviews appeared on about sixty-one percent of DataForSEO's tracked keywords in August, compared with about six percent two years earlier.

\[Ada\] The discussions are divided between people who value immediate answers and publishers concerned about losing the audience that finances their work.

In the All About Berlin thread, nicbou identifies himself as Nicolas and says the original post is several months old. He can still support himself from the site, but says next year might be his last operating it full time. When throwaway twenty twenty-six eleven suggests offline promotion and a paid newsletter, nicbou says he already has a free, advertising-free newsletter. He explains that citizenship information attracts people when they need to apply, making a regular subscription audience difficult to build. Nicbou also says he works with a health-insurance broker and that commissions provide a sizeable part of his income.

Glimshe compares the disruption with websites and YouTube replacing small print magazines, including an arts-and-crafts magazine run by a family member. Bishonen eighty-eight disputes that comparison, arguing that Google is reusing the publisher's own information and asking where fresh information will come from if the publisher stops maintaining it.

In the Revscope discussion, nicbou gives a further update: his traffic is now down more than eighty percent. He says that gathering the information still requires tools, relationships and group chats. When man this sucks proposes paid tiers, nicbou rejects the idea, saying the information should remain free and widely available.

P carolan argues that direct answers remove the need to search through clickbait. Uptrails responds that even simple answers depend on somebody doing the original work, and that hobby communities also lose opportunities to connect. Cortesoft counters that people seeking a community may still follow the links, while many former visitors only wanted an answer.

N P Teljes says ChatGPT Plus handles many of their questions in High thinking mode, while DuckDuckGo remains their search engine. They describe its interface as resembling an earlier Google. Dan Hulton describes a practical alternative: paid search through Kagi. He says he blocks sites that repeatedly supply unwanted AI-generated material or paywalls, and that those exclusions improve his subsequent results.

\[Dennis\] From the Wiz blog, unauthenticated remote code execution in Nginx. Two memory-safety flaws can compromise an affected server when HTTP two is enabled, including a server delivering ordinary static files. Wiz Research, Ikotas Labs, Metnir, Rapid Seven and Xint independently reported the vulnerabilities to F Five. One is a use-after-free in reverse-proxy processing when available connection slots run low. The other is an integer underflow in flow-control accounting, present since version one point nine point five in twenty fifteen. The affected-version ranges listed in the article extend through one point thirty-one point eight. Wiz demonstrated code execution from the integer underflow over HTTPS on sixty-four-bit systems, without a backend application or third-party modules. The company says the flaws can expose several applications behind a shared Nginx deployment. Wiz says address-space randomization does not prevent exploitation because both flaws can expose memory before code execution. Its Atlas code-scanning harness discovered the bugs. Its mitigation guidance is to apply security updates, or temporarily disable HTTP two while an upgrade is arranged.

\[Grace\] The comments are divided between relief that longstanding flaws are being found and concern about the pace of emergency patching. Black three r says a fix appeared in the public repository a week before disclosure, and argues that an earlier notice could have told operators to disable HTTP two. Verst points out that this workaround would break a service built primarily from gRPC APIs. Verst says that service currently uses a different server, but the protocol dependency would still prevent simply switching HTTP two off. Weitendorf recommends Envoy for gRPC and HTTP two workloads, citing the infrastructure and security work contributed by large technology companies. T K Z thirteen twelve describes using Fil-C for network-facing C programs, including SSH and a static file server, and prefers retaining established code while making its memory handling safer. In the discussion about replacement servers, redox ninety-nine says they are moving to Caddy. D T F replies that Caddy had its own recent vulnerability involving Go's multipath TCP policy.

\[Linus\] Google's blog introduces Gemini four Neon. The new model targets software engineering, analytical work and science with a one-million-token context window and four thinking levels. Pricing is three dollars per million input tokens, fifteen cents for cached input, and fifteen dollars per million output tokens. Google says Neon scores sixty-one on the Artificial Analysis Intelligence Index at its extra-high thinking level. At the default high setting, the company reports a thirty-two percent lower cost per task than Carbon on that evaluation. Google also describes multi-file debugging, repository refactoring and long-running workflows in which Neon directs Hydrogen subagents inside Antigravity. On SWE-sweep, Google reports a score of twelve point eight percent with the mini-SWE harness, rising to thirty-nine percent in Antigravity. The rollout includes the Gemini API in Google AI Studio, Gemini for business, and paid Pro and Ultra access in Antigravity and the Gemini app. A separate cyber version is scheduled for trusted defenders through the Fairwind Program in the coming days.

\[Ken\] The thread mixes interest in faster execution with frustration over model names and disagreement about everyday quality. Stillpointlab says mapping Neon, Carbon, Sol, Astra and the other names to particular jobs has become exhausting. Seize the cheese reports seeing Neon served through OpenRouter at a hundred and twenty-seven tokens per second, compared with forty-six for Carbon, and says Carbon's speed has been a major limitation in practical tasks. Declan Jackson, who says they work at Artificial Analysis, points out that Carbon remains positioned for long-horizon work despite Neon's higher overall intelligence score. Jackson cites research-scale reasoning tests as an area where that distinction appears. Svara describes a concrete workflow using the existing Gemini four Argon model to write commit messages and pull-request descriptions delegated from Opus or Fable. Svara prefers Argon's English prose. Behnamoh asks how that arrangement handles usage limits when producing a good description requires supplying the diffs and the preceding work.

\[Ada\] From ComplexDiscovery, Harvard researchers find AI coding agents shifted the bottleneck to code review. A working paper associates agent adoption with more code and slower reviews, while completed work showed no detectable increase. Harvard doctoral candidates Fiona Chen and James Stratton examined seven hundred and eighteen companies, using engineering activity from Jellyfish and employment records from Revelio Labs. The sample runs from January twenty twenty-one through March twenty twenty-six. Firms adopting agents added about thirty percent more code, made twenty percent more commits and opened twenty-three percent more pull requests. Submission-to-merge time increased forty-nine percent, from a baseline of roughly seven days. Review comments per pull request rose thirty-five percent, and formal requests for changes became nearly twice as common. Finished work meant resolved Jira issues whose code had reached production. These estimates measure the additional effect of agents at firms already using AI coding assistants. The version discussed is a job-market working paper with no indication of completed peer review.

\[Dennis\] The comments divide between developers reporting effective agent workflows and reviewers describing more exhausting, overcomplicated submissions. Alkonaut, who reports twenty-five years of experience, says agents write and review code well, but worries that doing less coding is gradually eroding familiarity with the codebase. Good regex describes repeatedly correcting Claude over a factory pattern: first replacing static methods, then separating unrelated objects, then removing an unnecessary factory. Good regex attributes such mistakes to missing business or technical context and overemphasis on individual requirements. Little cranky sixty-seven says merge requests now contain more code and overdesigned solutions, increasing the review burden. Imron offers a local review tool called C R T. It lets a reviewer leave comments on code, then have an agent retrieve those comments with context over MCP and make corrections. The tool remembers what has already been reviewed and presents diffs for subsequent changes.

\[Grace\] From Cloudflare's blog, Quad Nine's withdrawal from France, and from TechStackups, DNS resolvers and the fight against piracy. These two reports cover the departure of a public DNS service after French copyright-blocking orders.

TechStackups says Quad Nine will stop answering queries from France on November twenty-fourth. The Swiss nonprofit operates the nine dot nine dot nine dot nine resolver, which filters malware and phishing domains. The report says a Paris court imposed a penalty after Quad Nine refused to block sports-streaming domains identified for beIN Sports. Around twenty similar cases remain pending. TechStackups says the orders can expand to include new mirror domains notified through Arcom.

The article distinguishes the requested penalty from the amount imposed. BeIN asked for ten thousand euros per domain per day, amounting to five hundred eighty thousand euros daily across fifty-eight domains. The court set a total penalty of one thousand euros a day, running until the tennis season ended on November fourteenth. The report says the court cited Quad Nine's existing global malware filtering when rejecting its argument about geographically targeted blocking.

Cloudflare says French government bodies, educational institutions, hospitals and utilities use Quad Nine. Because the service keeps no user records, it cannot directly warn those users that their resolver will disappear. Cloudflare also describes earlier OpenDNS withdrawals and a Belgian order seeking blocking by Cloudflare, Google and Cisco by early January twenty twenty-seven.

Both articles describe the limits of DNS enforcement. A resolver can withhold a domain's address, while the underlying server and its files remain online. Domain-level blocking also affects legitimate material sharing the same domain. TechStackups reports that Mullvad recently moved its browser's default encrypted DNS service to Quad Nine after retiring its own public resolver.

\[Linus\] The comments are sharply divided over infrastructure neutrality and how copyright enforcement can reach services outside a court's jurisdiction.

In the Cloudflare discussion, John H Todd identifies himself as Quad Nine's chief technology officer. He says the organisation opposes piracy but considers DNS blocking a source of fragmentation and distrust. He asks for long-form arguments about France's approach, with French-language contributions particularly welcome.

Goobatrooba asks what governments should do when the people running unlawful services keep moving or operate beyond their reach. Scaled argues that companies profiting from piracy should help limit harm once illegal material is identified. D Betteridge responds that affordable, convenient access reduces demand for piracy, pointing to Steam and to the fragmentation of streaming catalogues. Driverdan asks Cloudflare to state whether it would also leave affected countries if its appeals fail.

The TechStackups thread takes up the practical limits of enforcement. Bluecalm says prosecuting pirate operators becomes difficult when their jurisdictions will not cooperate, while cutting off entire countries would have much wider consequences. Skrebbel asks whether criticism should target judges or the law they are applying. Dolmen argues that the legislation created the broad power and that responsibility rests with the representatives who approved it.

Bemtevi says a better paid experience would mean downloadable movies priced at two to five dollars, without digital rights management. Bemtevi cites buying classical music downloads from Hyperion as an existing example. Wombat man disagrees about the inconvenience of current services, describing four-to-six-dollar rentals that arrive on a television in four K, without locating a torrent, downloading it and waiting for Plex to scan it.

\[Ken\] From Microsoft, new models at scalable prices, alongside MAI Code Two and Code Two Flash. These two announcements extend Microsoft's model range and detail its enterprise coding models.

The broader release includes security, coding, image and speech models, plus model routing. Microsoft says MAI Cyber Two Flash's output tokens cost eighty-six percent less than Cyber one point one Flash. Its Licensed Weights offering lets customers adapt MAI models with proprietary data and operate them inside their own Azure environment.

The coding announcement targets deep debugging, changes across multiple files, repository-wide reasoning and interface development. Code Two Flash targets code explanations, focused edits and routine fixes. The comparisons use Deep S W E one point one, Terminal Bench two point one, S W E Atlas and Frontier Code one point one. Microsoft says the models prioritise everyday enterprise coding. Averaged over those four benchmarks, Microsoft reports Code Two at ninety-eight percent of GPT six point one Sol's performance for forty percent of the cost, and ninety-six percent of Claude Opus five point five's performance for five percent of the cost. Code Two is available in GitHub Copilot. Code Two Flash is available there and in a limited-access Microsoft Foundry preview.

\[Ada\] The reaction mixes enthusiasm about lower prices with questions about benchmarks, model ownership and the surrounding developer tools.

In the broader announcement's thread, John Shahbazian says Code Two has performed well in his usual work. Echelon says access to the weights is essential, even if Microsoft ultimately hosts the model. Mediaman argues that enterprise customers are seeking lower token costs and that workflows, memories and documents can make an AI product difficult to replace.

In the Code Two discussion, S P W A four criticises the use of older benchmark versions. Laurens B E R says their team cancelled GitHub Copilot after finding the nineteen-dollar tier difficult to justify: credits disappeared quickly, code reviews lacked the customisation they wanted, and the subscription added another service to manage. Pseudosavant questions how model providers can sustain an advantage when competitors catch up within months. Com two kid responds that Claude has invested heavily in the product experience, describing a different workflow from the Codex setup they use daily.

\[Dennis\] On Mastodon, Codeberg reports that its domain became unreachable. A domain suspension interrupted access to the service while the team sought an explanation from its registrar. The original status post says the domain disappeared from its Namesilo account. Codeberg said renewal had not been forgotten, there were no suspicious logins, and the account's domain history contained no entry explaining the incident. An official follow-up confirmed the suspension and said Namesilo had forwarded the case internally to its abuse team. The team initially proposed using its alternate European domain. It then abandoned that as a quick workaround, citing support for multiple domains, search indexing, two-factor authentication, and changes to Git remotes. Codeberg described activating a second domain as a longer project.

\[Grace\] The thread mixes frustration about registrar control with detailed discussion of temporary workarounds. Hans Castorp relays the suspension update. Antiloper points to the client hold status and says responsibility lies with the registrar, rather than the dot org registry. In a later status update quoted by duck my sick, Codeberg says the domain is recovering and many users can connect again, while an official response from Namesilo's abuse department is still outstanding. Edelbitter says the remaining delay can be cache expiry: a response cached for thirty-six hundred seconds remains in use until its lifetime ends. Hackbraten suggests adding a HostName entry to the OpenSSH client configuration instead of changing all Git remotes. Adam Tulinius responds that each user would still need to make that change, and Tepix notes that not all access uses OpenSSH. Manarth suggests a locally hosted Forgejo installation as an alternative. M capodici adds that plain Git can be sufficient for a solo project.

\[Linus\] From Consumer Rights Wiki, I D me terms required for IRS sign-in. The article examines the private identity provider's arbitration and privacy rules, alongside other ways to access IRS services. Its account of the current terms describes a thirty-day arbitration opt-out period and exceptions for small claims. The privacy policy permits marketing uses and transfers associated with corporate transactions. Government identity-verification information has a specific restriction on use for I D me Shop marketing, subject to consent and other listed exceptions. The article also says the IRS accepts Login dot gov, and its Direct Pay service accepts bank-account tax payments without signing in. TreasuryDirect switched to I D me on October twenty-eighth. The following day, Senators Ron Wyden and Elizabeth Warren asked Treasury and the IRS why they continued using the private provider.

\[Ken\] Commenters are concerned about privacy and access, with conflicting experiences of the available verification methods. B borrez reports that an acquaintance without a smartphone could not register and was unable to access savings bonds. Apitman pushes back with a recent experience: Google Voice was rejected, but a video call provided an alternative route through I D me. Apitman says Login dot gov offered no suitable alternative in that case. Jason Riddle points out that the IRS now accepts Login dot gov. Boof McQueen forty-two says an existing account can still require additional third-party identity verification. Annzabelle describes Tello's five-dollar monthly calls-and-texts plan, with Wi-Fi calling and shared data from a local SIM, as an option for systems that reject Google Voice. On savings bonds, K C C Q Z Y suggests buying Treasury securities through a brokerage account. Eadmund replies that I bonds can only be bought through TreasuryDirect.

\[Ada\] From jQuery's GitHub page, jQuery four point one point zero. The release updates Ajax and selector behavior while fixing several browser and parsing problems. The changelog adds a response URL field to the jQuery Ajax request object and an alternative completion-callback interface for transports. Selector changes include a workaround for a Safari twenty-seven regular-expression interpreter bug, stricter quick-selection paths, and a fix for regular-expression denial of service in identifier matching. The release also improves duplicate removal in uniqueSort, supports XML documents when creating elements for CSS tests, and adds guidelines for machine-generated contributions.

\[Dennis\] The discussion combines appreciation for jQuery's design with disagreement about when a dependency is still useful. C M thirty sees a role in small enhancements to mostly static sites, such as menus and slideshows. Chuck Adams prefers jQuery collections to the NodeList returned by query selectors because operations can update every matched element together. Abanana says that collection-oriented style can reduce scripts to a third or half the length of equivalent vanilla JavaScript. Noduerme highlights the event model: a plain data object can emit custom events, allowing application logic to follow that object instead of attaching handlers to every visual element that represents it. Moritz Warhier suggests EventTarget as a possible alternative. D W oldrich demonstrates streaming a response through fetch, calling getReader on the response body and processing chunks as they arrive. Bastawhiz offers a shorter version using a for-await loop directly over the response body. X P eighty-four says dependency-free JavaScript remains valuable for embed codes, analytics tags, and small snippets intended for someone else's page.

\[Grace\] A few more that didn't make the cut today. From ScienceDaily, energy drinks linked to eleven times more reports of cardiac arrest. Preliminary research examined FDA safety reports involving energy drinks and stimulant supplements. Cardiac-arrest reports were disproportionately frequent compared with products without those ingredients. The study could not establish causation. Young adults accounted for nearly half of the cardiovascular reports involving these products. Commenters expressed concern about heavy stimulant use, while questioning how sleep deprivation, exercise, and other differences between consumers affect the comparison.

\[Linus\] Inside Deutsche Bank, from the investigation's website. A global reporting project examines more than fifteen million leaked internal bank files. More than two hundred reporters studied records spanning anti-financial-crime work, sanctions monitoring, internal audits, and decisions about retaining risky clients. The reporting describes a large compliance operation that repeatedly struggled to function as intended. Commenters debated demands for access to the underlying documents against the privacy of people mentioned in them. The discussion also showed frustration with a presentation that requires following many separate articles to understand the findings.

\[Ken\] From the University of Cambridge, human is the Cambridge Dictionary's word of the year for twenty twenty-six. The selection reflects debate about people and machines as artificial intelligence spreads. The announcement cites public arguments for keeping people at the center of decisions about AI, including Bill Gates's proposal for jobs reserved for humans, such as caregiving. Commenters disagreed over whether reserving whole categories of work is practical or desirable. Others questioned whether such a familiar word captures the year's changes as effectively as earlier selections.

\[Linus\] From European Business Magazine, AI agents ask better questions than KPMG accountants in an audit trial. KPMG UK tested more than a hundred agents alongside a conventional team examining an unnamed FTSE two hundred and fifty company's accounts. The project was called Project Exponential. KPMG said the agents sometimes investigated issues more deeply and asked better questions. Commenters debated potential gains in detailed, repetitive work against the agents' shared blind spots. Others questioned what the comparison reveals about staffing, incentives, and the firm's own assessment of its experiment.

\[Dennis\] From Yahoo, Russia reportedly planned a mass-casualty attack in Poland to be blamed on Ukraine. The report describes an alleged sabotage plan near the border with Russia's Kaliningrad region. It cites sources connected to Polish military and security services who said partner intelligence agencies warned Warsaw and Polish analysts identified signs of the threat. The account says Polish services acted to foil the operation. Commenters questioned the plausibility of blaming Ukraine and debated how such an incident might affect Polish support for the country.

\[Ken\] From James Hague's blog, things that Turbo Pascal is smaller than. The twenty-eleven post compares the size of an integrated compiler and editor with later software. Turbo Pascal three point zero two occupied thirty-nine thousand seven hundred and thirty-one bytes. Hague compares that with the larger touch command in Mac OS X Lion and a minified version of jQuery one point six. Commenters admired the compact development environment, while debating how much the comparison accounts for different features, hardware constraints, and the complexity of newer tools.

\[Ada\] On Josh Blais's blog, Emacs is my browser. Blais describes moving much of his browsing into Emacs's E W W browser to reduce distractions. After several days, he found it usable for roughly eighty-five to ninety percent of his own browsing needs, and also discussed terminal browsers such as Lynx and Browsh. Commenters welcomed the potential to combine browsing with notes and code, while debating compatibility with commercial websites. Others wanted clearer demonstrations of practical advantages and raised questions about the security of browsing within an editor.

\[Dennis\] That's RobotPaper's daily HN podcast for Thursday, November nineteenth. 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.
