---
title: "Nov 5: AI exposes legacy software debt, and more"
show: "RobotPaper's daily HN podcast"
canonical: "https://podcasts.robotpaper.ai/hn/ep/2026-11-05/"
date: "2026-11-05T10:28:52+00:00"
duration: "29:21"
duration_seconds: 1761
audio: "https://podcasts.robotpaper.ai/hn/ep/2026-11-05/2026-11-05-hn.mp3"
chapters: "https://podcasts.robotpaper.ai/hn/ep/2026-11-05/chapters.json"
transcript: "https://podcasts.robotpaper.ai/hn/ep/2026-11-05.txt"
feed: "https://podcasts.robotpaper.ai/hn/feed.xml"
language: "en"
author: "Roy Osherove"
ai_generated_voices: true
---

# Nov 5: AI exposes legacy software debt, and more

RobotPaper's daily HN podcast · Nov 5, 2026 · 29 min

- Listen: [MP3, 28 MB](https://podcasts.robotpaper.ai/hn/ep/2026-11-05/2026-11-05-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/2026-11-05/
- Chapters (JSON): https://podcasts.robotpaper.ai/hn/ep/2026-11-05/chapters.json
- Transcript (text): https://podcasts.robotpaper.ai/hn/ep/2026-11-05.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
- `00:00:38` AI and legacy security
- `00:03:24` Data center ballots
- `00:05:43` Binance interview hack
- `00:08:12` C++26 ballot fails
- `00:10:25` AI coding workflows
- `00:13:22` Hinton on AI risks
- `00:15:28` RTX Spark on Linux
- `00:17:31` Windows menu controls
- `00:19:31` Stringman robot
- `00:21:08` White House proxies
- `00:22:47` Kitty terminal
- `00:24:37` Guerrilla bike lanes
- `00:25:19` AMD revenue outlook
- `00:25:55` UK high streets
- `00:26:30` AI obligations
- `00:27:06` Chat training controls
- `00:28:24` Conductor joins Meta
- `00:28:56` Outro

## Show notes

- [AI is forcing us to pay off 40 years of technical debt](https://spectrum.ieee.org/ai-in-cybersecurity-php2shell-vulnerability) · [HN discussion](https://news.ycombinator.com/item?id=50364336)
- [US voters roundly reject data centers in ballot measures](https://www.semafor.com/article/11/04/2026/us-voters-roundly-reject-data-centers-in-ballot-measures) · [HN discussion](https://news.ycombinator.com/item?id=50365982)
- [North Korean fake job interview led to $6.3B Binance heist](https://www.theregister.com/security/2026/11/03/north-korean-fake-job-interview-led-to-63b-binance-heist/5307398) · [HN discussion](https://news.ycombinator.com/item?id=50365755)
- [C++26 Disapproved](https://andreasfertig.com/blog/2026/11/cpp26-dis-failed-how-does-cpp-survive/) · [HN discussion](https://news.ycombinator.com/item?id=50360164) \| quoted: [nanolith](https://news.ycombinator.com/user?id=nanolith)

### AI coding workflows (2 stories)

From Ben Stolovitz's blog and Hacker News, changing workflows for AI agents.

- [How I use AI (Oct 2026)](https://ben.stolovitz.com/posts/how-use-ai-oct-2026/) · [HN discussion](https://news.ycombinator.com/item?id=50355767) \| quoted: [Dzugaru](https://news.ycombinator.com/user?id=Dzugaru)
- [Ask HN: What's your current AI agentic setup?](https://news.ycombinator.com/item?id=50366117) · [HN discussion](https://news.ycombinator.com/item?id=50366117)
- [Geoffrey Hinton: it might turn out we were better off without AI](https://www.theguardian.com/uk-news/2026/nov/03/ai-godfather-geoffrey-hinton-better-off-without-ai) · [HN discussion](https://news.ycombinator.com/item?id=50363554) \| quoted: [Aurornis](https://news.ycombinator.com/user?id=Aurornis)
- [Nvidia RTX Spark encrypted firmware slows inference speed to 35% on Linux](https://www.mindstudio.ai/blog/linux-nvidia-rtx-spark-laptop) · [HN discussion](https://news.ycombinator.com/item?id=50362905)
- [Microsoft fixes a Windows 11 annoyance as PC revenue declines](https://arstechnica.com/gadgets/2026/10/windows-11-gets-right-click-menu-improvements-as-pc-revenue-slides/) · [HN discussion](https://news.ycombinator.com/item?id=50351463)
- [Stringman, an $800 open-source room scale gripper robot](https://neufangled.com/) · [HN discussion](https://news.ycombinator.com/item?id=50354358)
- [Residential proxies detected on White House network](https://layer3intel.com/blog/residential-proxies-white-house-network) · [HN discussion](https://news.ycombinator.com/item?id=50371246)
- [Kitty 0.50](https://sw.kovidgoyal.net/kitty/changelog/#) · [HN discussion](https://news.ycombinator.com/item?id=50362980)
- [Guerrilla bike lanes: San Francisco makes illicit infrastructure permanent (2016)](https://99percentinvisible.org/article/guerrilla-bike-lanes-san-francisco-makes-illicit-infrastructure-permanent/) · [HN discussion](https://news.ycombinator.com/item?id=50366522)
- [AMD expects to outsell Intel in Q4 for the first time in history](https://www.tomshardware.com/pc-components/cpus/amd-expects-to-outsell-intel-in-q4-for-the-first-time-in-history-usd17b-guidance-would-exceed-intels-projections) · [HN discussion](https://news.ycombinator.com/item?id=50367125)
- ['It's just vape, nail, charity shops': Can anything save UK high street?](https://www.theguardian.com/business/2026/nov/04/what-can-save-the-british-high-street-rotherham) · [HN discussion](https://news.ycombinator.com/item?id=50367653)
- [US tech giants' hidden AI debt swells 90% to $3.2T in 6 months](https://asia.nikkei.com/business/technology/artificial-intelligence/us-tech-giants-hidden-ai-debt-swells-90-to-3.2tn-in-6-months) · [HN discussion](https://news.ycombinator.com/item?id=50358422)

### Chat training controls (2 stories)

On Hacker News, Gemini and Codex chat-training controls.

- [Tell HN: Gemini trains on chat history, even for subscribers (personal accounts)](https://news.ycombinator.com/item?id=50363461) · [HN discussion](https://news.ycombinator.com/item?id=50363461)
- [OpenAI seems to be silently re-enabling model training on Codex chats](https://news.ycombinator.com/item?id=50352757) · [HN discussion](https://news.ycombinator.com/item?id=50352757)
- [Conductor is joining Meta](https://www.conductor.build/announcement) · [HN discussion](https://news.ycombinator.com/item?id=50369503)

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 fifth, 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-assisted audits expose old vulnerabilities; maintainers face lengthy repairs.

\[Ada\] And I'm Ada. We also have C plus plus standardization, and Windows interface changes.

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

\[Ada\] From IEEE Spectrum: AI exposes decades of technical debt. AI-assisted security research is turning longstanding software bugs into working exploits, while maintainers face a slower repair process. IEEE Spectrum describes PHP2Shell, which exposed websites running PHP to data theft and remote code execution. The report names NASA, Wikipedia and the White House among websites exposed, alongside sites running WordPress or Drupal. The underlying bug had been reported to the PHP Foundation in twenty twenty-four. During an OpenAI-sponsored audit, a researcher at security firm Calif used AI tools to develop an exploit within two weeks. The researchers also had deep expertise in PHP and cybersecurity. They alerted the foundation, and a patch arrived before any known malicious exploitation. Columbia University's Suman Jana says finding an exploit requires detailed reasoning about memory locations and the order of writes. Fixing the problem then took four engineers two months, because they also had to preserve existing behavior. Jana describes another approach involving AI-generated patches and proofs, checked by a separate formal proof checker. Johns Hopkins professor Matthew Green's immediate advice in the article is to install security updates when they become available.

\[Dennis\] And the comments are divided between optimism about broader auditing and concern that rapid changes will introduce more vulnerabilities. Arbor theus hopes automated discovery will expose vulnerabilities stockpiled by governments and spyware companies, making software more robust after the current patching effort. Swat coder questions whether this debt can be permanently retired: substantial fixes may require restructuring, and each rewrite can introduce another defect. In the discussion of PHP itself, traceroute sixty-six argues that criticism based on PHP four and five overlooks substantial work in versions seven and eight. Traceroute sixty-six also names ColdFusion, CGI and ASPX as the alternatives developers faced at the time. Keeda challenges the assumption that publicly visible code automatically gets an effective audit, arguing that a skilled security researcher such as Tavis Ormandy can contribute more than a large group of general developers. Lil big doot proposes stronger operating-system permissions as a practical safeguard: explicitly whitelist which files a program may access, reducing the burden on application code to validate every path perfectly.

\[Ken\] Semafor reports that US voters rejected data centers in local ballot measures. Communities approved restrictions on new facilities as opposition to the AI infrastructure buildout reached the ballot box. Semafor says residents in localities across nine states backed moratoriums, bans and other limits on Tuesday. Some measures directly targeted projects already underway. In Edgewater, Florida, voters approved putting a data-center ban in the city charter. Perry, Ohio, passed financial and operational constraints on hyperscalers seeking tax breaks. The report says lenders and asset managers are already treating local opposition as a credit risk, while hyperscalers recruit community liaisons. Companies are also trying the label compute campuses for their facilities. Semafor describes an election campaign in which data-center opposition appeared in advertising and pushed some candidates toward promises of tougher regulation.

\[Grace\] And the thread is sharply divided between supporters of local restrictions and commenters who see them as barriers to investment and growth. Twick low describes a proposed Kevin O'Leary data center near their Utah community and says it would use one hundred sixty-two million fewer gallons of water annually than the farm previously on the land. Twick low also says the project would generate its own electricity and send excess power to the grid. Guitar limeo challenges that reassurance, arguing an operator could switch to grid power whenever buying electricity becomes cheaper than generating it. N diddy outlines proposals attributed to Abdul El-Sayed: transparently negotiated community-benefit agreements, no increase in utility rates, revenue reserved for power reliability, protection of waterways and union construction labor. N diddy argues small local governments can be outmatched by corporate lawyers from companies such as Google, xAI and Oracle. P J Mukh argues economic growth requires actual construction in communities and worries voters expect the gains while pushing development elsewhere. Better Unix two proposes payments to the surrounding community in proportion to a data center's electricity consumption, and higher payments for facilities operating off the grid to discourage pollution from their own generators.

\[Linus\] The Register reports that a fake job interview preceded the Binance heist. Binance says attackers compromised employee computers and deployment credentials before stealing six point three billion dollars on October twenty-ninth. The Register reports that the FBI attributed the theft to North Korea's TraderTraitor operation. According to Binance's incident account, an engineer downloaded a malicious Homebrew installation in June. The company replaced that device and rotated affected credentials, but information about the wallet team had been exposed. A different engineer later entered a fake recruitment process. During a second interview, the supposed chief executive supplied a coding assignment whose setup executed malware on the engineer's work computer. The malware replaced the legitimate pyenv development-tool component. Attackers subsequently used that computer and an internal AI tool to reach wallet engineers' machines, stole a container-deployment token, and used it to launch the attack. Binance says it rebuilt the affected systems, replaced compromised devices and decommissioned the engineers' credentials.

\[Ada\] And the comments combine criticism of using a work laptop for interviews with concern about the access available through a single deployment credential. Gandreani argues deployments should run only through continuous integration, providing a clean environment to build software. Gandreani specifically objects to workflows that require copying a token onto a developer's computer and remembering to delete it afterward, and recommends paid intrusion testing. Ben oh ow describes moving comparable processes off a team's computers in a previous job, where a mistaken command-history selection could otherwise affect production. Alex I T C says social engineering already features in intrusion tests and suggests fake interviews could be included. Bilekas argues work-issued laptops should be reserved for activities within an employee's role. Paxy, describing experience as an interviewer using Zoom, says candidates sometimes joined with company accounts and exposed email inboxes, calendars and proprietary windows during screen sharing. Paxy also reports candidates completing interview exercises inside their employers' internal development environments and tools.

\[Dennis\] On Andreas Fertig's blog: the C++ twenty-six draft failed its ballot. National standards bodies rejected the draft after objections to its contracts feature, leaving the next release delayed. Fertig, a member of Germany's standardization delegation, says the October twenty-third ballot failed to reach the required two-thirds majority. Contracts had previously cleared the C++ working group's process, following years of discussion. Fertig says national bodies objected that the feature was not ready for release. He lists several possible next steps: revise the draft for another ballot, return it to an earlier committee stage, publish a revised technical specification, or cancel the project. Each route requires support inside the working group. Fertig describes the disagreement as a breakdown in trust, while noting that these are his personal views. The next committee meeting is scheduled for November sixteenth.

\[Ken\] And the thread is split over the contracts design and whether the national bodies properly exercised their authority. Low watt asks why a second approval stage should be considered a betrayal when its voters reject a proposal. J Andrew Rogers, describing work on high-reliability C++, argues existing constexpr and C++ twenty concepts already cover much of the useful ground. Quote master disputes that comparison, distinguishing concepts, which describe types, from contracts, which provide structured runtime assertions. Nanolith offers a concrete alternative from their own work: C code with custom macros and the CBMC model checker, expressing function contracts and invariants for compositional verification. Nanolith favors a simpler future contracts proposal and says C's relative simplicity makes external analysis tooling easier to build. In a follow-up, nanolith points readers to their BSDCan twenty twenty-six talk, which includes slides and example code. Uecker reports that CBMC automatically proved memory safety and assertions in small test programs, while larger programs required splitting the work into pieces using a contract-like mechanism.

\[Grace\] From Ben Stolovitz's blog and Hacker News, changing workflows for AI agents. Two posts examine how developers delegate work and check the results. Stolovitz, now a research engineer at Salesforce AI Research, says he manages five or six coding agents, each working in a separate Git worktree. His tools include Claude Code, OpenCode, Codex and Pi, and he reports spending about three thousand dollars monthly on tokens. Agents prototype features, investigate bugs and prepare pull requests. He still reviews architecture and tests results manually. His strongest improvement this year involved replacing about four thousand lines of generated code with a few hundred handwritten lines. Managing several simultaneous conversations also leaves him exhausted. Beyond coding, he reports that days spent discussing training-data annotations with models were resolved by one conversation with a teammate. For checking his understanding, he has asked Claude to quiz him on a long code review. He uses OpenWhispr and Handy for local dictation when talking to coding agents, but reports inconsistent hotkeys. He avoids AI drafting and outlining for his own writing. The companion Ask HN post on Hacker News asks about subscriptions, locally hosted models, coding interfaces and orchestration. It specifically invites comparisons between standard tools and custom setups, including OpenCode, Pi and Aider.

\[Linus\] The threads combine enthusiasm about faster development with concern about verification and fatigue. In Stolovitz's discussion, M E F fifty one asks how developers will acquire the judgment agents still require if they stop practicing programming. M Z Haase describes an agent updating a database by copying the entire database into memory, changing it there, deleting the original and copying everything back. Tests passed, and the problem emerged only during inspection. User four three nine two eight argues that agents make later architectural changes cheaper, while still requiring substantial quality assurance before production. Dzugaru reports finding incorrect assumptions in generated code daily, especially in data analysis and complex systems, despite spending more than three thousand dollars monthly on tokens. In the setup discussion, The Fourth Chime uses Claude Code Desktop with Opus five point five at medium effort and recommends specifying exactly what a completed result should contain. D G B twenty three uses Pi with self-hosted GLM Flash, separates planning, testing and implementation, and turns recurring tasks such as filtering logs and checking status into a project command-line tool and local agent skills.

\[Ada\] From The Guardian: Geoffrey Hinton says it might turn out we were better off without AI. The computer scientist is calling for independent safety testing before powerful models reach the public. Hinton told the newspaper that AI could produce enormous abundance or cause unemployment, inequality and severe social disruption, depending on how its risks are managed. He warned that a major incident could arrive this year or next, giving examples of an attack on electricity generation or a financial system outage lasting a week. Hinton supports mandatory testing and common international standards. He wants companies to present statistical evidence to an independent authority that a model does more good than harm before release, comparing that process with drug approval. He also said he remains hopeful that people can retain control if enough effort goes into addressing the risks. Hinton left Google in twenty twenty-three so he could speak freely about AI risks.

\[Dennis\] The discussion is divided between criticism of commercial incentives and arguments that research would advance under different companies anyway. Mister weasel argues that presenting language models as improved IntelliSense, able to suggest implementations, or as better Grammarly would have created less opposition. Mister weasel says that approach would require building businesses over time instead of chasing rapid, enormous valuations. Aurornis replies that researchers and competing entrepreneurs would continue pushing capabilities even if existing frontier labs stopped at autocomplete and writing assistance. Aurornis also disputes the idea that company founders mainly want a quick payout and retirement. Mirmor twenty-three describes using AI to extend a protocol across a desktop application and an internet of things device, maintaining the shared premise during development and testing. Mirmor twenty-three also reports that it has been particularly useful for debugging systems with multiple threads.

\[Ken\] From MindStudio, native Linux performance problems on an RTX Spark laptop. Tests on an MSI Prestige laptop found that plugging into wall power sharply reduced GPU speed. With GPT O S S one twenty B, MindStudio reports sixty five point four tokens per second on Windows, sixty five point one through Windows Subsystem for Linux, and thirteen point nine on native Linux. GPU clocks fell to roughly seven hundred eighty megahertz when plugged in. On battery, a separate LM Studio test reached sixty one to sixty seven tokens per second. MindStudio attributes the problem to missing Linux support for Nvidia's Platform Controllers and Framework and an encrypted power-management table. Disabling that controller in Windows reproduced the slowdown; enabling it restored performance. WSL retained the Windows driver's boost behavior. The tested setup also needed its default fourteen-gigabyte WSL memory allocation increased for large-model workloads. Reducing the Windows-side GPU memory reservation from ninety six to four gigabytes made one hundred twenty gigabytes visible to Linux.

\[Grace\] The discussion is divided between criticism of platform restrictions and arguments about missing driver support. None the Wiser points to the MSI Prestige and notes that these RTX Spark laptops ship with Windows. Callam Delaney argues Linux should deliver equal performance from day one, or owners should have the ability to make it do so. A Farah One highlights Nvidia's move to open-source kernel modules. K L L R N O H J responds that proprietary firmware still carries functionality previously inside the driver. Barkel proposes a different setup for private inference: keep a DGX Spark or similar machine powered at home, connect through Tailscale, and use a phone hotspot while traveling. Anonym twenty nine names Strix Halo and Mac Studio as other home inference options.

\[Linus\] Next, from Ars Technica: Microsoft fixes a Windows eleven annoyance as PC revenue declines. A preview update gives users control over the commands in the right-click menu. The cumulative update covers Windows eleven versions twenty-six H two and twenty-five H two, with a gradual rollout. Its Customize Menu button can toggle built-in commands, rearrange the main menu and manage third-party extensions. Previously, familiar operations were split between icons and the Show More Options submenu. Cloud Rebuild can also reinstall Windows and drivers from the cloud when the system fails to boot; existing applications and files are excluded. Microsoft reported that Windows OEM revenue fell twenty-seven percent, citing last year's Windows ten support deadline, inventory reductions and weaker PC demand.

\[Ada\] The comments are mostly frustrated with Windows changes, with discussion turning to Linux alternatives and practical ways around interface problems. Glimshe argues that Windows ten already worked well and that later interface changes removed features without improving productivity. Recursive call explains a use for Copy as Path: supplying script locations to a program in another folder, including AutoHotKey or Task Scheduler. Simoncion offers a technique for running Windows software under Linux through Steam. Choose Add a Non-Steam Game to My Library, browse to an executable, and let Steam create the Proton environment. Simoncion says the executable associated with that entry can be changed afterward, and suggests trying the installer first and then pointing the entry at the installed application. A N T H K names Lutris as another tool that handles this. For native art software, A N T H K says Krita has a Flatpak and runs well under Fedora Atomic.

\[Dennis\] From Neufangled, Stringman, an open-source room-scale gripper robot. The seven hundred ninety five dollar system moves a suspended gripper using four motors and cables anchored near a room's ceiling. Neufangled lists an average payload of one point nine kilograms, reaching two point two near the floor, with travel around zero point four meters per second. Its mechanical designs, circuit boards and control software are published. The interface and AI models can run locally, with optional remote access. Wall power reaches the gripper through a support cable. The company says active swing cancellation steadies the hanging gripper without additional motors. A rotating wrist positions its two fingers around objects.

\[Ken\] Commenters are interested in clearing household clutter, while questioning cable routing and physical safety. Finance Tech Bro suggests coordinating it with a Roomba to clear a path before vacuuming. Monk Grilla questions whether cables would intersect furniture when reaching a toy beside a couch. Rider asks about a physical release and whether someone with limited mobility could pull hard enough to trigger it. Creator N H Nifong says each line has a twenty eight newton tension limit, above which the motors release with damped motion. N H Nifong says the threshold can be lowered at the expense of payload capacity, and the work area's ceiling can be lowered to exclude a ceiling fan. James Hart explains that three anchors would confine movement beneath their triangle, while four corners suit a rectangular room.

\[Linus\] On the Layer Three Intel blog: residential proxies detected on the White House network. The company reports proxy traffic exiting through address space registered to the Executive Office of the President. Between October fifth and November third, its monitoring recorded two hundred thirty-six exits through six addresses, across twenty-six days and twenty-eight proxy services. Layer Three Intel says Spur and Synthient independently identify those addresses as residential proxy exits. The company found them through global rotating access, without targeting this network. Its data identifies exits but does not identify the device or person behind them. About half the recorded activity occurred between five p.m. and midnight Eastern time.

\[Ken\] The thread combines concern about network access with skepticism about identifying the physical location from an internet address. Lakeviewterrace says the addresses could belong to a security lab or a guest network for news media. Terr replies that the security question is whether an outsider can send and receive traffic through a node with excessive trust, regardless of its building. C P Burns two thousand nine reports that residential proxies overwhelmed an employer's website, and that web application firewall captcha rules largely stopped the traffic. Random block one offers a check for users investigating their own connection: open IPinfo's What Is My IP page and scroll to Anonymization. According to Random block one, if something appears there, monitor the network to identify what is responsible.

\[Grace\] From Kovid Goyal's Kitty release notes, Kitty zero point fifty. The terminal adds docked windows, session restoration and new controls for named workspaces. Docked panes can hold file trees or status bars. Session restoration includes scrollback; macOS encrypts the saved state using Keychain, while Linux requires the Inhibit desktop portal. Programs can report working, finished, failed or blocked states in the tab bar. The release also fixes a password-authentication bypass in remote control. For failed remote-file uploads, Kitty now preserves the local edited copy and displays its recovery path. The transfer kitten now receives files into temporary files and renames them into place, preventing interrupted transfers from leaving truncated destination files.

\[Linus\] The thread is enthusiastic about features and customization, with differing views on Ghostty and platform integration. Jemmy W says missing Kitty features brought them back after trying Ghostty. Paaloeye prefers Ghostty's more native feel on macOS. Ivo Limmen highlights Control Shift G for opening the last execution log. No Milk uses Control Shift H to navigate terminal history and configured it to open in Neovim when searching thousands of server-log lines. Remedan reports resolving SSH problems by adding a Host star section to the SSH configuration and setting the TERM environment value to xterm dash two five six color with SetEnv. Jonnie Cache describes another customization: having the Fish shell emit a bell on a nonzero command exit, then connecting that bell to a screen-shake shader so failed commands trigger a visible effect.

\[Dennis\] A few more that didn't make the cut today. From Ninety-Nine Percent Invisible, guerrilla bike lanes become permanent in San Francisco. This article from twenty sixteen describes activists installing flexible posts along Golden Gate Park. The city said it would replace them with official posts and leave the temporary installation during its review. Similar efforts in Seattle initially met removal before officials agreed to restore the delineators. Commenters were broadly supportive of faster street-safety improvements, with disagreement over the risks of unauthorized construction and how existing road space should be divided.

\[Ken\] From Tom's Hardware, AMD expects quarterly revenue to overtake Intel's. AMD's fourth-quarter forecast would put its sales ahead of Intel for the first time. AMD projects approximately seventeen billion dollars, compared with Intel's forecast of fifteen to sixteen billion. Data-center sales reached eight point five five billion dollars in the third quarter, supported by Epyc processors and Instinct accelerators. Commenters welcomed stronger competition while debating how much the comparison reflects AMD's progress, Intel's difficulties, and their different positions in AI hardware.

\[Ada\] In The Guardian, what can save Britain's high streets? Reporting from Rotherham examines how retail parks have drawn shoppers away from town centers. Around a quarter of the town center's retail units are vacant. The article describes competition from larger stores and free parking outside town, alongside proposals for housing, social activities and other uses in town centers. Commenters were frustrated by the decline, with disagreement over the roles of rents, online shopping, parking and pedestrian access, and how much conventional retail could return.

\[Grace\] From Nikkei Asia, tech giants' hidden AI obligations reach three point two trillion dollars. Nikkei's analysis puts off-balance-sheet commitments at almost double their level six months earlier. It covers Alphabet, Microsoft, Amazon, Meta and Oracle. Roughly sixty percent consists of purchase commitments, with the rest comprising leases that have yet to begin. The report traces financing links between these companies and cloud startups. Commenters were concerned about financial exposure and spending, while disagreeing over whether investors were adequately pricing the risks.

\[Linus\] On Hacker News, Gemini and Codex chat-training controls. Two posts describe concerns about saved conversations being used for training. In the Gemini post, James L K cites Google's documentation linking model training to the Keep Activity setting on personal accounts, including paid subscriptions. The post identifies turning that setting off or using temporary chats as options. In the Codex post, user zero five two nine one nine reports that the Improve the Model for Everyone setting had become enabled for colleagues who had previously disabled it. The threads show frustration over privacy controls and disagreement about providers' data practices. In the Gemini discussion, rich richardsson reports that turning history off stopped phone reminders from working even though Gemini had said a reminder was set. In the Codex discussion, Chris Kanan says the setting was enabled despite two earlier approved requests to disable training. Jamie N K quotes Gemini's setting description: with activity off, Google still saves chats for seventy-two hours to respond and maintain safety.

\[Ken\] From Conductor, the team is joining Meta. The announcement says a new interface is planned for early next year. Conductor will continue operating until then, while Conductor Cloud will shut down over the next few weeks. The company has provided a way to move cloud workspaces to local machines. Commenters combined appreciation for the product with disappointment about the acquisition and cloud shutdown. The discussion also showed strong interest in replacements, with differing priorities around local development and isolated cloud environments.

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