---
title: "Nov 3: Astra edits Qwen's model weights, and more"
show: "RobotPaper's daily HN podcast"
canonical: "https://podcasts.robotpaper.ai/hn/ep/2026-11-03/"
date: "2026-11-03T03:32:13+00:00"
duration: "29:01"
duration_seconds: 1741
audio: "https://podcasts.robotpaper.ai/hn/ep/2026-11-03/2026-11-03-hn.mp3"
chapters: "https://podcasts.robotpaper.ai/hn/ep/2026-11-03/chapters.json"
transcript: "https://podcasts.robotpaper.ai/hn/ep/2026-11-03.txt"
feed: "https://podcasts.robotpaper.ai/hn/feed.xml"
language: "en"
author: "Roy Osherove"
ai_generated_voices: true
---

# Nov 3: Astra edits Qwen's model weights, and more

RobotPaper's daily HN podcast · Nov 3, 2026 · 29 min

- Listen: [MP3, 28 MB](https://podcasts.robotpaper.ai/hn/ep/2026-11-03/2026-11-03-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/2026-11-03/
- Chapters (JSON): https://podcasts.robotpaper.ai/hn/ep/2026-11-03/chapters.json
- Transcript (text): https://podcasts.robotpaper.ai/hn/ep/2026-11-03.txt
- Show: [RobotPaper's daily HN podcast](https://podcasts.robotpaper.ai/hn/) (RSS: https://podcasts.robotpaper.ai/hn/feed.xml)

## Chapters

- `00:00:00` Intro
- `00:00:36` Editing model weights
- `00:02:53` Ubuntu patch dispute
- `00:04:58` Postgres bug backlog
- `00:07:18` Freezer cyberattacks
- `00:09:25` Changing baby names
- `00:11:30` Silent speech glasses
- `00:13:29` Open SaaS alternatives
- `00:15:38` November hiring
- `00:18:25` Anthropic IPO finances
- `00:23:06` BYD electric sales
- `00:25:08` Tech jobs in the boom
- `00:25:51` Soviet math textbooks
- `00:26:32` Labor share declines
- `00:27:09` Vehicle backup power
- `00:27:44` UK phone migration
- `00:28:23` Outro

## Show notes

- [Astra can read and write weights to improve Qwen](https://rikiyat.github.io/blog/weight-space-programming/) · [HN discussion](https://news.ycombinator.com/item?id=50341174)
- [WTF Ubuntu?](https://chalmovsky.com/posts/wtf-ubuntu) · [HN discussion](https://news.ycombinator.com/item?id=50337608) \| quoted: [BoppreH](https://news.ycombinator.com/user?id=BoppreH), [ASalazarMX](https://news.ycombinator.com/user?id=ASalazarMX)
- [Postgres 19 is late because so many more bugs are being found now](http://rhaas.blogspot.com/2026/10/postgres-19-is-late-and-its-ais-fault.html) · [HN discussion](https://news.ycombinator.com/item?id=50312408) \| quoted: [Gigachad](https://news.ycombinator.com/user?id=Gigachad), [CharlesW](https://news.ycombinator.com/user?id=CharlesW)
- [Supermarket freezers across Europe hit by second suspected cyber-attack](https://www.theguardian.com/world/2026/nov/02/supermarket-freezers-across-europe-hit-by-second-suspected-cyber-attack) · [HN discussion](https://news.ycombinator.com/item?id=50335615)
- [US boy names that became girl names and vice versa](https://namedaisy.com/blog/boy-names-that-became-girl-names) · [HN discussion](https://news.ycombinator.com/item?id=50339397)
- [Show HN: Subvocal – Talk to your computer without making noise](https://news.ycombinator.com/item?id=50340088) · [HN discussion](https://news.ycombinator.com/item?id=50340088) \| quoted: [aarushyadav](https://news.ycombinator.com/user?id=aarushyadav), [IHLayman](https://news.ycombinator.com/user?id=IHLayman)
- [OccupySaaS](https://occupysaas.com/) · [HN discussion](https://news.ycombinator.com/item?id=50331402) \| quoted: [MonkeyClub](https://news.ycombinator.com/user?id=MonkeyClub)

### November hiring (2 stories)

Two stories on Hacker News cover November hiring: Who wants to be hired, and Who is hiring.

- [Ask HN: Who wants to be hired? (November 2026)](https://news.ycombinator.com/item?id=50338726) · [HN discussion](https://news.ycombinator.com/item?id=50338726) \| quoted: [natthan](https://news.ycombinator.com/user?id=natthan)
- [Ask HN: Who is hiring? (November 2026)](https://news.ycombinator.com/item?id=50338728) · [HN discussion](https://news.ycombinator.com/item?id=50338728) \| quoted: [justicz](https://news.ycombinator.com/user?id=justicz), [AdriaanvRossum](https://news.ycombinator.com/user?id=AdriaanvRossum)

### Anthropic IPO finances (4 stories)

From CNBC, Anthropic reports seventeen billion dollars in quarterly revenue and nearly nine hundred billion in commitments.

- [Anthropic reports $17B in Q3 revenue, commitments of $895B in revised S-1](https://www.cnbc.com/2026/11/02/anthropic-ant-s-1-quarter-financials.html) · [HN discussion](https://news.ycombinator.com/item?id=50337439) \| quoted: [cmiles8](https://news.ycombinator.com/user?id=cmiles8), [imeanwhatdoikno](https://news.ycombinator.com/user?id=imeanwhatdoikno)
- [Anthropic's internal AI use, at retail prices, now exceeds its employee pay](https://www.businessinsider.com/anthropic-claude-internal-ai-use-exceeds-employee-pay-tokens-2026-11) · [HN discussion](https://news.ycombinator.com/item?id=50339494) \| quoted: [Aurornis](https://news.ycombinator.com/user?id=Aurornis), [JohnMakin](https://news.ycombinator.com/user?id=JohnMakin), [crazygringo](https://news.ycombinator.com/user?id=crazygringo), [cmiles8](https://news.ycombinator.com/user?id=cmiles8)
- [Anthropic's growth just flattened](https://www.regenerator1.com/p/anthropics-growth-just-flattened) · [HN discussion](https://news.ycombinator.com/item?id=50344958) \| quoted: [chunky1994](https://news.ycombinator.com/user?id=chunky1994), [subtlesoftware](https://news.ycombinator.com/user?id=subtlesoftware)
- [The SEC Should Block Anthropic's IPO to Protect the Public](https://www.techpolicy.press/the-sec-should-block-anthropics-ipo-to-protect-the-public/) · [HN discussion](https://news.ycombinator.com/item?id=50344919) \| quoted: [marsten](https://news.ycombinator.com/user?id=marsten)
- [BYD sold more EVs last month than the US did all quarter](https://electrek.co/2026/11/02/byd-sold-more-evs-last-month-than-the-us-did-all-quarter/) · [HN discussion](https://news.ycombinator.com/item?id=50344181) \| quoted: [Aozora7](https://news.ycombinator.com/user?id=Aozora7)
- [Jobless tech workers are being left out of San Francisco's AI boom](https://www.bloomberg.com/news/features/2026-09-17/san-francisco-s-ai-wealth-boom-skirts-engineers-as-openai-anthropic-ipos-near) · [HN discussion](https://news.ycombinator.com/item?id=50337454)
- [Why are Soviet math textbooks so hardcore in comparison to US textbooks? (2017)](https://www.quora.com/Why-are-Soviet-mathematics-physics-textbooks-so-insanely-hardcore-in-comparison-to-US-textbooks/answer/Scott-Miller-307?share=1) · [HN discussion](https://news.ycombinator.com/item?id=50337725)
- [U.S. workers took home their smallest share of capital since at least 1947](https://fortune.com/article/why-did-us-workers-just-take-home-smallest-share-of-capital-since-1947-lowest-ever-recorded/) · [HN discussion](https://news.ycombinator.com/item?id=50330939)
- [Bidirectional charging during grid outages](https://www.jefftk.com/p/bidirectional-charging-during-grid-outages) · [HN discussion](https://news.ycombinator.com/item?id=50319749)
- [Landline phones will switch off across Britland in 100 days, no extensions](https://www.theregister.com/networks/2026/11/01/landline-phones-will-switch-off-across-britland-in-100-days-no-extensions-openreach/5306673) · [HN discussion](https://news.ycombinator.com/item?id=50326609)

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 Tuesday, November third, 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, researchers use Astra to edit a smaller model's weights.

\[Ada\] And I'm Ada. We also examine Ubuntu's security response, and cyberattacks on supermarket freezers.

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

\[Linus\] First up, from Rikiya Takehi's blog: Astra can interpret and modify weights to improve Qwen. Researchers report that a frontier language model corrected a smaller model's reasoning loops by editing a small fraction of its neural network weights.

The team calls the approach Weight-Space Programming. In experiments with small game-playing networks, Astra identified what games the networks had learned from their weights alone. It also generated networks for simple games without training or testing them beforehand.

The Qwen experiment used a different setup. Astra had a Python environment for inspecting and changing weights, plus a limited allowance of forward passes on practice problems. Gradient descent and training were prohibited. With fifty forward passes, it located a direction associated with repetitive reasoning and changed about zero point zero one six percent of the model's weights. Under the researchers' evaluation settings, MATH five hundred accuracy rose from seven point nine percent to forty-eight point one percent. Instruction following improved too.

The distinction is that the edit helped Qwen finish answers it could already work out. A separate attempt with a smaller Gemma model that rarely looped did not establish a mathematics improvement. These results suggest a possible route to targeted model repairs, while leaving broader capability gains unproven.

\[Ada\] The discussion combines excitement about interpreting weights with caution about the experiments. Vessenes found the prospect of recognizing and repairing another network's behavior deeply surprising.

L9o emphasized that repetition blocked Qwen's existing mathematical ability, while the Gemma experiment failed to establish improvement.

Jsenn distinguished identifying tiny networks without tools from investigating Qwen through a Python harness and experimental feedback. Presenting both as one tool-free capability would blur the evidence.

Ex-aws-dude asked whether targeted modifications preserve behavior across other domains. Improved benchmark scores leave a further question about whether the same edits introduce different failures elsewhere.

\[Ken\] Next, on the Chalmovsky blog: W T F, Ubuntu? A server operator criticizes Ubuntu's response to a severe Linux kernel vulnerability and says the delay has destroyed their trust in the distribution.

The post concerns C V E twenty twenty-six, nine eight four three two, a flaw in the kernel's T C P code. The author cites a severity score of nine point eight and Anthropic's report of an exploit capable of gaining root access. Debian, the post says, shipped a correction on October sixteenth.

Writing on November second, the author said Ubuntu's supported releases were still listed as vulnerable, criticized its promises of timely security updates, and described taking personal servers offline while preparing to migrate. That broad description of Ubuntu's patch availability is challenged in the supplied discussion.

Administrators need to understand which systems are affected and when an approved update becomes available, especially when a distribution sells security support. The issue attracting attention is the relationship between vulnerability severity, release testing and communication.

\[Grace\] The thread is critical of Ubuntu, but commenters dispute patch status and urgency. BoppreH criticized the high rather than critical priority and the inadequate explanation.

Throw0101c challenged the post's blanket claim, saying patched kernels were in proposed repositories for long-term-support releases, and the twenty-six L T S kernel reached updates and security repositories on October twenty-eighth. They still questioned why other releases were waiting.

ASalazarMX defended cautious kernel releases because emergency updates can cause failures, and practical exploitability matters alongside severity scores. Xbar replied that this does not excuse silence or delay for a highly rated network-stack flaw.

The debate concerns both release decisions and the information administrators receive while waiting for approved updates. Clearer explanations would help them understand the choices being made.

\[Dennis\] From Robert Haas's blog: Postgres nineteen is late because many more bugs are being found. The PostgreSQL developer attributes the delayed release to an increase in AI-assisted bug discoveries that has outpaced the maintainers' ability to review and integrate fixes.

Haas writes that the release is likely to arrive in November, beyond the project's usual September-to-October window, and could slip further. Several substantial features were removed late in beta, including property graph queries and operations for merging and splitting partitions.

His comparison of corresponding development periods shows back-patched commits increasing by more than sixty percent. Fixes reaching an earlier stable release increased by more than seventy-four percent. His interpretation is that many longstanding defects are now becoming easier to discover, rather than developers suddenly making substantially more mistakes.

Finding problems and safely fixing them remain different tasks. Generated patches may be incorrect or fail to meet PostgreSQL's standards, so human review is still necessary. Security work on older releases also took time away from reviewing the new version. Haas considers the immediate delay a consequence of prioritizing quality, while asking how the project can sustain innovation under a continuing maintenance burden.

\[Linus\] Commenters welcome better bug discovery but question how maintainers can manage the workload. CharlesW favored delays over shipping longstanding defects.

Gigachad described a workplace where urgent security tasks had halted feature development, illustrating the competition between maintenance and new capabilities.

Tibbar described AI-assisted PostgreSQL investigations that found serialization anomalies, foreign-key integrity problems and planner bugs. They had to prioritize which issues to report and praised the maintainers' responsiveness.

Droppedasbaby identified verification as the constraint: additional reports do not expand the small group available to validate them and review changes. Finding problems faster therefore does not automatically produce faster fixes.

\[Ada\] Next up, from The Guardian: supermarket freezers across Europe hit by a second suspected cyberattack. Retailers in at least seven countries reported refrigeration failures on Monday, with several companies confirming hacking and some temperatures reaching fifty degrees Celsius.

The affected countries include Belgium, Estonia, Ireland, Finland, Luxembourg, the Netherlands and Sweden. Ahold Delhaize reported incidents in dozens of shops in Belgium and Luxembourg. Ireland's SuperValu said a specific refrigeration model was affected in thirteen stores, and removed chilled and frozen goods from sale.

The incidents follow last month's attack in France, which affected as many as seven hundred supermarkets and caused food losses estimated in the tens of millions of euros. In the latest Belgian incidents, one refrigeration operator said evaporators had been switched into defrost mode.

Authorities have not identified who was responsible, and there have been no claims of responsibility. The article reports no financial demands. The wider concern is that remotely accessible control systems can translate a software intrusion into immediate damage to food supplies and shop operations.

\[Ken\] Commenters are frustrated about exposed equipment but disagree over necessary connectivity. Pbohun favored freezers without Internet access and criticized unnecessary automation.

Chromaton, who said they worked in a related field, explained that industrial systems require more than a household thermostat. Defrost cycles, compressor behavior and temperature ranges affect energy use and maintenance.

Xoa argued that restricted networks and virtual private networks preserve useful remote access without exposing controllers to everyone. Prennert countered that retailers face many vendors, legacy equipment and difficult coordination, favoring immediate inbound disconnection to contain damage.

The debate is how to restrict access while preserving necessary operations across these complex systems.

\[Grace\] Next up, from the NameDaisy blog: U.S. boy names that became girl names, and vice versa. An analysis of Social Security records finds that changes in girls' naming patterns account for most of the shifts in both directions. The study covers nineteen thirty-six through twenty twenty-five. Forty-two names met its definition of moving from mostly boys to mostly girls, compared with twelve moving the other way. Thirty-four of the shifts toward girls mainly reflected increased use among girls. Nine of the shifts toward boys mainly reflected declining use among girls. The definition requires a sixty-percent share for one sex across two consecutive five-year periods, then the same threshold for the other sex, with minimum birth counts. The researchers also account for uncertainty where rare annual names are omitted. The broader visualization also includes names that changed substantially without meeting that test. Ashley, for example, was too uncommon during its predominantly masculine years to qualify. The distinction matters because a changing proportion can reflect either new popularity or the decline of an older naming fashion.

\[Dennis\] The comments combine interest with questions about the categories. Billyp-rva sees girls' naming trends as more dynamic. Conductr suggests family traditions keep boys' names more stable, citing a son named after a grandfather. Gregates emphasizes that predominantly masculine names were not necessarily popular. Their Ashley example contrasts a small earlier population with its later surge among girls, alongside growth among boys. Ecshafer notes that identical spellings can represent different names across languages, as with Jan and Jean. Mikepurvis says information about languages spoken at home would help interpret those cases. The discussion distinguishes proportions from popularity and cultural context.

\[Linus\] Posted on Hacker News: Subvocal, talk to your computer without making noise. The Subvocal Company is developing smart glasses that turn silent mouth movements into text using radio-frequency sensing. Founders Aarush and Arnav describe moving away from facial electrodes because sweat, interference and placement changes complicated the signal. Their glasses instead detect how radio waves change as they interact with the tongue, lips and soft palate. Models translate those patterns into text. The founders say they expanded training data across users to address differences in facial anatomy, and their demonstration includes a short calibration. The hardware remains a prototype, with shipping targeted for the first quarter of twenty twenty-seven. The system senses articulation, and the post distinguishes that from the automatic movements associated with thinking words. Closed-mouth articulation is a further experimental mode. It seems the appeal is dictation in libraries, offices or public places without speaking aloud.

\[Ada\] The thread is enthusiastic, but cloud processing and readiness are central concerns. Sourpanda highlights accessibility and hands-free uses. Founder aarushyadav says decoding uses a companion app and cloud processing, usually with latency below five hundred milliseconds. Initially, a key on the connected device activates text input. Vessenes also asks about adjustable fit. IHLayman objects to cloud transmission and questions operation when connectivity drops. Vessenes argues local processing would reduce privacy concerns. For experimental closed-mouth articulation, aarushyadav says testing involved only the founders and one employee. Learning the behavior takes too long for it to be a primary initial feature, distinct from the demonstrated silent mouthing.

\[Ken\] From the project's own site, OccupySaaS. The initiative proposes open-source replacements for commercial software services, built from public documentation with AI assistance. It presents the work as a response to layoffs, vendor lock-in and per-seat pricing. Its method separates specification writers from implementers, excludes confidential source material, and tests compatibility using vendors' public client libraries. Humans are meant to review changes. The initial projects are ZeroAuthy, an Auth0-compatible identity server, and SalesCompel, a Salesforce-compatible system with a Go and PostgreSQL core and a Vue record editor. The listed capabilities include identity and token flows on one side, and account and contact operations with limited queries on the other. Both are self-hostable and released under the Apache two-point-oh license. The site's FAQ explicitly says neither project is production-ready. Their broader compatibility ambitions are therefore goals still being developed. The interest seems to center on whether cheaper implementation can make switching providers easier while preserving existing integrations.

\[Grace\] The discussion splits between support for alternatives and skepticism about enterprise requirements. Temp123477 argues that enterprise customers purchase a continuing service rather than just the software code itself. Ozim cites security updates, regression testing, backups and performance work. Akersten counters that some companies mainly need software, with support available separately. Baubino questions starting new projects when established free-software alternatives exist. Wackget highlights compatibility with existing client libraries and APIs as the intended distinction. MonkeyClub worries that responding to layoffs with more AI reinforces the same shift. Swiftcoder argues newcomers should be welcomed regardless of former roles. The debate includes whether these replacements can become dependable over time.

\[Dennis\] Two stories on Hacker News cover November hiring: Who wants to be hired, and Who is hiring. The paired threads let individual job seekers and employers describe their needs directly. In Who wants to be hired, participants list their location, remote-work preference, willingness to relocate, technologies and résumé information. The post restricts participation to people personally looking for work, with agencies and recruiting businesses excluded. The separate Who is hiring post asks for one entry per company, written by someone who actually works there. Companies must be actively filling a position and committed to replying to applicants. They are also asked to state whether a job is onsite or remote, including any country restrictions. The job-seeker post asks readers to use contact details only for work opportunities. The employer post similarly asks that applicants be personally interested in the position. These details matter because the same broad job title can describe very different working arrangements. The posts make location and eligibility part of the initial description, alongside skills and experience. Together they offer a view of the particular opportunities and preferences people are willing to publish this month.

\[Linus\] The comments are practical introductions with varied preferences. In the job-seeker thread, natthan describes a Vancouver search with Unix experience, excluding AI, surveillance, marketing and cryptocurrency work. Yzy98, based in China, seeks remote full-stack or senior frontend work, preferably in AI, developer tools or software services. Sepiligo, in Melbourne, describes former Google experience and a computer-science doctorate, with flexibility on relocation and employment arrangements. Namusanga, in Kampala, focuses on GPU kernels and performance, with previous remote work for British and American companies. On the employer side, justicz advertises onsite software and hardware positions at Charge Robotics in the Bay Area, building robots for solar-farm construction. AdriaanvRossum offers worldwide remote DevOps work at Simple Analytics, focused on its own servers. Ayewo asks how advertised monthly pay relates to reduced hours, and AdriaanvRossum confirms it scales with contracted hours. Those exchanges show how location and working arrangements matter alongside technical specialization. The preferences span both sector choices and different employment terms.

\[Ada\] From CNBC, Anthropic reports seventeen billion dollars in quarterly revenue and nearly nine hundred billion in commitments. Four stories examine the company's finances and internal AI use as it prepares to go public. The amended filing covers the third quarter, which was missing from the initial prospectus. CNBC reports that revenue rose about forty-seven percent from the previous quarter. That is substantial growth, but slower than the more than doubling recorded in each of the preceding quarters. Revenue for the first nine months reached thirty-three point two four billion dollars. The reported net loss for that period was one hundred four billion dollars, mostly attributed to changes in the fair value of convertible notes and derivative liabilities. Those accounting losses should not be confused with an equivalent cash outflow. Separately, purchase commitments and lease agreements total about eight hundred ninety-five point four billion dollars, substantially all for computing and infrastructure. These are aggregate commitments, not one quarter's expenses. CNBC also describes investor concern about financing costs and whether computing capacity will arrive quickly enough.

Business Insider adds a different measure of AI use inside Anthropic. In April, internal Claude consumption, valued at customer prices, exceeded monthly employee compensation excluding stock awards. That does not mean Anthropic paid more cash for its own AI than for salaries. The article reports that September usage was more than two hundred fifty times January's, and that Claude wrote more than ninety percent of code added in September. Its comparison of revenue with headcount growth is evidence of different growth rates, rather than a direct measurement of individual productivity.

A post on Regenerator argues that Anthropic's growth has flattened. The author combines earlier reported quarterly figures with a reported July annualized revenue rate to estimate modest growth during August and September. These are inferred monthly figures, not disclosed monthly accounts. The post considers seasonality, capacity shortages and weaker spending growth as possible explanations.

Finally, at Tech Policy Press, Matt Scherer argues that the SEC should block Anthropic's IPO until it supplies two consecutive quarters of audited financial statements and fully discloses material spending commitments. He describes competitive pressure, infrastructure costs and risks to retirement savers. That is his policy recommendation, rather than an announced regulatory decision. It seems these reports are drawing attention because rapid revenue growth, long-term obligations and evidence of internal AI adoption answer different questions about the company's prospects.

\[Ken\] The threads are divided over valuation, accounting and what investors should be allowed to decide for themselves. Under the CNBC report, cmiles8 argues that the spending commitments and the proposed scale of the valuation make the slowdown concerning. That commenter also says losses linked to convertible instruments can matter through dilution, even when they are not immediate cash expenses. In reply, imeanwhatdoikno argues that the revenue already being generated makes the lower valuation suggested in that discussion implausible.

The Business Insider thread focuses on the retail-price comparison. Hungryhobbit argues that an internal tool should be measured by its actual cost to the company. AI-x counters that computing used internally has an opportunity cost if it could instead be sold to customers. Aurornis says retail pricing also gives outside readers a usable comparison. JohnMakin asks why such heavy automation accompanies continued headcount growth, while crazygringo argues that better productivity can make both additional staff and additional computing worthwhile.

Under the Regenerator post, subtlesoftware points out that the analysis used earlier reported numbers and contrasts them with the newer quarterly disclosure. Modeless expects the late-September model release to support further usage, while chunky1994 questions whether Anthropic's advantage over competitors is as strong as claimed.

In the Tech Policy Press thread, marsten argues that the SEC's role is to ensure adequate disclosure, rather than decide whether a company is a good investment. Gerdesj pushes back by emphasizing investor protection. Other commenters debate whether index-fund investors can meaningfully avoid the risks being discussed.

\[Grace\] From Electrek, BYD sold more electric vehicles in October than the United States sold during the entire third quarter. The comparison puts one manufacturer's global monthly sales against a national market over three months. BYD sold three hundred five thousand seven hundred forty-three battery electric vehicles in October, up thirty-seven percent from a year earlier. Electrek compares that with Kelley Blue Book's estimate of just over two hundred sixty-two thousand five hundred electric vehicles sold in the United States during the third quarter. BYD's broader new-energy-vehicle total, which also includes plug-in hybrids, reached four hundred ninety thousand one hundred sixty-eight. The categories are different and should not be combined when making the electric-only comparison. Battery electric vehicles accounted for about sixty-three percent of BYD's total sales in October, while plug-in hybrid sales declined. The company also reported exporting about one hundred seventy thousand vehicles during the month. It seems this is drawing attention because BYD's electric sales and overseas expansion contrast with the smaller American electric-vehicle market.

\[Dennis\] The comments divide between criticism of American policy and debate about what buyers value in a car. Paxy attributes weakness in the American market to changes in subsidies, charging support and regulation. Iamjake648 pushes back on the claim that domestic manufacturers have largely abandoned electric vehicles, pointing to continuing product and charging investments. A separate exchange focuses on affordability and quality. Diath describes disappointing interiors in BYD vehicles they used as a passenger. Dabinat replies that lower purchase and running costs can matter more to many buyers than premium materials. Aozora7 shifts the discussion toward overnight charging at home, while noting the difficulty for people in apartments.

\[Linus\] A few more that didn't make the cut for full coverage. From Bloomberg, jobless tech workers are being left out of San Francisco's AI boom. The September report follows unemployed workers gathering for hikes and interview preparation as demand shifts toward AI-related roles. It describes declining demand for conventional software engineers alongside growth in jobs helping customers deploy AI. Commenters debated whether investment priorities, automation or hiring practices best explain the disruption, while others focused on the difficulty of staying in an expensive city during a prolonged job search.

\[Ada\] On Quora, why were Soviet mathematics textbooks so demanding compared with American textbooks? In an answer from twenty seventeen, Scott Miller proposes that scarce computers and laboratory equipment encouraged analytical problem-solving. He draws on time spent in Protvino in nineteen ninety-two and praises inexpensive, well-written Mir textbooks. His account offers a personal explanation, rather than a comprehensive comparison of education systems. Commenters disagreed over demanding standards versus helping more students succeed, and several cautioned against judging an entire system by its strongest students or books.

\[Ken\] From Fortune, American workers' share of economic output falls to its lowest recorded level since nineteen forty-seven. The report puts labor's share at fifty-two point eight percent in the second quarter, down from fifty-three point six percent in the previous quarter. It examines automation, employment and immigration as possible influences, while noting uncertainty about AI's contribution to productivity. Commenters worried about concentrated wealth and consumer spending, while others stressed that a falling percentage of output does not, by itself, establish a fall in workers' absolute income.

\[Grace\] On the Jefftk blog, bidirectional charging during grid outages. The author examines whether an electric car can power a home and also recharge from solar while the grid remains down. The post finds that several systems separate those functions or depend on specific vehicles and equipment, making installation less straightforward than the basic promise suggests. Commenters compared batteries with longer-term backup options and described difficult installations. They also debated interoperability, battery wear and whether simpler arrangements can meet a household's needs.

\[Linus\] From The Register, Openreach warns that Britain's legacy telephone network will retire on January thirty-first, twenty twenty-seven. This concerns the old switched network, while landline service can continue through digital voice or transitional arrangements. The article highlights care alarms, security systems and payment terminals that may need migration work, and calls for providers to help vulnerable customers. Commenters focused on reliability during power cuts and the maintenance of backup batteries, while others compared the change with telephone migrations already experienced in other countries.

\[Dennis\] That's RobotPaper's daily HN podcast for Tuesday, November third. Today we covered model-weight research, security and maintenance pressures, and the financial questions surrounding Anthropic's planned public offering. The article links and discussion threads are in the show notes. 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.
