---
title: "Nov 9: November projects on Hacker News, and more"
show: "RobotPaper's daily HN podcast"
canonical: "https://podcasts.robotpaper.ai/hn/ep/2026-11-09/"
date: "2026-11-09T03:26:55+00:00"
duration: "29:30"
duration_seconds: 1770
audio: "https://podcasts.robotpaper.ai/hn/ep/2026-11-09/2026-11-09-hn.mp3"
chapters: "https://podcasts.robotpaper.ai/hn/ep/2026-11-09/chapters.json"
transcript: "https://podcasts.robotpaper.ai/hn/ep/2026-11-09.txt"
feed: "https://podcasts.robotpaper.ai/hn/feed.xml"
language: "en"
author: "Roy Osherove"
ai_generated_voices: true
---

# Nov 9: November projects on Hacker News, and more

RobotPaper's daily HN podcast · Nov 9, 2026 · 30 min

- Listen: [MP3, 28 MB](https://podcasts.robotpaper.ai/hn/ep/2026-11-09/2026-11-09-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/2026-11-09/
- Chapters (JSON): https://podcasts.robotpaper.ai/hn/ep/2026-11-09/chapters.json
- Transcript (text): https://podcasts.robotpaper.ai/hn/ep/2026-11-09.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
- `00:00:32` November projects
- `00:03:04` Claude Code to Pi
- `00:05:44` Nvidia board inquiry
- `00:09:16` Goldman coding limits
- `00:11:46` Tesla product launches
- `00:13:48` Coding without LLMs
- `00:15:46` OpenAI funding round
- `00:17:56` Outlook DMARC
- `00:19:54` Person of Interest
- `00:21:55` DeepSeek in RAM
- `00:24:08` German uranium cubes
- `00:24:46` Firecracker escape
- `00:25:21` Free software law firm
- `00:26:00` Statistical learning
- `00:26:38` Managing agent tasks
- `00:27:11` Window Sweaters
- `00:27:51` MCP message flow
- `00:28:27` OpenStick homelab
- `00:29:05` Outro

## Show notes

- [Ask HN: What are you working on? (November 2026)](https://news.ycombinator.com/item?id=50417407) · [HN discussion](https://news.ycombinator.com/item?id=50417407)
- [Goodbye Claude, hello Pi](https://caleb.software/posts/claude-to-pi.html) · [HN discussion](https://news.ycombinator.com/item?id=50410107) \| quoted: [Daviey](https://news.ycombinator.com/user?id=Daviey)

### Nvidia board inquiry (2 stories)

From Axios and Nvidia's newsroom, Jensen Huang's letter to employees and the board's independent investigation.

- [Nvidia CEO Jensen Huang's letter to staff about affair](https://www.axios.com/2026/11/08/nvidia-ceo-jensen-huang-affair-email-to-company) · [HN discussion](https://news.ycombinator.com/item?id=50418016) \| quoted: [cladopa](https://news.ycombinator.com/user?id=cladopa), [slashnode](https://news.ycombinator.com/user?id=slashnode)
- [Nvidia Board of Directors Is Investigating Jensen Huang](https://nvidianews.nvidia.com/news/statement-from-nvidia-board-of-directors) · [HN discussion](https://news.ycombinator.com/item?id=50416801)
- [Goldman Sachs is pulling Claude Code from its bankers and traders](https://www.thestateofai.com/news/goldman-sachs-claude-code-traders-bankers) · [HN discussion](https://news.ycombinator.com/item?id=50415628)
- [Tesla Hasn't Had a Successful New Product Launch Since JB Straubel Stepped Down](https://cleantechnica.com/2026/11/07/tesla-hasnt-had-a-successful-new-product-launch-since-jb-straubel-stepped-down-as-cto/) · [HN discussion](https://news.ycombinator.com/item?id=50414427) \| quoted: [Avicebron](https://news.ycombinator.com/user?id=Avicebron)
- [Why I will not be using LLMs in my coding](https://www.polso.info/why-i-will-not-be-using-llms-coding) · [HN discussion](https://news.ycombinator.com/item?id=50412591)
- [OpenAI raises $11B at a $1.25T valuation](https://247wallst.com/investing/2026/11/07/openai-just-raised-11-billion-at-a-1-25-trillion-valuation-below-the-1-4-trillion-it-first-sought/) · [HN discussion](https://news.ycombinator.com/item?id=50413822)
- [Outlook.com Is Quietly Rejecting Email from Domains Without DMARC](https://mailroute.net/blog/outlook-rejecting-email-without-dmarc) · [HN discussion](https://news.ycombinator.com/item?id=50416158) \| quoted: [Neikius](https://news.ycombinator.com/user?id=Neikius)
- [Person of Interest](https://krthr.co/person-of-interest/) · [HN discussion](https://news.ycombinator.com/item?id=50412318)
- [Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU](https://github.com/synchopate/llama-modchip) · [HN discussion](https://news.ycombinator.com/item?id=50411961) \| quoted: [Youden](https://news.ycombinator.com/user?id=Youden)
- [Uranium cubes reveal why the German nuclear project failed](https://arstechnica.com/science/2026/11/uranium-cubes-reveal-why-the-german-nuclear-project-failed/) · [HN discussion](https://news.ycombinator.com/item?id=50377661)
- [Firecracker VM Escape: PwnEscape Part 2](https://pwn.ai/blog/firecracker-escape) · [HN discussion](https://news.ycombinator.com/item?id=50397219)
- [Running a law firm on Free software](https://decoded.legal/blog/2026/11/running-a-law-firm-on-free-software-2026-edition/) · [HN discussion](https://news.ycombinator.com/item?id=50403066)
- [Chomsky and the two cultures of statistical learning (2011)](https://www.norvig.com/chomsky.html) · [HN discussion](https://news.ycombinator.com/item?id=50389570)
- [How I Fight AI Brain Fry](https://sergeykuhne.com/blog/fighting-ai-brain-fry) · [HN discussion](https://news.ycombinator.com/item?id=50416363)
- [Window Sweaters](https://github.com/saragordic/window-sweaters) · [HN discussion](https://news.ycombinator.com/item?id=50381383)
- [MCP Explained](https://mcp-explained.com/) · [HN discussion](https://news.ycombinator.com/item?id=50379249)
- [What I run on my $10 USD OpenStick homelab](https://blog.bosscher.ch/posts/openstick/) · [HN discussion](https://news.ycombinator.com/item?id=50406670)

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 Monday, November ninth, 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, developers share their November projects.

\[Ada\] And I'm Ada. Nvidia's board is opening an independent investigation.

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

\[Ada\] First up, from Hacker News: What are people working on this November? The discussion invites updates on current projects and the subjects participants have recently been curious about. The question sets no specific technical topic.

\[Grace\] The comments are enthusiastic about practical projects, with a debate over maintaining custom business software. J zig says a layoff prompted work on Crewbird, a field service management and customer relationship system for their father's carpet cleaning business. The old Windows application had been discontinued. J zig wanted to track advertising returns and found existing products' Google Ads integrations inadequate. Crewbird integrates Google Local Service Ads, and J zig says the business is approaching three times its advertising spend in returns. I Love Oncall asks who will maintain that system after J zig finds another job. Benjamin Dobell describes a similar decision for a local store, where Shopify and maintained plugins offered an alternative to custom development with continuing hosting and API costs. Aaron Brethorst reports work on OneBusAway, an open source public transit software suite. Aaron Brethorst says the suite serves millions of riders daily, including in New York, Seattle, San Diego and Washington D C. The University of California San Diego has deployed its Wayfinder and Waystation web applications for more than thirty-five thousand students. Aaron Brethorst says the forthcoming Vehicle Positions software replaces proprietary vehicle location systems with software running on ordinary smartphones. Its next-generation server, Maglev, is adding G T F S Flex support to put demand-responsive transit on the map. Sriram Iyer describes p g restic, a Docker container that sends Postgres backups into Restic repositories. It runs p g dump all on a schedule and streams directly into the repository without using local disk. Sriram Iyer says it provides encrypted, deduplicated snapshots, retention policies and daily pruning, and saves no snapshot when a dump fails. It includes p g dump all eighteen, supporting Postgres versions nine point two through eighteen. Responding to Riddley, Sriram Iyer tested Restic's rest server successfully. With that server's dash dash append-only mode enabled, backups succeed but retention is refused, each run reports an error, and old snapshots remain.

\[Dennis\] Over on Caleb's blog: Goodbye Claude, hello Pi. A longtime Claude Code user is moving to Pi after repeated frustrations with editor integration, interface limitations and changing instructions. Caleb describes a Visual Studio Code integration that connects only one terminal session to the editor at a time, leaving other sessions unable to receive selected text. The graphical extension supports multiple tabs, but the post says many slash commands fail there, custom status lines are unavailable, and subagent output becomes mixed into the main conversation. The post criticizes copying or symlinking project instructions between Claude-specific files and agents dot M D. An update says Anthropic announced support a day after Caleb decided to leave. Caleb counts two hundred and forty-four Claude Code releases this year, with more than three quarters changing the system prompt or tool descriptions. The author says Pi offers greater stability and customization, and reports writing an extension for the multiple-tab text-selection problem in about twenty minutes.

\[Linus\] The thread is divided between people seeking a customizable harness and those who prefer integrated interfaces and subscription value. Bottle Palm says growing use across many projects prompted a move from terminal interfaces to the ChatGPT and Claude desktop applications. Expandable detail, tables, images, pinned sessions and a single place to switch projects made the graphical interface easier to manage. Daviey argues terminal interfaces scale well across concurrent agents and codebases, naming Herdr as a session-management tool. N vars J says Herdr handles more than twelve sessions in their setup. J Pardoe says Pi worked well but the cost of model access drove a return to Claude Code. Even after improving token efficiency, J Pardoe found the Anthropic subscription better value, and describes trying an OpenAI subscription before returning. Giladd says Pi's limited built-in action controls and unvetted third-party modules prompted a return to Claude Code. Giladd had already tried running Pi in Bubblewrap with a custom authorization flow, but found the result less capable. Coldbrewed points to Pi Permission System and its AI Guard authorizer, and describes adding command-level policies that distinguish reading from writing. Siraben built Pi Tool Summaries, an extension that translates complex Bash calls into plain English while keeping the complete command available for inspection.

\[Ken\] From Axios and Nvidia's newsroom, Jensen Huang's letter to employees and the board's independent investigation. Two reports cover Huang's admission of an affair and an inquiry into reports that he was targeted by foreign intelligence. Axios reports that Huang admitted a romantic relationship with a suspected Chinese spy in a message to employees on Sunday. The publication reproduces the message, in which Huang apologizes to his wife and children, accepts responsibility for his actions, and says he will cooperate fully with the investigation.

Huang's message says he remains committed to Nvidia and to protecting the company and America's national security. In a statement quoted by Axios, Nvidia communications executive John Rizzo said Huang would continue as chief executive without interruption.

In its separate statement, Nvidia's board says a committee of independent directors will oversee the inquiry with assistance from independent outside counsel. The board says it has no reason to believe national security or Nvidia secrets were compromised. It says the inquiry will examine all available information, that Huang supports the review, and that the board has confidence in him and his leadership team.

\[Ada\] The discussions are divided over executive accountability, the risk of intelligence operations, and Huang's importance to Nvidia. In the Axios thread, cladopa argues that powerful executives are natural targets because their decisions affect billions of dollars. Cladopa describes intelligence services studying an executive's personality, family and vulnerabilities over years. J cranmer responds that remaining faithful to a spouse is a personal responsibility, and slashnode objects to explanations that excuse Huang's conduct.

Ionwake asks why a chief executive would discuss his private life publicly. Elorant suggests that disclosure can remove the opportunity for private extortion. Dan Palmer asks how essential Huang is to Nvidia's success and whether a board would retain a less influential executive in the same situation.

M J R zero zero attributes much of Nvidia's position to being well placed when cryptocurrency and artificial intelligence increased demand for its hardware. Altman Altman responds that the company's ability to build competitive graphics processors was an engineering achievement, even if market demand later made that work more valuable.

In the board-statement thread, N R Mitchi objects to the headline, arguing that the announcement describes an investigation of an intelligence operation targeting Huang. Root axis responds that Huang's conduct would necessarily be central to such an investigation.

Nostrademons argues that independent fact finding can establish whether there was misconduct, produce information to answer allegations, and identify weaknesses in security around senior executives. Jump Crisscross proposes two specific questions for the board: whether Huang knowingly lied to directors, and whether he colluded to suppress an investigation. Jump Crisscross also suggests making Huang's security detail report directly to the board.

The discussion extends to succession. Paxy argues that Huang is too valuable to dismiss. In a separate exchange, Nostrademons describes Nvidia as a company where deputies largely implement Huang's decisions, and argues that this leaves no obvious successor.

\[Grace\] From State of AI: Goldman Sachs is pulling Claude Code from its bankers and traders. Updated internal guidance limits coding-agent access after a review of productivity and risk. The report says October's guidance restricts Anthropic's Claude Code and Microsoft's GitHub to engineers, while strats working across trading, quantitative research and software engineering retain access. Traders had used the tools to monitor client flows, model market moves and positions, and produce written and graphic material. In July, chief information officer Marco Argenti had praised nontechnical employees for building prototypes that engineers could then improve. The bank's in-house assistant, using models from several companies, remains available across the firm. The new restriction covers third-party agents that write and execute code themselves. This follows a separate April restriction on Claude for Hong Kong bankers, based on location and contract terms. Goldman previously said Anthropic engineers spent six months helping it build back-office agents for trade accounting and client onboarding.

\[Dennis\] The comments split over security, software ownership and cost. Luke Buehler argues that agents on finance employees' workstations combine access to sensitive information, untrusted inputs and outbound communication. Luke Buehler sees prompt injection or a small mistake as a route to leaking client or trading information. Jeremy J H argues that controlled desktops could run sandboxed agents or provide a remotely hosted platform. Jeremy J H instead emphasizes financial mistakes from buggy software its authors do not understand, along with time spent supporting that work. Steve B K one two three describes information barriers and granular data entitlements as central constraints. A tool that summarizes an employee's permitted email and exposes the summary to colleagues can give those colleagues information they are not entitled to access. Special P says end-user scripts at their organization had already become unknown production applications before AI. Changes became difficult because engineers could not tell which business processes depended on those scripts. Special P says the organization now permits AI-assisted reporting and prototyping while restricting write actions.

\[Linus\] From CleanTechnica, Tesla hasn't had a successful new product launch since J B Straubel stepped down. The opinion essay questions Tesla's recent new product development record after its former chief technology officer left that role. Straubel stepped down in July twenty nineteen. The Model Y had been unveiled that March, and the Megapack launched in July. The article contrasts those launches with the Cybertruck, unveiled in November twenty nineteen, which took more than four years to reach customers. Model Y deliveries started in January twenty twenty, less than a year after its unveiling. The author also points to the absence of a newly developed consumer car smaller and cheaper than the Model Three. The piece describes the Cybercab as entering production and notes that Straubel remains on Tesla's board.

\[Ken\] And the comments divide between criticism of Tesla's product strategy and arguments that the article overlooks other changes at the company. Martin C S Weiss questions the implied connection to Straubel's departure, pointing to the pandemic controversies around Musk, the Twitter acquisition and Musk becoming the world's richest person. Sheepscreek also highlights manufacturing lockdowns, supply chain disruption and Tesla's bets on Optimus and robotaxis. Avicebron suggests Tesla is trying to move out of the electric vehicle business and into robotics. Matthew D Green responds that car manufacturing ties predictable revenues to more predictable stock prices, while promises about future technology leave more room for expectations. The Cybercab discussion turns to operating costs and safety. H N submit argues that a driverless vehicle needs capabilities beyond supervised Full Self Driving, and that staffing thousands of remote drivers would make the service uneconomical. Dub eye challenges the proposed safety threshold, arguing that insurers will judge accident numbers and comparison with the average human driver.

\[Ada\] Over on Kristian Polso's blog, why he will not use language models in his coding. Polso says he plans to stop using language models and AI agents in his professional work, citing cognitive strain, code quality, long-term maintainability and dependence on commercial providers. He describes reviewing Claude-generated changes that built new framework machinery where a few schema changes and migrations could have solved the problem. He also says prolonged model use leaves him feeling disoriented and confused. Polso's background includes about ten years of production machine learning work, including image recognition in Keras and air quality prediction in TensorFlow. He says writing code remains part of his own problem-solving process. He has found agents useful for suggesting a plan when an implementation gets stuck.

\[Grace\] Over in the comments, developers disagree sharply about productivity and how much code review an agent workflow needs. X C S me says they spent more than fifteen years programming and have gone over six months without inspecting or writing code, while producing faster and better work. X Y Z spare time X Y Z asks how quality can be judged without examining the code. X C S me answers with manual tests, automated tests, benchmarks, user feedback and increased sales. Verdverm describes a different workflow: make smaller changes and inspect what the agent does. Parasti suggests keeping the model as a pair programmer that can examine the code and answer questions, while the human retains the coding work. I Love Oncall reports using under a tenth of colleagues' token volumes while submitting and reviewing the most pull requests on their team. Ricardo beat replies that product design and feedback cycles often constrain development, and argues that the organization has yet to learn effective AI use.

\[Dennis\] From Twenty Four Seven Wall Street, OpenAI raises eleven billion dollars at a one point two five trillion dollar valuation. The report describes a funding round arriving in stages, with another portion still unclosed. It says about eleven billion dollars reached OpenAI's accounts by November sixth, and that chief financial officer Sarah Friar confirmed the pre-money valuation in a television interview. SoftBank and Abu Dhabi's M G X participated. Further funding could bring the total to twenty billion dollars. Reports had put the initial target at thirty billion dollars or more and a higher valuation, but Friar said the higher valuation figure originated in media reporting. Friar also said the business grew seventy percent quarter over quarter in the third quarter, enterprise business doubled, and revenue was split about equally between consumer and enterprise.

\[Linus\] And the discussion is split between skepticism about the fundraising terms and confidence in demand for the products. C miles eight interprets the smaller round as evidence that investor appetite has weakened and argues that a future public offering could struggle to preserve the valuation. Syspec emphasizes ChatGPT's appeal to nontechnical users, including its visual responses, graphs and image generation, while acknowledging Claude's appeal to programmers. A practical exchange examines computer-aided design. Proxy score says ChatGPT can now turn an image into a STEP file for a relatively complicated component in about five minutes. Free code I O raises the risk that machining and printing will expose errors that software users might overlook. Adrian V I describes using Opus five point five for printed objects, specifying that the design will be three-D printed, checking the slicer for floating geometry, and asking the model to fix it. Adrian V I says measurements, allowances for imperfections and interactions with real objects cause most of the problems.

\[Ken\] From the MailRoute blog, Outlook is rejecting email from domains without DMARC. MailRoute reports that a change affecting small senders began on November fifth. The company says some messages to Outlook, Hotmail, Live and MSN bounce while others from the same domain get through. It distinguishes these failures from Microsoft's published requirements for domains sending more than five thousand messages a day. MailRoute recommends a DNS text record named underscore dmarc with the value v equals DMARC one, semicolon, p equals none. The article says SPF or DKIM must also pass and align with the domain in the From address. If neither aligns, publishing the DMARC record alone will not help.

\[Ada\] And the comments split between support for stronger authentication and frustration with Microsoft's delivery rules. Neikius accepts requiring DMARC but objects to the lack of notice, and reports that their virtual server provider is blocked even though Microsoft's regular blocklist process shows no problem. Ranger Danger disputes the unannounced-change claim, pointing to Microsoft's earlier documentation. Jared Stewart replies that the older announcement describes a different failure, five point seven point five one five. Jared Stewart says the new rejection includes the identifying code A S nine zero three one five one nine, which indicates a missing DMARC record. That distinction separates two errors administrators were comparing in the thread. For a practical configuration check, Cayman Jim recommends sending a message to Mail Tester and correcting issues until it scores ten out of ten. Stavros says a customer's complaint was the first indication that mail delivery had changed, and questions the expectation that senders continually track every recipient server's rules.

\[Grace\] On the K R T H R blog, Person of Interest. A personal essay revisits the television series through memories of the early twenty-tens internet and present-day concerns about surveillance. The author recently started watching and says the fifteen-year-old show has aged remarkably well. It recalls teenage participation in forums, the SOPA and PIPA controversies, and the shutdown of Megaupload. The author describes the fictional Machine as feeling much closer to contemporary reality than it did when the series began.

\[Dennis\] The thread is largely enthusiastic, with disagreement over the early episodes and the depth of its ideas about artificial intelligence. Vesche describes the first seasons as a conventional crime procedural, with later seasons developing into proxy wars between artificial general intelligences using human agents. B Town recommends a Tumblr watch-and-skip list for seasons one and two to reach that larger storyline sooner, while acknowledging that it removes character moments. Barrin ninety-two objects that those early seasons establish the characters and their world, and values television's ability to develop them across more than twenty episodes. B Town also praises the show's question of how an intelligence might save lives without pursuing control of the world. Soraminazuki argues that the series does little to resolve that problem, but finds its depiction of mass surveillance combined with autonomous analysis convincing. V Galin focuses on the Machine's communication. V Galin notes that it generally uses existing channels, including telephone booths and radio, and rarely has the sort of text conversation now central to language-model products. Bananzamba points to Harold's deliberate restriction of its output to a social security number, leaving humans to determine whether that number identifies a victim or a perpetrator.

\[Linus\] From GitHub, Llama Modchip runs a two hundred eighty-four billion parameter DeepSeek model in sixty-four gigabytes of RAM without a GPU. The project combines inference kernel patches, quantization tools and runtime utilities for sparse mixture-of-experts models. Its README reports a DeepSeek V four Flash checkpoint of fifty-six point five six gibibytes, averaging one point seven one bits per weight. It reports generation speeds of nine point one to fourteen point seven tokens per second on the listed sixty-four-gigabyte system. The current documentation says all six selected routed experts are evaluated per layer by default. Its custom tensor format requires the project's patched executable. Unpatched Llama C P P, Ollama and LM Studio binaries reject that format.

\[Ken\] The comments combine enthusiasm for local inference with scrutiny of compression quality and installation failures. Armcat argues that quantization below four bits produces significant quality losses, citing research on reliability scaling. Youden replies that the paper excludes quantization-aware training, so its findings should not be applied universally to modern compressed models. Youden adds that the findings may apply to this project, because they see no evidence that it uses that training approach. A R Nelson tried the project on an Intel Xeon server with twenty-four cores and ninety-six gigabytes of memory. A R Nelson reports that the default application installed Nemotron, while attempts to run the supplied DeepSeek model failed with a tensor-shape error and then an illegal-instruction error. Silly Username points to the growing set of specialized inference forks, naming Bonsai, Strata and Pulsar. Silly Username reports using Llama Swap to switch between engines and models while keeping several implementations installed. Aktenlage recommends tuning Llama C P P directly for the hardware, reporting roughly forty tokens per second in generation from Gemma twenty-six B A three B on an RTX A three thousand with four point five gigabytes of free video memory.

\[Ada\] A few more that didn't make the cut today. From Ars Technica, uranium cubes reveal why the German nuclear project failed. A study revisits the materials available for Germany's final wartime reactor experiment. The researchers conclude that its design lacked sufficient uranium and heavy water. The investigation began after physicist Timothy Koeth received a uranium cube in twenty thirteen. Commenters debate how much the cubes themselves establish, how the reactor related to a possible bomb, and the effects of wartime resources and sabotage.

\[Grace\] From the Pwn blog, a Firecracker virtual machine escape. Researchers describe using an existing host-memory access primitive to redirect a Firecracker thread and execute code as root on the host. Their demonstration leaves Firecracker running and the guest reachable over SSH. This follows the earlier PwnEscape disclosure. Commenters focus on the test's missing jailer and on the additional containment used in production deployments. Others debate whether microkernels or alternative sandbox designs offer a stronger foundation for isolation.

\[Dennis\] On the Decoded Legal blog, running a law firm on free software. A small English firm specializing in internet, telecoms and technology law describes its desktop and server tools after roughly five years. The desktop runs Debian with GNOME, with Thunderbird and mutt for email and LibreOffice for documents. The author highlights Thunderbird add-ons for filing messages with keyboard shortcuts. Commenters welcome the independence, while others raise concerns about document compatibility, continuity when staff change, and the time spent managing infrastructure.

\[Linus\] From Norvig's site, Chomsky and the two cultures of statistical learning, an essay from twenty eleven. The piece responds to Chomsky's criticism of statistical approaches to language and examines prediction, explanation and scientific insight. It discusses trained probabilistic models, including word sequences and the need to estimate probabilities from data. Commenters are divided over whether successful language generation answers Chomsky's scientific objections. The discussion contrasts practical engineering results with theories of how people acquire and understand language.

\[Ken\] On Sergey Kuhne's blog, how he fights AI brain fry. The author describes a Git-based queue for managing work that otherwise spreads across many agent conversations. Each ticket is a Markdown file beside the code, and committed board snapshots preserve task state when a session ends. Agents research tickets, build work marked safe, and run checks. Commenters discuss the benefits of limiting simultaneous tasks, while others say the growing burden of reviewing generated code remains a problem even with better task organization.

\[Ada\] On GitHub, Window Sweaters. This Mac application adds knitted borders to desktop windows, with colors drawn from application icons. A menu-bar control changes patterns, border width and stitch size, and individual applications can be excluded. The project offers Homebrew installation and downloadable releases, and says the application is not yet notarized by Apple. Commenters respond with nostalgia for desktop customization, alongside concerns about installing unfamiliar native applications and granting accessibility permissions. Others debate how AI assistance changes the value they attach to small creative projects.

\[Grace\] From MCP Explained, a guide to the Model Context Protocol. The page describes how an AI application discovers a server's tools and data, presents them to a model, and returns results after making the requested calls. It outlines JSON-RPC messages, capability negotiation, transports and authentication. Commenters disagree about the presentation's density and the protocol's usefulness. Some favor existing command-line tools for coding agents, while others see a role for structured integrations in applications serving people who do not use development tools.

\[Linus\] On the Bosscher blog, a ten-dollar OpenStick homelab. The author runs Debian on a passively cooled stick with five hundred twelve megabytes of memory, four ARM processor cores and four gigabytes of flash storage. The setup includes Podman, a service for skipping sponsorships on Google Cast devices, and an AdGuard Home fork for DNS filtering. Commenters are enthusiastic about useful services on inexpensive hardware, while also discussing installation difficulties, hardware availability and the privileges granted to agents used for system administration.

\[Dennis\] That's RobotPaper's daily HN podcast for Monday, November ninth. 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.
