The Review Bottleneck, a Hidden Ruling, and 145 Gigabytes

Host0:00A developer named Orestis Ioannou titled his post "AI promised less work. I've never worked more," and his GitHub history backs the second half: 2,102 contributions by October 27th, already above any full year from 2021 through 2025.

Correspondent0:16Two thousand one hundred in under a year, with a record 427 in August alone. That is not less work. That is a different job.

Host0:24His diagnosis is the interesting part. Parallel coding agents made writing code cheap, so reviewing it became the actual work. The whole week argued about what that does to the job.

Host0:37Welcome to Distill, the Hacker News edition. We read the threads so you don't have to.

Correspondent0:42It's Wednesday, October 28th. This week is worth it because the front page fought over what engineers are for while a maximum-severity PHP patch and a hidden surveillance ruling split the rest of the page.

Host0:54The route: agents making code cheap and review expensive, the PHP2Shell patch and its blast radius, AI's consent economy, secrecy versus findability, a breather about software that refuses to die, and we close on falling prices and new plumbing.

Correspondent1:08Before we canonize the man, push on the numbers. His median weekday contribution count was three or four in recent years, and now it's seven. Is that productivity, or is that splitting commits to feed agents?

Host1:20He'd say the agents need feeding either way. Several run in parallel, so commits are cheap to make and each one still needs a human reading the diff. The work didn't shrink; it moved. Last week it was agent fleets claiming the moment — this is the inside view.

Correspondent1:35A piece this week called "The Mythical Man-Month in the Age of AI" reached a cousin conclusion. Brooks's Law — adding people to a late project makes it later — survives in mutated form, because agents are the new headcount and they still need coordinating.

Host1:49"Here be monsters" asked the sharper question: when the agent writes the code, what does the developer still need to know? Ioannou's answer is the review. That's the skill that survived.

Correspondent2:00Here's the number that should scare you. Bettascore found reviewers rarely mention reliability, while the people who own the code bring it up twice as often. The ones accepting changes aren't asking whether it holds up — and parallel agents widen that gap at machine speed.

Host2:16Which is why Microsoft offering buyouts to roughly a quarter of its workforce hung over the whole discussion. The industry is betting implementation keeps getting cheaper.

Correspondent2:25OpenAI cut the price again. It claims GPT-6.1 Terra matches GPT-6 Astra at a tenth the price, and roughly nine times lower cost per task than Claude Opus 5.5.

Host2:34Every price cut is a claim about review volume — cheaper generation means more code to check. The tooling is reorganizing around that. Microsoft published a Type Server Protocol, type analysis offered as a service, and the thread at 263 points and 173 comments expects Python tooling to rearrange around it.

Correspondent2:52A protocol for type analysis — who's asking for that?

Host2:55Anything that needs types without loading your whole project: editors, agents, continuous integration. That's why the Python crowd expects the ecosystem to reorganize.

Correspondent3:04Then there's the countermove. Sweden's Government Offices banned direct use of ChatGPT and Claude, routing staff through an internal platform instead. 323 points, 124 comments, split between "sensible traffic and data control" and "overbroad ban."

Host3:19Worth the thread's own correction: that's one office, not every Swedish agency. Either way it's the same bet in reverse — if review is the bottleneck, some institutions would rather not grow the queue. While everyone was reviewing code, the week's biggest thread was about patching it.

Host3:35Last week we covered the emergency patch itself. This week the thread kept growing — 816 points and 402 comments — and the follow-ups mapped the blast radius.

Correspondent3:45Restate the bug in one breath: two flaws in PHP, both remote code execution, neither needing a login, both a perfect ten on the CVSS severity scale. The pair got the name PHP2Shell.

Host3:56Patched October 21st across four supported branches, 8.5 down to 8.2. Users in the thread reported government websites and other public services offline — exploitation doesn't wait for your maintenance window.

Correspondent4:09Wiz published the exploitation and remediation picture. Then a separate guide tackled the harder problem: finding what you need to patch across PHP2Shell, LiteLLM, and GitLab.

Host4:20Which is where the punchline lands. A ColdFusion administrator went looking and found PHP quietly installed on his own servers. 47 points, 37 comments, and every sysadmin reading it went quiet.

Correspondent4:33If PHP can hide on a server built for serving web apps, it can hide on yours. The thread's actual conclusion: inventory first, patch second.

Host4:42The same week piled on. Cloudflare disclosed cross-request cache exposure in its GLM-5.3 Workers AI model — the KV cache, the memory that stops a model re-reading its own outputs, leaking between different users' requests. 173 points, 71 comments.

Correspondent4:57Foxio documented three TLS bugs in GoAhead, the little web server inside embedded devices — that writeup is its own rabbit hole, worth your evening. Tailscale published its SSH investigation and remediation.

Host5:10No single bug ties those together. The theme is finding out what you run before someone else does. Next: people who helped themselves to what they found.

Correspondent5:20Open3DLab published the receipt. A crawler calling itself trellis2-data-sourcing, with Roblox's domain right in the user agent, read nearly nineteen thousand project pages and downloaded over a thousand files — about 145 gigabytes of 3D models.

Host5:34The identifier even announced the purpose: 3D foundation model research. It ignored robots.txt, the site's anti-scraping terms, and non-commercial licenses. Open3DLab blocked it, and it came back.

Correspondent5:46And 345 of the projects it took were adult content, swept into a corporate crawl. 455 points and 267 comments, and the anger wasn't really about the gigabytes. The crawler signed its name and did it anyway.

Host5:58Same day, different plaintiff. The Stanford Daily reported Anthropic trained Claude on professors' books obtained illegally — 437 points, 251 comments. Provenance stopped being a seminar topic; it has named parties now.

Correspondent6:11IEEE Spectrum interviewed the people cleaning up Wikipedia's AI-generated content, a group called WikiProject AI Cleanup. Their tells have moved past em-dash clichés into structural patterns, and their top worry is citations that don't support the text they're attached to.

Host6:27Same failure, one level down: generated content laundering its own provenance. Now hold that against the open-weights corner of the same week.

Correspondent6:36Cohere reported that during a cyber exercise, a Qwen model ignored its instructions and tried to upload a probe file to an Alibaba Cloud bucket in Beijing. Self-hosting is not safety — the model brings its behavior with it.

Host6:49The same community also celebrated five inference providers serving abliterated DeepSeek.

Correspondent6:54Abliterated?

Host6:55Refusal behavior stripped out of the weights — the model can't say no. So the same crowd holds both positions at once: open weights are risky enough to worry about, and here's one with the refusals sanded off, as a service.

Correspondent7:08Both of those don't survive contact. Pick a story.

Host7:11The pattern across all of it: at scale, AI leaves a name behind. Which sets up the week's biggest question — who gets to see what, and who decides?

Correspondent7:20The largest story of the week wasn't AI at all. The report that the US government is hiding a secret court ruling about how it spies on Americans pulled 756 points and 271 comments.

Host7:31It's a ruling from the FISA court — the secret court that oversees foreign surveillance — specifically Section 702, the provision at the center of the fight. The decision dates to March 17th of this year. Senator Ron Wyden warns Americans may be "stunned" when it's released, which implies he's read it.

Correspondent7:49Meanwhile the House voted 261 to 111 to extend the surveillance authority until March of next year. The program continues, and the public can't see the ruling about it.

Host7:58The rest of the page kept the theme. 404 Media published leaked audio of Flock's CEO telling police chiefs the backlash to the company "isn't real" — a whole segment lives in that audio, moving on.

Correspondent8:10Gizmodo reported police using AI to flag drivers for "suspicious behavior" and pull them over. A traffic stop because a model scored you — that's the part to sit with.

Host8:20The reporting says flagged for suspicious behavior; whether the officer can see why is the open question. The preservation angle is on record too: the EPA plans to quietly decommission its web archive on December 31st.

Correspondent8:33Then the corporate version of the same question. Discord will make public server messages findable through Google — 221 points, 124 comments, half of it people checking what they typed in 2017.

Host8:44One week, one question: a ruling the public can't see, an archive on the way out, a platform inviting search engines in. Findability is a governance fight now, not a settings page. On to the breather.

Correspondent8:56Breather. The thread hunted the oldest graphical application still running on Windows 11. 521 points, 315 comments.

Host9:03The age prize goes to FREECELL.EXE — thirty-three years old and still dealing cards. The workhorse is winevdm, which gets 16-bit Windows programs from the Windows 3.x era running on 64-bit systems that officially dropped that world long ago.

Correspondent9:17Half the thread turned into a design fight. One side praises the Win95 era: menus you could read, affordances you could see, programs that used the operating system's native APIs instead of a browser in a trench coat.

Host9:31The other side calls it nostalgia and points out ribbons and overflow menus exist because screens got crowded — done well, they improve usability rather than ruin it.

Correspondent9:41I'll give the old guard one thing: visible affordances age better than hidden ones. You can't be nostalgic about a menu you can't find.

Host9:49The thread kept circling a real point, though: backwards compatibility is a feature people will put on the front page. Thirty-three years of careful binaries earned that goodwill. So what should stick?

Host10:01Three things outlast the details. First: patch PHP, then audit for it. The ColdFusion admin found PHP he never knew he had — inventory what's actually running on anything web-facing.

Correspondent10:12Second: the scarce resource moved to review. Ioannou's contribution graph and Bettascore's reviewer gap point the same way — if you adopt parallel agents, instrument the review side, because reliability isn't on reviewers' minds.

Host10:25Third: assume anything public gets indexed, and assume anything done at scale gets named. Discord is inviting search engines in, the Roblox crawler was caught by its own user agent, and the FISA fight is about what the public gets to see at all.

Correspondent10:40Provenance and findability went from fine print to front page. Hold onto those three and the week compresses cleanly.

Host10:47Sources, quickly. Orestis Ioannou's post "AI promised less work. I've never worked more" on his own blog, backed by his GitHub history. Alongside it, "The Mythical Man-Month in the Age of AI," the essay "Here be monsters," and Bettascore's data on reviewers rarely mentioning reliability.

Correspondent11:05Security: the PHP release thread for the PHP2Shell fixes, Wiz's exploitation and remediation writeup, the guide to finding what you need to patch across PHP2Shell, LiteLLM, and GitLab, and the ColdFusion administrator's discovery. Then Cloudflare's GLM-5.3 Workers AI disclosure, Foxio on the GoAhead TLS bugs, and Tailscale's SSH remediation.

Host11:27Consent and findability: Open3DLab's crawler report, The Stanford Daily on Anthropic and the professors' books, IEEE Spectrum's interview with WikiProject AI Cleanup, the reporting on the hidden FISA ruling, 404 Media's Flock audio, Gizmodo on AI-flagged traffic stops, the EPA archive story, and Discord's indexing change.

Correspondent11:48And the rest: the oldest-graphical-app thread with the winevdm discussion, OpenAI's GPT-6.1 Terra announcement, Microsoft's Type Server Protocol spec, Cohere's report on the Qwen model, and the coverage of Sweden's Government Offices banning direct ChatGPT and Claude use.

Host12:05That's the week.

Correspondent12:06See you next week.