---
title: "Nov 18: OpenAI launches GPT-6.2 Astra and Sol, and more"
show: "RobotPaper's daily HN podcast"
canonical: "https://podcasts.robotpaper.ai/hn/ep/2026-11-18/"
date: "2026-11-18T03:29:25+00:00"
duration: "29:59"
duration_seconds: 1799
audio: "https://podcasts.robotpaper.ai/hn/ep/2026-11-18/2026-11-18-hn-v2.mp3"
chapters: "https://podcasts.robotpaper.ai/hn/ep/2026-11-18/chapters.json"
transcript: "https://podcasts.robotpaper.ai/hn/ep/2026-11-18.txt"
feed: "https://podcasts.robotpaper.ai/hn/feed.xml"
language: "en"
author: "Roy Osherove"
ai_generated_voices: true
---

# Nov 18: OpenAI launches GPT-6.2 Astra and Sol, and more

RobotPaper's daily HN podcast · Nov 18, 2026 · 30 min

- Listen: [MP3, 29 MB](https://podcasts.robotpaper.ai/hn/ep/2026-11-18/2026-11-18-hn-v2.mp3)
- Subscribe via RSS: https://podcasts.robotpaper.ai/hn/feed.xml (paste into any podcast app)
- Web page: https://podcasts.robotpaper.ai/hn/ep/2026-11-18/
- Chapters (JSON): https://podcasts.robotpaper.ai/hn/ep/2026-11-18/chapters.json
- Transcript (text): https://podcasts.robotpaper.ai/hn/ep/2026-11-18.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
- `00:00:37` GPT-6.2 models
- `00:03:02` School phone bans
- `00:05:22` NSA network outage
- `00:07:42` Apple service failures
- `00:09:38` CMU free courses
- `00:11:46` Optimizely malware
- `00:14:09` Comcast fiber claims
- `00:16:22` AI investment returns
- `00:18:34` Open source upkeep
- `00:20:41` Claude verification
- `00:22:36` OpenCode Browse
- `00:24:40` Loopholes and ethics
- `00:25:15` Nvidia Singapore sales
- `00:25:54` Learning through Rust
- `00:26:30` Huang return statement
- `00:27:06` Geohot on AI power
- `00:27:42` Java Base64 flaw
- `00:28:19` License plate warrants
- `00:28:56` Majestic ranking error
- `00:29:34` Outro

## Show notes

- GPT‑6.2 Astra and Sol
- New Evidence Shows School Phone Bans Are Gradually Improving Test Scores
- Highly classified US national security communications network goes down
- Apple Services Down
- Carnegie Mellon is putting computer science courses on YouTube for free
- Optimizely compromised, served malware on Zoom, GitLab
- Comcast's "fiber-powered Wi-Fi" is cable, AT&T says in false advertising case
- Baidu, Xiaomi CEOs warn of growing gap between AI hype and returns
- We don't need more apps, we need better open source
- How the Claude Code team thinks about verification
- OpenCode Browse
- Loopholes Are for Assholes
- Nvidia stopped disclosing Singapore sales as revenues hit $104B
- Rust changed me as a developer
- Jensen Huang says he’s open to returning in first statement since resignation
- My Empire of Dirt
- Java4Shell: RCE in OpenJDK when processing Base64
- We're Asking an Appeals Court to Require a Warrant for ALPR Searches
- What is w2rt6byl5nf1opmc-a.com?

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 Wednesday, November eighteenth, 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, OpenAI introduces Astra and Sol; its new GPT six point two models.

\[Ada\] And I'm Ada. We've also got school phone bans, and an extended outage at the NSA.

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

\[Dennis\] First up, from OpenAI, GPT six point two Astra and Sol. OpenAI's new models target complex professional projects and everyday work while reducing the cost of completing tasks. The company says Astra improves documents, spreadsheets, code and frontend design. On Agents' Last Exam, its test of multistep professional work, Astra scores five point two higher than the previous Astra at less than a third of the cost. On Terminal-Bench-Science, a test of scientific research workflows in a terminal, OpenAI reports a nine-percentage-point improvement over the previous Astra at less than one third of the cost. Astra's API prices are eight dollars per million input tokens, forty cents for cached input and forty dollars for output. The input and output rates are twenty percent below the previous Astra's. Sol keeps the previous Sol's token prices. On Cognition's FrontierCode benchmark, OpenAI says Sol finishes within half a point of the new Astra at about one fifth of the cost, with both using maximum reasoning effort.

\[Ken\] The comments welcome the lower costs, while questioning how well benchmark gains carry over to everyday work. Chriddy P reports that a data analytics exam fell from about twenty dollars to four dollars with Astra, and from two dollars sixty to one dollar twenty with Sol. Chriddy P found similar accuracy and fewer tool calls, although some runs stopped before finding the requested answer. Topfi criticizes the announcement's slide and workbook examples, saying older models had produced better results in personal use. M Linsey offers a practical evaluation method: replay eight previous projects from their original specifications, compare designs through a blind voting page, and have a review agent search for problems and edge cases that actually occurred. M Linsey says that suite is tailored to personal work and had not yet produced results for either new model. Eli adds a reproducibility precaution: use git archive to supply only the files from the chosen commit, because an agent with the repository can explore later commits and branches. Eli also recreates an isolated container for every run, preventing files left in temporary directories from supplying clues.

\[Linus\] From The New York Times, new evidence shows school phone bans gradually improving test scores. Research on strict school-day phone storage finds modest academic gains, alongside disagreement about how reliably those gains can be measured. A Harvard-led study analyzed three point three million students at forty-six hundred schools. Schools requiring locked pouches throughout the day gained the equivalent of roughly one or two weeks of learning over a school year. The analysis used math and reading scores from grades three through ten across five academic years. A separate study published in May found an overall effect near zero, with modest positive effects in high schools over three years. That study used different tests, phone-location activity, teacher surveys and pouch-provider data. Researcher Thomas Dee questions extrapolating long-term gains from small estimates covering only a few years.

\[Ada\] The thread divides between parents reporting better school experiences and commenters scrutinizing the size of the measured effect. Robot Jesus says children aged sixteen and thirteen enjoy their schools' bans because lunch and other breaks have become more social. Robot Jesus also describes the communication arrangement: during non-class periods, students can go to the office to check a message or send a text. Yorwba puts the study's first-year difference at about one hundredth of a standard deviation, with no detectable additional difference the next year. Rahim Nathwani focuses on the treatment of extreme scores. According to Rahim Nathwani, keeping those observations reduces the estimate from zero point zero zero eight seven to zero point zero zero five eight standard deviations, and it loses statistical significance. Tootie reports a different practical experience: children can no longer photograph the board or check Google Classroom assignments on their phones, and are more likely to miss work. Japhyr, who says they taught high school during the twenty-tens, argues that consistent enforcement must extend from district administrators and principals to individual teachers, because exceptions in one classroom undermine the policy elsewhere.

\[Grace\] Next, from CNN, a highly classified American national security communications network goes down. The National Security Agency restricted its campus to essential personnel while restoring its internal communications network. CNN reports that the agency first staggered arrivals, briefly restored normal reporting instructions, and then closed the campus to most employees on Tuesday night. NSANet carries classified communications within the agency. The NSA attributed the disruption to scheduled maintenance and said mission-critical capabilities remained operational. Director of National Intelligence Jay Clayton said patching major software vulnerabilities had left some systems unable to restart as expected. CNN identifies those vulnerabilities as affecting PHP. Sources told CNN that backup communications could still carry urgent intelligence to the president, and that the Five Eyes intelligence-sharing system was unaffected. A source said network restoration was ongoing and incomplete.

\[Dennis\] The discussion is anxious about infrastructure reliability, with competing explanations centered on staffing, software changes and heightened attention to outages. B Trees of Spring argues that deep technology staffing cuts remove the backup personnel who keep complex systems operating, even when the immediate effects look manageable. The Phyber describes a contact at a government contractor working with technology five to ten years old, spending much of the time getting urgent patches tested and approved or handling changing requirements. Lucas C Dot Net asks whether the apparent surge reflects more failures or greater awareness of them. Someone twelve thirty-four cites shortened testing windows and weak recovery preparation as alternative explanations for recent incidents. Kube System corrects a separate premise in the discussion: the first reporting came from people familiar with the matter, rather than an initial public announcement by officials. Dr Stewart asks critics for objective reliability measurements extending beyond isolated examples. C E D W S adds that organizations need to hire juniors and graduates to maintain their future supply of experienced staff.

\[Ken\] From Apple's developer site, Apple services down. All services showed available at five thirty-four Israel time. Resolved outages included App Store In-App Purchases, Developer ID Notary Service, App Store Connect Analytics and Sales and Trends, Apple Music Connect, Apple Music for Artists, and Xcode Cloud. App Store Connect listed two outages. News Publisher's performance issue was resolved. No incident times were visible.

\[Linus\] The comments are frustrated about interrupted transactions and the dependence of everyday tasks on online services. C J B Connor says an Apple Store could not process orders or returns, and newly purchased iPhones could not be activated. Ruther four D reports driving an hour to collect an order and finding that the store could make no transactions. Ruther four D also says the appointment had not been cancelled despite the outage. Paul Hebert describes testing a new Sign in with Apple flow when authentication stopped working. The unchanged production version failed too, even though it had not been modified for weeks. Apparent reports that a login attempt could not deliver two-factor authentication through any of the available methods. Apparent considered resetting the account, but waited after noticing other unusual page behavior and suspecting a wider problem. Frenchie forty-one eleven initially reported no personal problems, then added that Apple's status page showed a partial outage. On the broader reliability question, Blop calls for more local and diverse alternatives to centralized services. Xyst replies that self-hosting brings costs and maintenance requirements, especially when something like Immich is exposed publicly to share family photos outside a VPN. Xyst argues that those demands make decentralized services difficult to maintain for ordinary users.

\[Ada\] From Carnegie Mellon's website, computer science courses are coming to YouTube for free. CS Learn is releasing full classroom lectures, with self-paced courseware and AI tutoring planned for later. The offering draws on faculty across seven departments, covering subjects including AI, robotics, systems and biology. Machine Learning with Large Datasets and Principles of Computing are already listed as available. Software Architectures and Quality Management are scheduled for November eighteenth, followed by AI Agents and Machine Learning in Production on November nineteenth. The January twenty twenty-seven schedule includes Probability and Computing, Multimodal Machine Learning, and Requirements for Embedded Systems. The planned courseware will provide lessons based on Carnegie Mellon courses, instant feedback and badges. AI tutoring is listed as a forthcoming addition to that courseware.

\[Grace\] The thread welcomes the access, while commenters question the recording quality and ask for practical exercises. P Compeau, who says he runs CS Learn and teaches at Carnegie Mellon, explained that students built tools to automate production, allowing a small team to release courses quickly while people review every lecture. He said these are full semesters of current courses, with more subjects arriving in January. E C Shafer argued that lectures need projects and feedback, and asked for graduate-level material as well as introductory courses. Jacob Rast replied that coursework for Machine Learning with Large Datasets is already available, with test cases in most assignments for checking answers. Robbak reported a recording problem in Principles of Computing: the lecturer's voice became difficult to hear when she moved away from the microphone. Traceroute sixty-six asked for the lectures to be cross-posted beyond YouTube, citing sign-in demands, advertising and unrelated recommendations. In that discussion, Jakzurr reported fewer viewing problems using Windows ten, Chrome and uBlock Origin Lite, and suggested Firefox with uBlock Origin as another option.

\[Dennis\] From Report U R I's blog, an Optimizely compromise served malware on Zoom, GitLab and other major websites. Attackers inserted malicious JavaScript into customer experiments distributed through the platform's trusted infrastructure. Report U R I says visitors to one hundred seventy-eight organizations received malicious code on November sixteenth. The attacker modified the Project JavaScript field in customer snippets, which were then delivered through Optimizely's content delivery network. The payload replaced cryptocurrency addresses, skimmed payment cards, and stole accessible cookies and screenshots. It also delivered the Coruna and DarkSword exploit kits to iPhones and iPads. Report U R I documented cryptocurrency transfers to the attackers. Payment details entered inside hosted Stripe or Adyen payment frames were outside the card skimmer's reach. At twenty-one fifty-two UTC, the researchers found malicious code remaining in one of the one hundred eighty-eight infected snippets.

\[Ken\] The comments are sharply critical of third-party scripts, with disagreement over how to enforce safeguards. Stickfigure said tag managers let marketing teams add code without engineering involvement. Goalieca described the difficulty of maintaining a content security policy when trackers and connection domains change frequently. Robomc replied that a restrictive policy can force review before those additions work. M C Oliver proposed auditing vendor scripts, hosting copies locally and adding delays and phased releases before deploying changes. M C Oliver also called for automatic security checks and internal release controls. Thought-gap recommended NoScript, explaining that it can allow JavaScript from a business's own domain and content delivery network while blocking other domains. Mywittyname reported having previously allowed Optimizely because some sites would not work without it. Stavros distinguished subresource integrity, which checks that a loaded script matches its expected contents, from a content security policy that restricts additional script loading. Jach described enforcing a content security policy, removing unsafe-eval and requiring customers to explicitly allow domains.

\[Linus\] From Ars Technica, AT and T is challenging Comcast's fiber-powered Wi-Fi advertising. A federal lawsuit argues that the wording gives customers a misleading impression of fiber delivered directly to their homes. Comcast's main residential network combines fiber with coaxial cable for the final connection. AT and T filed the case in the Northern District of California and seeks an injunction stopping the fiber-powered description, along with financial relief. The article also recalls AT and T's own advertising of four G service as five G Evolution, which the advertising industry's review body found likely to mislead consumers. In June, that body found Comcast accurately described its mix of cable and fiber in a disclosure, but placed the disclosure too inconspicuously. Comcast lost one hundred forty-four thousand residential broadband customers in the third quarter, compared with ninety-one thousand a year earlier.

\[Ada\] Commenters criticize the marketing, while disagreeing about which network differences customers should prioritize. Userbinator pointed out that DOCSIS can provide symmetric speeds and that fiber service can be configured asymmetrically. Jacobgold cited Comcast's DOCSIS four point zero offerings with symmetric one-gigabit and two-gigabit connections, arguing those can meet customers' expectations of fast service. Ruler Of described coaxial faults caused by water infiltration, bad crimps and unterminated connections, which can make intermittent failures difficult to diagnose. Userbinator replied that acquaintances had also experienced intermittent failures after upgrading to fiber. Segmondy reported a T-Mobile installation with fiber entering the house and one-gigabit upload and download speeds. Bityard said T-Mobile's offer in their neighborhood provided twice the speed for half the price of their existing Comcast plan. Bri three D explained that T-Mobile Fiber often uses acquired or partner networks, identifying Intrepid Fiber Networks as the underlying provider in their area.

\[Grace\] From the South China Morning Post, Baidu and Xiaomi executives warn that AI investment is outrunning commercial returns. Their remarks focus on the difficulty of turning stronger models into profitable applications. Baidu's Robin Li and Xiaomi's Lei Jun spoke at the APEC CEO Summit in Shenzhen on Tuesday. Li called for high-value applications and changes to organizational structures that would let businesses benefit from AI. Lei said successful commercial deployments remain limited, while predicting opportunities in smartphones, vehicles and humanoid robots. Baidu reported a forty-six percent decline in adjusted net profit for the September quarter, to one point four billion yuan, as AI gains failed to offset falling advertising revenue. Li cited Baidu's work at the Port of Qingdao, where specialized agents identify efficient ways to load shipping containers. Baidu also plans a one-thousand-twenty-four-chip Tianchi supernode next year.

\[Dennis\] The thread splits between disappointment with current returns and arguments that adoption takes time. Secret Dreams criticized the repeated production of similar applications and the lack of original ideas. J N Watson argued that current investment requires productivity gains across many fields, beyond software development. Origini described integration work still needed inside companies, including agent environments, connectors and security approval. Will Adams supplied a concrete example of limits in document processing. While automating the renaming of scanned checks and invoices, Microsoft upload limits reduced throughput to eighty to one hundred documents daily. Will Adams then tried a local model, which took four hundred five seconds to process two files on a work laptop. By hand, assisted by a batch file assembled from a spreadsheet, Will Adams could process about two hundred per hour. In another branch, P Salaun described building a family recipe application to replace Airtable and a work prioritization tool layered over an existing ticketing service.

\[Ken\] Over on Jordan Krueger's blog, We Don't Need More Apps, We Need Better Open Source. The post calls for AI-assisted coding to improve the installation, performance and upkeep of existing public software. Krueger helps nonprofits with technology and runs a Matrix community using Synapse, Element and a moderation bot. He describes capable software whose deployment needs exceed the staff and budgets of small organizations. His nonprofit clients often lack a part-time system administrator. His proposed work includes simpler installers, lower memory use, built-in backups, safer upgrades and error messages that explain the next step. He also asks contributors to discuss changes with maintainers, measure improvements and keep patches small and tested. The post cites curl's experience with AI-assisted security reports, where even useful submissions can create more review work than maintainers can handle.

\[Linus\] The comments are divided over whether more automated contributions will help maintainers or overwhelm them. Josh data says their projects were barely sustainable when submissions addressed serious logic problems. They now face patches for annoyances, hypothetical problems and performance edge cases, and may stop accepting contributions. Lin somniac proposes project-specific triage rules that let an agent identify issues already rejected as won't-fix. Lin somniac also describes using an agent to reproduce a slow APT package-cache problem by trying different cache versions, package updates and client configurations, requiring about two minutes of personal effort. Necovek responds that running agents on strangers' pull requests creates an attack surface and can spend the maintainer's money. Valdik S S suggests a Social dot M D file that explains a project's support and maintenance expectations. Valdik S S says an open-source license alone does not promise ongoing fixes or fulfillment of feature requests.

\[Ada\] From the Claude blog, How the Claude Code Team Thinks About Verification. The team describes automated checks and testing of running applications as the basis for delegating more coding work to agents. It says Claude writes and reviews about ninety percent of its code. The article recommends end-to-end tests for complete user flows, property tests that generate many inputs, and differential tests that compare old and new systems. It names fast-check for JavaScript and TypeScript, Hypothesis for Python, and proptest for Rust. The team also puts instructions for launching and inspecting an application into a verify skill near the relevant code. When an agent cannot fix a failing check after three rounds, it brings in a person. Code owners write and refine the checks. The team maintains a reliability checklist with thirty-nine incident-linked items.

\[Grace\] The thread is skeptical about token costs, with disagreement over how much verification should depend on an agent. Other Shrezzing argues that recurring browser checks can be written once in Playwright and run deterministically at almost no marginal token cost. X nor swap recommends starting with static analysis and other cheap deterministic checks. P U P E counters that an agent examining an error with the full application context could discover different problems and handle edge cases more smoothly. Alien Robot describes another failure mode: generated unit tests that mock every call and assert how often implementation methods run. Alien Robot says those tests break when the implementation changes while checking behavior that often does not matter. Koolba disputes the claim that nobody can review every change. Koolba recommends breaking work into pieces small enough for human review, as with manually written changes.

\[Dennis\] Next up, from OpenCode, OpenCode Browse. An MCP server and Chromium extension connect agents to an existing browser session, including its signed-in sites and extensions. Agents can open background tabs in collapsed groups, inspect pages, click, type, capture screenshots and record tabs. When a login, two-factor prompt or payment requires human participation, Browse hands the tab back and resumes after the person clicks Continue. A toolbar button can lend an existing tab to an agent, and allowed-site settings limit where its tab can go. The product page lists Chrome, Edge, Brave and other Chromium browsers, with clients including OpenCode, Claude Code, Cursor, Codex and Zed. The extension also captures network traffic, and the product page says it works with any MCP client. Installation starts with the browser extension, which guides the remaining setup.

\[Ken\] Commenters welcome access to existing sessions but disagree sharply about the security model. Tornikeo reports two weeks of heavy use without encountering bugs, highlighting background tabs and switching between browser profiles. Tornikeo uses its command-line interface. Test test me test criticizes OpenCode's release practices and points to Chrome's existing DevTools Protocol as an alternative, naming the package go-rod for browser control. Zap the impaler argues for enforcing restrictions in an external sandbox to avoid repeated permission prompts. R punk F U responds that filesystem restrictions do not control everything an agent can send to GitHub or an external API. Jeremy Jacob says Playwright's MCP server also grants access to already-open pages and sessions, and recommends separate profiles for agent use. I H S W adds that Firefox has its own MCP server with profile selection, identifying that control as important for isolation.

\[Linus\] A few more that didn't make the cut today. From Acts of Volition, Loopholes Are for Assholes. The essay examines everyday rule-breaking through the question of what would happen if everyone behaved the same way. The author recalls once admiring an adult's advice to park on a shop's sidewalk, then later rejecting that advice. Commenters debated civic responsibility and the limits of universal moral rules. Much of the disagreement concerned whether ignoring an unenforced parking restriction counts as a loophole.

\[Ada\] From the Financial Times, Nvidia stopped disclosing Singapore sales as revenues reached one hundred four billion dollars. The company changed its geographic reporting from customer billing addresses to customer headquarters. Its Singapore subsidiary booked that revenue in twenty twenty-five, up seventy-two percent in a year. The subsidiary serves customers around the world, and only a small quantity of chips passes through Singapore. Commenters split between concerns about reduced transparency and explanations of international invoicing, with disagreement over what these figures reveal about the chips' final destinations.

\[Grace\] Over on the Startifact blog, How Rust Changed Me as a Developer. The essay describes how a different programming language can reshape familiar approaches to software design. After more than six years using Rust, including more than four professionally, the author highlights enums and compiler-enforced handling of different variants as useful additions to their approach to abstraction. Commenters welcomed learning across language families, while debating Rust's type design and safety guarantees. Other reactions questioned how these lessons fit a profession increasingly shaped by coding models.

\[Dennis\] From CNBC, Jensen Huang says he is open to returning to Nvidia. His first public statement since resigning promises cooperation with the board's independent investigation. Huang says he regrets the problems he created for the board and would return as chief executive if directors decided that was in the company's interests. CNBC's sources said the board was not considering his return. Commenters argued over corporate governance, accountability, and the evidence behind the allegations, with sharply different views of whether a comeback would be justified.

\[Ken\] From the Geohot blog on GitHub Pages, My Empire of Dirt. The post argues that an economy dominated by autonomous agents could exclude human workers, customers, and even the companies developing them. The author criticizes Anthropic's approach and predicts that competition over powerful AI could produce a world war. Commenters were divided over those predictions, with many rejecting the war scenario and others debating the economic consequences of automation. The discussion also included concern about the post's pessimism and arguments over geopolitical competition.

\[Linus\] From Calif, Java four Shell in Open JDK's Base sixty-four decoder. Researchers describe a critical memory-corruption flaw in optimized decoding code. They say malformed input can overwrite neighboring memory and enable information leaks or remote code execution without credentials. The report gives the vulnerability a severity score of nine point two out of ten. Commenters praised the technical research while debating the practical exploitation constraints, the range of affected systems, and how urgently operators should apply patches or temporary mitigations.

\[Ada\] From the Electronic Frontier Foundation, warrants for license-plate database searches. The organization and the ACLU have filed a brief in a federal appeals case over police tracking a suspect's movements without a warrant. Investigators searched Flock and ELSAG databases using a car's make, model, and color after robberies around Cleveland. The groups argue that this tracking is a search under the Fourth Amendment. Commenters supported stronger privacy protections while debating retention limits and the investigative value of historical location records.

\[Grace\] From Jack's blog, the mystery of a highly ranked domain. An investigation traced an unexplained entry in website rankings to an internal experiment. The domain appeared eighth in Majestic's rankings despite failing to resolve. In a November sixteenth update, the author quotes a Majestic employee saying internal experiments reached public results because a filter was missing. Commenters examined the ranking anomaly with curiosity and skepticism, while debating the article's presentation. The thread also expressed frustration with the difficulty of establishing what the domain actually did.

\[Dennis\] That's RobotPaper's daily HN podcast for Wednesday, November eighteenth. 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.
