---
title: "Oct 28: AI agents bring more work, and more"
show: "RobotPaper's daily HN podcast"
canonical: "https://podcasts.robotpaper.ai/hn/ep/2026-10-28/"
date: "2026-10-28T03:18:49+00:00"
duration: "29:45"
duration_seconds: 1785
audio: "https://podcasts.robotpaper.ai/hn/ep/2026-10-28/2026-10-28-hn.mp3"
chapters: "https://podcasts.robotpaper.ai/hn/ep/2026-10-28/chapters.json"
transcript: "https://podcasts.robotpaper.ai/hn/ep/2026-10-28.txt"
feed: "https://podcasts.robotpaper.ai/hn/feed.xml"
language: "en"
author: "Roy Osherove"
ai_generated_voices: true
---

# Oct 28: AI agents bring more work, and more

RobotPaper's daily HN podcast · Oct 28, 2026 · 30 min

- Listen: [MP3, 29 MB](https://podcasts.robotpaper.ai/hn/ep/2026-10-28/2026-10-28-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/2026-10-28/
- Chapters (JSON): https://podcasts.robotpaper.ai/hn/ep/2026-10-28/chapters.json
- Transcript (text): https://podcasts.robotpaper.ai/hn/ep/2026-10-28.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
- `00:00:30` AI and workload
- `00:02:56` Genie 4
- `00:05:15` Apple M6 Macs
- `00:09:07` Windows compatibility
- `00:11:24` Software quality
- `00:13:45` Brooks law and AI
- `00:16:00` Secret FISA ruling
- `00:18:20` Exponential Time
- `00:20:41` Plate reader scrutiny
- `00:23:13` Roblox model scraping
- `00:25:33` Claude book dispute
- `00:26:13` Ghostty release
- `00:26:53` Python JIT removal
- `00:27:26` Wikipedia AI cleanup
- `00:28:06` Printable websites
- `00:28:41` Amazon AI refocus
- `00:29:20` Outro

## Show notes

- [AI promised less work. I've never worked more](https://oioannou.com/blog/ai-promised-less-work/) · [HN discussion](https://news.ycombinator.com/item?id=50257325) \| quoted: [nkrisc](https://news.ycombinator.com/user?id=nkrisc), [conception](https://news.ycombinator.com/user?id=conception), [ak39](https://news.ycombinator.com/user?id=ak39), [matsemann](https://news.ycombinator.com/user?id=matsemann), [khasan222](https://news.ycombinator.com/user?id=khasan222), [thatguymike](https://news.ycombinator.com/user?id=thatguymike)
- [Genie 4](https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/) · [HN discussion](https://news.ycombinator.com/item?id=50261319) \| quoted: [inkcapmushroom](https://news.ycombinator.com/user?id=inkcapmushroom), [cvoss](https://news.ycombinator.com/user?id=cvoss), [daniel31x13](https://news.ycombinator.com/user?id=daniel31x13), [captainclam](https://news.ycombinator.com/user?id=captainclam), [kouteiheika](https://news.ycombinator.com/user?id=kouteiheika), [dack](https://news.ycombinator.com/user?id=dack), [sosodev](https://news.ycombinator.com/user?id=sosodev), [overfeed](https://news.ycombinator.com/user?id=overfeed)

### Apple M6 Macs (3 stories)

From Apple: a new iMac with M six, and a new fourteen-inch MacBook Pro with M six.

- [New iMac with M6](https://www.apple.com/imac/) · [HN discussion](https://news.ycombinator.com/item?id=50258538) \| quoted: [chrisguilbeau](https://news.ycombinator.com/user?id=chrisguilbeau), [trollbridge](https://news.ycombinator.com/user?id=trollbridge), [driverdan](https://news.ycombinator.com/user?id=driverdan), [porphyra](https://news.ycombinator.com/user?id=porphyra), [littlecranky67](https://news.ycombinator.com/user?id=littlecranky67), [JaggerJo](https://news.ycombinator.com/user?id=JaggerJo), [hbn](https://news.ycombinator.com/user?id=hbn)
- [New 14‑inch MacBook Pro with M6](https://www.apple.com/in/newsroom/2026/10/apple-introduces-the-new-14-inch-macbook-pro-with-m6/) · [HN discussion](https://news.ycombinator.com/item?id=50261352) \| quoted: [at1as](https://news.ycombinator.com/user?id=at1as), [epistasis](https://news.ycombinator.com/user?id=epistasis), [swiftcoder](https://news.ycombinator.com/user?id=swiftcoder), [porphyra](https://news.ycombinator.com/user?id=porphyra), [bigwheels](https://news.ycombinator.com/user?id=bigwheels), [hbn](https://news.ycombinator.com/user?id=hbn)
- [New 14‑inch MacBook Pro with M6](https://www.apple.com/newsroom/2026/10/apple-introduces-the-new-14-inch-macbook-pro-with-m6/) · [HN discussion](https://news.ycombinator.com/item?id=50258521) \| quoted: [makmanalp](https://news.ycombinator.com/user?id=makmanalp), [Hamuko](https://news.ycombinator.com/user?id=Hamuko)
- [Oldest GUI application that still runs on Windows 11](https://blog.enlyth.com/posts/oldest-windows-gui-app-win-11/) · [HN discussion](https://news.ycombinator.com/item?id=50248055) \| quoted: [drnick1](https://news.ycombinator.com/user?id=drnick1), [dpark](https://news.ycombinator.com/user?id=dpark), [lolpanda](https://news.ycombinator.com/user?id=lolpanda), [the\_\_alchemist](https://news.ycombinator.com/user?id=the__alchemist), [jen20](https://news.ycombinator.com/user?id=jen20), [mattbee](https://news.ycombinator.com/user?id=mattbee)
- [Software may be getting worse, but it used to be horrible](https://www.bitlog.com/2026/10/27/software-may-be-getting-worse-but-it-used-to-be-horrible/) · [HN discussion](https://news.ycombinator.com/item?id=50257548) \| quoted: [xmddmx](https://news.ycombinator.com/user?id=xmddmx), [brookst](https://news.ycombinator.com/user?id=brookst), [malfist](https://news.ycombinator.com/user?id=malfist), [quacked](https://news.ycombinator.com/user?id=quacked), [z0r](https://news.ycombinator.com/user?id=z0r), [drnick1](https://news.ycombinator.com/user?id=drnick1)
- [The Mythical Man-Month in the Age of AI](https://j11g.com/the-mythical-man-month-in-the-age-of-ai) · [HN discussion](https://news.ycombinator.com/item?id=50255298) \| quoted: [freefaler](https://news.ycombinator.com/user?id=freefaler), [aleph\_minus\_one](https://news.ycombinator.com/user?id=aleph_minus_one), [rimeice](https://news.ycombinator.com/user?id=rimeice), [Fordec](https://news.ycombinator.com/user?id=Fordec), [matsemann](https://news.ycombinator.com/user?id=matsemann), [breckenedge](https://news.ycombinator.com/user?id=breckenedge), [yread](https://news.ycombinator.com/user?id=yread), [kian](https://news.ycombinator.com/user?id=kian)
- [US Government is hiding a secret court ruling about how it spies on Americans](https://www.wired.com/story/the-us-government-is-hiding-a-secret-court-ruling-about-how-it-spies-on-americans/) · [HN discussion](https://news.ycombinator.com/item?id=50259604) \| quoted: [edot](https://news.ycombinator.com/user?id=edot), [koops](https://news.ycombinator.com/user?id=koops), [jandrewrogers](https://news.ycombinator.com/user?id=jandrewrogers), [Matl](https://news.ycombinator.com/user?id=Matl), [NewJazz](https://news.ycombinator.com/user?id=NewJazz), [tancop](https://news.ycombinator.com/user?id=tancop), [TheCoelacanth](https://news.ycombinator.com/user?id=TheCoelacanth), [Tangurena2](https://news.ycombinator.com/user?id=Tangurena2), [WillPostForFood](https://news.ycombinator.com/user?id=WillPostForFood)
- [The Exponential Time Hypothesis Is False \[pdf\]](https://github.com/openai/math/blob/main/preprints/The-Exponential-Time-Hypothesis-is-False-October-25-2026/paper.pdf) · [HN discussion](https://news.ycombinator.com/item?id=50251661) \| quoted: [cochne](https://news.ycombinator.com/user?id=cochne), [lisper](https://news.ycombinator.com/user?id=lisper), [mudita](https://news.ycombinator.com/user?id=mudita), [strommec](https://news.ycombinator.com/user?id=strommec), [rokob](https://news.ycombinator.com/user?id=rokob), [jibalt](https://news.ycombinator.com/user?id=jibalt), [tomaskafka](https://news.ycombinator.com/user?id=tomaskafka), [brookst](https://news.ycombinator.com/user?id=brookst), [fxwin](https://news.ycombinator.com/user?id=fxwin)

### Plate reader scrutiny (2 stories)

From Gizmodo: police are using AI to pull over drivers for suspicious behavior.

- [Police are using AI to pull over drivers for 'suspicious' behavior](https://gizmodo.com/police-are-using-ai-to-pull-over-drivers-for-suspicious-behavior-2000831220) · [HN discussion](https://news.ycombinator.com/item?id=50261819) \| quoted: [JumpCrisscross](https://news.ycombinator.com/user?id=JumpCrisscross), [taurath](https://news.ycombinator.com/user?id=taurath), [nonethewiser](https://news.ycombinator.com/user?id=nonethewiser), [simoncion](https://news.ycombinator.com/user?id=simoncion)
- [In Leaked Audio, Flock CEO Tells Cops Backlash to the Company Isn't Real](https://www.404media.co/in-leaked-audio-flock-ceo-tells-cops-backlash-to-the-company-isnt-real/) · [HN discussion](https://news.ycombinator.com/item?id=50248005) \| quoted: [alexjplant](https://news.ycombinator.com/user?id=alexjplant), [CGMthrowaway](https://news.ycombinator.com/user?id=CGMthrowaway)
- [Roblox downloaded 145GB models from us (including adult content) without asking](https://open3dlab.com/news/roblox-downloaded-145-gb-of-models-from-our-sites/) · [HN discussion](https://news.ycombinator.com/item?id=50256985) \| quoted: [phn](https://news.ycombinator.com/user?id=phn), [amiga386](https://news.ycombinator.com/user?id=amiga386), [hypfer](https://news.ycombinator.com/user?id=hypfer), [cube00](https://news.ycombinator.com/user?id=cube00), [jeremyjh](https://news.ycombinator.com/user?id=jeremyjh), [offbeatport](https://news.ycombinator.com/user?id=offbeatport), [semiquaver](https://news.ycombinator.com/user?id=semiquaver)
- [Anthropic trained Claude on illegally obtained Stanford professors' books](https://stanforddaily.com/2026/10/27/professors-books-pirated-to-train-claude/) · [HN discussion](https://news.ycombinator.com/item?id=50260814)
- [Ghostty 1.4](https://ghostty.org/docs/install/release-notes/1-4-0) · [HN discussion](https://news.ycombinator.com/item?id=50261998)
- [Python's experimental JIT to be removed](https://lwn.net/SubscriberLink/1102102/0c00e4d550f25f8c/) · [HN discussion](https://news.ycombinator.com/item?id=50257419)
- [Wikipedia's Cleanup of AI-Generated Content](https://spectrum.ieee.org/wikipedia-wikiproject-ai-writing-cleanup) · [HN discussion](https://news.ycombinator.com/item?id=50257326)
- [Websites Should Be Printable](https://brennenputh.me/blog/printable-websites/) · [HN discussion](https://news.ycombinator.com/item?id=50241664)
- [Amazon Is Gutting Its AI Division After Sustained Failure](https://futurism.com/artificial-intelligence/amazon-ai-division-failure-nova) · [HN discussion](https://news.ycombinator.com/item?id=50261589)

### Music

"Sneaky Snitch" Kevin MacLeod (incompetech.com)\
Licensed under Creative Commons: By Attribution 4.0\
http://creativecommons.org/licenses/by/4.0/

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, October twenty-eighth, 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, coding agents bring more work; Apple introduces M six Macs.

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

\[Ada\] First up, from the Oioannou blog: AI promised less work. I've never worked more. An engineering leader describes how coding agents expanded the work they take on and increased the strain of switching between tasks. The author leads engineering at Bold dot org and now uses agents throughout their engineering work. Several can run at once, implementing changes, investigating production questions, reviewing code or doing research. Each eventually needs a decision or a review, requiring them to reload a different problem. The post contrasts this with an earlier routine of handling a few substantial tasks in sequence, where the author could concentrate on a feature or migration before moving to the next problem. Their GitHub history shows activity returning toward the level of their earlier full-time feature development role. By October twenty-seventh, they had recorded two thousand one hundred and two contributions, more than in any full year from twenty twenty-one through twenty twenty-five. The author explicitly calls contributions a rough activity signal, not a measure of value. The pattern is uneven: bursts of experimentation followed by periods of pulling back. Cheaper implementation has made previously deferred bugs, improvements and projects feel worth attempting, while review and attention remain their responsibility.

\[Linus\] And the thread is divided between enthusiasm for the tools and concern about the workload they create. nkrisc argues that employers will generally expect more output rather than pay the same salary for less work. conception pushes back on that framing, saying collective action could also turn productivity gains into shorter working time. The practical discussion focuses on boundaries. ak39 describes setting a finite daily task list and moving planning away from the keyboard when new ideas threaten to expand it. matsemann adds that apparently useful side projects also create reviews and discussions for colleagues. khasan222 welcomes spending less time on routine debugging and more on broader engineering problems. thatguymike agrees that the work can be more engaging, but says routine tasks also offered breathing room: the remaining work demands sustained judgment, interaction and context switching.

\[Ken\] Next up, from Google DeepMind's blog: Genie four, a new generation of world simulation. Google's experimental world model generates interactive environments with native sound and dialogue, while Gemini models help steer events. Users navigate with keyboard controls and alter scenes with text prompts, changing weather, camera positions or objects as the simulation runs. Google says Gemini Omni helps ground appearance and behavior. Gemini three point eight Flash observes generated frames, breaks goals into steps and adjusts its plan as the scene changes. The announcement also describes experiments linking Genie with Gemini four Argon and robotics research. Generated characters can speak and react to user input, while ambient audio changes with the setting of the simulation. Trusted testers are trying film previsualization, explorable game prototypes, historical scenes and scenario generation. Google is rolling out a research preview to trusted testers before planned access for consumers and developers, beginning with Google AI Ultra subscribers in Project Genie and early API and AI Studio partners. The company describes the model as early and experimental.

\[Grace\] Over in the comments, the demonstrations draw admiration alongside detailed challenges to their consistency and usefulness. inkcapmushroom points to objects changing after leaving the frame, including a rider gaining a helmet, and questions whether the examples show a strong world simulation. cvoss argues that preserving objects when they disappear from view remains a central technical problem. sosodev questions whether visually convincing scenes are reliable enough for robot training. overfeed replies that usefulness depends on the task and the level of detail required. Another discussion concerns creative control. daniel31x13 is excited about the possible convergence of films and games. captainclam counters that a film's deliberate artistic choices are part of its value and that more interactive possibilities do not automatically improve it. kouteiheika is frustrated by limited access and the absence of technical detail. dack summarizes that concern as wanting both an explanation of how the model works and a chance to use it.

\[Dennis\] From Apple: a new iMac with M six, and a new fourteen-inch MacBook Pro with M six. Apple is refreshing its all-in-one desktop and its base-chip professional laptop, with orders opening today and availability beginning November sixth. Three stories cover these announcements, including two submissions of Apple's MacBook Pro release. The iMac product page describes a twenty-four-inch, four point five K Retina display, up to thirty-two gigabytes of unified memory and a dual sixteen-core Neural Engine. It offers a nano-texture glass option, up to four Thunderbolt four ports and support for up to two external six K displays. Apple's MacBook Pro newsroom release gives the laptop's M six chip a twelve-core CPU and a twelve-core GPU, each with two more cores than the previous generation. Apple claims up to thirty percent faster graphics and up to twenty-five percent faster on-device AI workloads than M five. The laptop starts with sixteen gigabytes of unified memory and can be configured with thirty-two. Stated memory bandwidth is up to one hundred and seventy gigabytes per second. For code compilation, Apple reports Xcode project builds running up to one point three times as fast as on M five, and up to two point five times as fast as on the thirteen-inch M one MacBook Pro. The release also lists Wi-Fi seven, Bluetooth six and battery life of up to twenty-four hours. Its display retains a nano-texture option, and the machine includes a twelve-megapixel Center Stage camera. The additional MacBook Pro submission repeats the same launch announcement, including Apple's comparisons with M one systems and its emphasis on local AI processing. These performance and battery figures are Apple's stated results.

\[Ada\] Across the Apple threads, commenters generally respect the hardware but disagree about the value of upgrading and the limits of each design. In the iMac discussion, chrisguilbeau says older machines remain physically useful after software support ends and favors separating the computer from the display. trollbridge replies that older Intel iMacs can still serve useful roles with other operating systems. driverdan argues that a Mac mini plus separate peripherals offers better value. littlecranky67 defends the iMac's integrated camera, speakers and simple desk setup, while JaggerJo wants a larger display and the ability to reuse the computer as an external monitor. In the main MacBook Pro thread, at1as plans to keep an M one Pro longer. epistasis likewise finds an older MacBook Air useful despite some slowdown when running virtual machines. swiftcoder finds the overlap between newer base chips and older Pro or Max chips confusing. porphyra notes that different models can perform better on different workloads, rather than one being the clear best choice. bigwheels questions how the new model's memory bandwidth compares with a Max chip for local language models. hbn pushes back on comparing different chip tiers, asking instead for the previous base model's bandwidth. In the additional MacBook Pro thread, makmanalp describes repeatedly postponing an upgrade while waiting for the next product cycle, even though storage is becoming tight. Weystrom likes the hardware but prefers Linux and KDE to macOS. Gniv welcomes a stated nineteen-hundred-and-ninety-nine-dollar price, comparing it with an earlier M one Pro purchase. Hamuko replies that the new machine uses the base M six chip, rather than an M six Pro. Together, those comments keep the distinctions between chip generation, chip tier and the operating system central to the buying discussion.

\[Linus\] Next up, from the Enlyth blog: Oldest GUI application that still runs on Windows eleven. An investigation found a CD player from a nineteen ninety-two Windows NT build that still launches on the current operating system. The search initially turned up FreeCell dated February twenty-second, nineteen ninety-three, running without changes. Earlier Windows NT programs proved more difficult. Some used an incompatible draft executable format, while others expected a different calling convention for Windows API functions. The CD player, linked on November sixteenth, nineteen ninety-two, avoided another obstacle by using a system runtime library that Windows eleven still includes. It launched, but the author's computer had no CD drive. Playing audio required an additional library that emulated a drive with fake audio data. The post also demonstrates Reversi from Windows one point oh one, dated nineteen eighty-five. That example uses otvdm to run sixteen-bit software, so it falls outside the original search for unmodified applications running without emulation.

\[Ken\] The comments are nostalgic about older applications, with disagreement over whether modern interfaces have improved. drnick1 argues that desktop design peaked around Windows ninety-five and criticizes disappearing menu bars. dpark responds that many applications still have menus and says browsers remove them to conserve screen space. Another discussion concerns whether developers should still choose the Win32 API. lolpanda describes building a responsive image viewer with it, while the\_\_alchemist favors cross-platform toolkits to avoid separate implementations. jen20 says native behavior matters more than appearance, citing keyboard shortcuts and accessibility integration. mattbee offers a practical limit to the compatibility story. A Windows three point one children's game launched on Windows eleven, but restoring it required fixes for palette behavior, a disk-space calculation and threading bugs. In that account, keeping the executable runnable was only part of making the application work correctly.

\[Grace\] From Bitlog: Software may be getting worse, but it used to be horrible. The author challenges nostalgia for old software by recalling routine crashes, lost documents and weak security on earlier Windows computers. Buggy device drivers could destabilize the entire machine, while resource-management problems made reliability deteriorate over time. Applications often lacked autosave, and some that had it could still corrupt files. The author describes joining the Google Docs team in twenty ten and measuring how often people pressed Control S, even though the service saved automatically. It ranked among the most common non-editing actions. The post also recalls advertising malware and password-protected screensavers that could be dismissed through Task Manager. It acknowledges one constraint on older releases: software distributed on physical media needed to function when it reached customers. But that shipping requirement did not eliminate the failures the author remembers.

\[Dennis\] The thread is divided over what software quality should measure, with reliability, responsiveness and user control producing different answers. xmddmx says the article overlooks latency, describing everyday delays after typing or clicking in both websites and installed applications. quacked agrees that reliability has improved but argues that interface design and controllability have declined. Their examples include disappearing customization options, inconsistent menus and animation delays. brookst challenges the article's opening assertion that software is degrading, asking for measurements and evidence of causation. They report writing more tests and improving accessibility with newer tools. malfist replies that comparing lines of test code with application code is a poor quality measure. Another commenter, z0r, focuses on ownership: older software could be installed locally and kept, while newer products can depend on logins or remote servers and change without warning. drnick1 responds that high-quality open-source software is more plentiful now, and z0r agrees that this is an exception to their criticism.

\[Ada\] Moving on, from J11g: The Mythical Man-Month in the Age of AI. The post argues that adding coding agents preserves the coordination problems described by Brooks's Law, even when some human overhead disappears. Frederick Brooks formulated the law in his nineteen seventy-five book, drawing on his experience managing IBM's OS three sixty project. Its central observation is that adding people to a late software project can make it later. The author asks how that principle changes when the additional workers are AI agents. Sequential work still has to happen in sequence, and someone must define completion, review generated code, identify hallucinations and integrate the results. Without clear orchestration, agents working in the same codebase may overwrite one another's changes. The proposed update to the law is that adding agents to a project with unclear requirements produces the wrong answer faster.

\[Linus\] The discussion splits between people who see faster experimentation as an advantage and those who see review and coordination becoming bottlenecks. rimeice argues that generating a wrong answer quickly can help uncover the right requirements. Fordec expands on that idea, contrasting an empty page with a cheap prototype whose flaws make the missing requirements visible. matsemann counters that speed can obscure whether a feature solves the intended problem, especially when nobody has internalized the context. freefaler reports a different divide within their team: stronger programmers produce good solutions faster, while weaker ones submit work without checking architectural requirements. aleph\_minus\_one pushes back on equating AI adoption with ability, saying skilled programmers vary widely in their preferences, particularly when using obscure languages. In the parallel-agent discussion, yread describes agents interfering with one another and performance tests becoming noisy. breckenedge suggests separate worktrees and minimal local testing, while kian likewise recommends plans and worktrees to coordinate concurrent work.

\[Ken\] From Wired: The US government is hiding a secret court ruling about how it spies on Americans. Senators Ron Wyden and Adam Schiff say the administration has missed a deadline to release a surveillance ruling while seeking renewal of Section seven oh two. According to the report, the Foreign Intelligence Surveillance Court objected in March to filters that exposed Americans' intercepted messages without required approvals, written justifications or records. It nevertheless approved another year of collection. Section seven oh two lapsed in June, but surveillance continues under that court approval. Wyden and Schiff say the administration's published one-hundred-and-eighty-day review window expired on September thirteenth. They are seeking an explanation and the names of responsible officials by November sixth. Civil-liberties groups have also urged Congress to withhold reauthorization while the ruling remains secret. Wired reports that a rebuilt FBI participants filter had resumed operating after earlier searches went uncounted. The current court authorization expires in March, while intelligence officials want lawmakers to renew the program before the end of this year.

\[Grace\] The comments are strongly critical of surveillance secrecy, with a separate dispute over how to assign political responsibility. edot argues that support for spying on Americans crosses party lines. koops responds that Wyden, who is raising the objection, is a Democrat. jandrewrogers replies that Wyden's position does not erase other Democrats' support for expanding intelligence powers. Matl asks how a classified interpretation of law can be justified in a democracy. NewJazz argues that secret interpretations follow from having secret courts and warrants, while tancop rejects secrecy around legal interpretations and argues for narrow, time-limited restrictions on genuinely sensitive information. TheCoelacanth calls for abolishing secret courts. In another exchange, Tangurena2 invokes the Sixth Amendment's protections against secret proceedings. WillPostForFood responds that the amendment concerns criminal prosecutions and that the surveillance court is not a criminal court. Much of the debate therefore turns on both public accountability and which legal safeguards apply.

\[Dennis\] Next up, from OpenAI's GitHub: a preprint titled The Exponential Time Hypothesis Is False. The paper, dated seventeenth of October and revised on the twenty-fifth, gives a deterministic algorithm that counts the satisfying assignments of any formula in k-CNF form, for every fixed k of at least two. It runs in subexponential time and polynomial space, roughly two to the n divided by a power of the logarithm of n slightly above one half. The Exponential Time Hypothesis, known as ETH, says that deciding three-SAT on n variables requires exponential time in the worst case. The authors say their algorithm refutes it. They also state a limit: because reductions between NP problems can inflate the number of variables, the result does not imply that general NP problems have subexponential algorithms. They do derive subexponential counting algorithms for independent sets, proper graph colorings and perfect matchings, arbitrarily accurate approximations for Max-k-SAT, and a circuit lower bound for the class E to the NP.

\[Ada\] The thread opens with amazement and a caveat. cochne calls the result amazing and unexpected, but quotes the paper's limit and explains that a subexponential solution to three-SAT does not carry over to every NP problem once conversion inflates the input. lisper asks whether this is nearly as big as P equals NP. mudita answers by citing a comment on Scott Aaronson's blog, in which he raises his personal probability of P equals NP from three percent to five percent. strommec asks whether a result about ETH could imply faster self-attention for language models, since a 2022 paper ties quadratic attention to the Strong Exponential Time Hypothesis. rokob and jibalt reply that the implication runs only one way, so refuting it implies nothing about attention. Another line of discussion starts with tomaskafka and fxwin, who note that results across mathematics came out with nothing touching cryptography, and suspect a veto on publication. brookst objects that this treats absence of evidence as proof.

\[Linus\] From Gizmodo: police are using AI to pull over drivers for suspicious behavior. License plate tracking is being used to flag travel patterns even when drivers are not suspected of a specific crime. Two reports cover these surveillance systems and the backlash against them. Gizmodo recounts an ACLU account from Grace Knowles, who was stopped during a road trip. The officer initially cited left-lane driving, then said her plate had been flagged for suspicious travel patterns. Knowles said he showed her a map with locations and photographs of her plate. The ACLU warns that these encounters can be frightening and dangerous. Plate readers collect vehicle information alongside times and locations, allowing police to search movements across a camera network. Gizmodo also reports that a federal jury in Texas sided with a man challenging a traffic stop triggered by an overnight trip between Houston and Carrizo Springs. From Four Oh Four Media: in leaked audio, Flock's CEO tells police that backlash to the company is not real. Garrett Langley told police chiefs at a private conference that Russian bots drove much of the online opposition. The report says he supplied no evidence and repeated a disproven theory connecting reduced bot activity with Ukrainian strikes on Russian data centers.

\[Ken\] The comments are strongly critical of surveillance, with disagreement over how to describe the searches and the opposition. In Gizmodo's thread, JumpCrisscross emphasizes that the system tracks movements across locations, rather than detecting erratic driving. taurath calls the practice a dragnet. nonethewiser pushes back that observing plates differs from indiscriminate stops and searches, and describes the reported stop as targeted. simoncion replies that indiscriminate collection of information can itself constitute a dragnet, even without stopping everyone. These commenters disagree over the meaning of the term as well as the surveillance itself. In the Four Oh Four Media thread, alexjplant objects to Langley's remarks about making mistakes while refusing to apologize. CGMthrowaway argues that genuine opponents and organized bot activity can coexist. Slowin rejects that explanation of the backlash, saying the people they know oppose Flock. That exchange centers on whether Langley's characterization reflects the opposition people see around them.

\[Grace\] Next up, from Open3DLab: Roblox downloaded a hundred and forty-five gigabytes of models, including adult content, without asking. The model-sharing site says a crawler identifying itself with a Roblox research address copied creators' files despite its scraping restrictions. Open3DLab reports that the crawler accessed almost its entire catalogue across Open3DLab, SFMLab and SmutBase, downloading a thousand and twenty-seven files from seven hundred and twelve projects. Nearly half the downloaded projects contained adult content. The site says the crawler retrieved but ignored its robots dot text file, disregarded terms forbidding scraping, and collected projects carrying noncommercial licenses. After administrators blocked it, further requests continued but failed. The site says Roblox did not answer its requests for comment. Open3DLab suspects collection for three-dimensional model training or evaluation, citing the crawler's identifier and a researcher associated with Roblox's Cube work. It explicitly says it cannot prove that purpose or whether any downloaded files entered a model.

\[Dennis\] The thread is divided between criticism of bulk scraping and challenges to the site's copyright position. phn points to assets extracted from commercial games and questions the site's moral argument. amiga386 responds by quoting the site's acknowledgement that fan communities inhabit copyright gray areas, while arguing that industrial copying by large companies creates an unequal system. offbeatport emphasizes the site's restrictions and creators' licenses. semiquaver counters that downloading alone does not demonstrate the feared downstream use, and asks for evidence of a license violation. A practical discussion follows. hypfer suggests blocking traffic from major cloud providers to reduce bots. cube00 warns that this can also exclude legitimate visitors using corporate or personal VPN gateways. jeremyjh adds that scrapers already use residential proxies. The debate leaves the alleged collection and the unproven training purpose as distinct questions.

\[Ada\] A few more that didn't make the cut today. From The Stanford Daily: Anthropic trained Claude on illegally obtained Stanford professors' books. The report describes a copyright dispute separating pirated acquisition from training on legally acquired books. It says Judge William Alsup barred piracy to build a permanent central library while treating training on legally acquired books as fair use, and reports a one-point-five-billion-dollar settlement. Commenters are angry about compensation and unequal treatment, while debating the distinction between obtaining books and using them for training.

\[Linus\] From Ghostty: version one point four. The terminal release brings performance and memory improvements, expanded graphics support and security fixes. Its release notes describe a complete implementation of Kitty graphics, including animation and relative placement, alongside improved SSH integration. Fixes address paste protection and malicious terminal output that could cause crashes or, in limited cases, access files outside permitted temporary directories. Commenters welcome the improvements, while others question switching from established terminals and point to features or usability details they still miss.

\[Ken\] From LWN: Python's experimental just-in-time compiler is to be removed. Maintainers have withdrawn the proposal intended to establish a supported path for the compiler. The steering council had paused work pending a proposal addressing maintenance, compatibility, success metrics and a completion timeline. The experiment entered CPython with version three point thirteen in twenty twenty-four. Commenters express disappointment at the apparent end of development, while disagreeing over governance requirements and discussing other approaches to speeding up Python.

\[Grace\] From IEEE Spectrum: Wikipedia's cleanup of AI-generated content. Volunteers are adapting their methods as machine-written articles become harder to identify. WikiProject AI Cleanup began in twenty twenty-three to remove generated material that violates Wikipedia's guidelines. Founding member Ellia Lebleu describes a shift from conspicuous stock phrases toward subtler structural clues, with volunteers developing detection tools. Commenters broadly appreciate the cleanup, while debating readability, the reliability of moderation and whether future AI systems could produce verifiable encyclopedic material.

\[Dennis\] Over on the Brennenputh blog: websites should be printable. A post argues for making web writing usable on paper as well as on a screen. The author describes preferring printed documents for annotation, then finding oversized text and an awkward header when testing the site's own print output. Commenters share frustration with browser printing and documents that must work in both HTML and PDF. Others question changing fonts for print, or whether printing deserves the same priority as screen navigation.

\[Linus\] From Futurism: Amazon is gutting its AI division after sustained failure. The report describes reduced ambitions for Nova models and a narrower research focus. It says text, image and video models are receiving minimal maintenance resources, while staff and computing capacity shift toward a frontier-model effort led by Pieter Abbeel. Commenters criticize Amazon's research culture and debate the economics of developing models versus hosting other companies' systems. The discussion also includes objections that the report is several months old.

\[Ada\] That's RobotPaper's daily HN podcast for Wednesday, October twenty-eighth. Thanks for listening. 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.
