0818 | GitHub Incident, Wiz Red Agent Hits Snowflake, Amazon Book Scanning, Apple Antitrust

Mia: Welcome to HackerNews Daily on Bri Radio. I'm Mia, and joining me is Milo. We've got a packed show today.

Milo: Hey everyone. We're covering an incident affecting GitHub.com, plus a major breakthrough by an AI-powered security tool that found and exploited a critical GitHub Actions vulnerability.

Mia: We'll also dig into Amazon quietly buying up rare books at scale, a big antitrust decision on Apple's personalized ads in Germany, and a court ruling that could return a massive trove of data to a St. Louis public broadcaster.

Milo: And with a new OpenAI model available at a steep discount, plus a guide on dodging intrusive AI, we've got plenty to get to. Let's jump in.

Mia: GitHub hit a reported capacity crunch over the past couple of days, with users landing on an outright error page saying no server was currently available to service their request. At first there was no incident listed on GitHub's status page at all — that notice only appeared after people started posting about it, and one developer who hit the wall titled their post “GitHub Is Overloaded.”

Milo: And that spike comes at a moment when people are already fed up. Another developer pointed out that GitHub has been down consistently over the last few months and openly asked whether it's time to switch to alternatives. That thread drew recommendations for lighter-weight options — one that was pointed to is a single-binary CI tool written in Go, with a built-in container runner and pipelines written in ordinary programming languages instead of YAML.

Mia: So the reliability complaints and the overloaded message are feeding the same conversation. The real question hanging over all of it: if GitHub's struggles keep recurring, do developers start treating one of the most central services in the industry as something they need a fallback for?

Mia: Wiz Research's autonomous, AI-powered security tool called Red Agent discovered and exploited a critical GitHub Actions workflow vulnerability in snowflakedb/snowflake-connector-net, a public Snowflake repository, doing so through Snowflake's HackerOne bug bounty program. The flaw let an unauthenticated user run arbitrary commands inside a GitHub Actions runner simply by opening a GitHub issue with a specially crafted title. It was introduced on June 18, just five days before discovery, by commit 4a1b8ce in PR 1218, co-authored by GitHub Copilot Autofix — the AI fix removed the repo's existing safe pattern and replaced it with direct interpolation of the issue title into a shell script. When the agent's first exfiltration attempt hit a bash syntax error, it adjusted its payload, closed the shell block, and received the callback with credentials that authenticated as qa@snowflake.net and granted read access across Snowflake's engineering systems.

Milo: Let me walk through what happened, because it's genuinely unsettling. The flaw sat in Snowflake's.NET connector repository, in a GitHub Actions workflow. An unauthenticated user could trigger arbitrary command execution on a GitHub runner just by opening an issue with a carefully crafted title.

Mia: And here's the part that should really give people pause — the vulnerability was introduced just five days before it was found, by a commit co-authored by GitHub Copilot's Autofix. That AI-driven fix removed an existing safe pattern, which passed the issue title through a secure environment variable and built the payload with a JSON tool. It replaced that with direct interpolation of the issue title straight into a shell script.

Milo: Even the guardrail put in place didn't help. There was a protective-looking condition checking for a pull request, but on issue events that field is always empty — so every GitHub user sailed right through the gate. And when Red Agent's first attempt to grab credentials hit a syntax error, the agent didn't fail — it read the error, adjusted its payload to properly close the shell block, and got the callback with base64-encoded credentials from an Azure server.

Mia: Those stolen credentials turned out to authenticate to Snowflake's Atlassian workspace, granting read access across the engineering department. The takeaway here is uncomfortable: an autonomous AI tool found a real vulnerability, exploited it end-to-end, and the vulnerable code was itself generated by an AI autofix. That's a feedback loop the security world is going to be chewing on for a while.

Milo: TechCrunch reports, citing 404 Media, that Amazon is buying up large quantities of rare books, cutting off their spines, and scanning them for AI training. A tracking device 404 Media placed in one rare book ultimately arrived at an Amazon facility in Las Vegas called VGT3, which Amazon told 404 Media purchases books through commercial channels to improve the products customers use. The point of buying rare and out-of-print books is that they hard this are to find online and predate AI-generated text, which matters because training on AI-generated text carries a risk of model collapse — though commenters disputed the legal framing, with some noting rare books aren't necessarily under copyright and that destroying the original doesn't change the copyright status of the scanned contents.

Mia: The reporting even tracked a physical device placed inside one of those rare books, and it eventually arrived at an Amazon facility in Las Vegas that identifies itself with a symbol of a dinosaur holding a book. Amazon's response was that it purchases books through commercial channels to improve the products and services customers use.

Milo: So why rare books? The argument is that out-of-print material that's hard to find online is especially valuable because it predates AI-generated text — and training on AI output carries the risk of something called model collapse, where the model degrades by learning from generated content. The piece also notes a cited case where Anthropic allegedly pirated books.

Mia: The commenters really split on the legal questions here. One pointed out that rare books are far from necessarily under copyright, and that destroying the original doesn't change the copyright status of what's scanned — while another pushed back on the framing entirely, arguing Amazon has always been profit-driven and that destroying the originals exploits a loophole whose payoff outweighs the outrage. Whatever the exact legal reading, the image of cutting up physical artifacts to feed a model is one that hits differently for people who care about the material record.

Mia: Germany's competition regulator, the Bundeskartellamt, has concluded its proceeding against Apple over personalized advertising in apps, with Apple committing to change its rules. The authority objected to how Apple designed different consent requests for its own offerings and for third-party apps: Apple's App Tracking Transparency framework requires third-party apps to get extra consent through an Apple-defined prompt for certain cross-company data use, while that requirement doesn't apply to Apple's own offerings. Apple maintains the framework is compliant with competition law, but offered commitments the regulator has now declared binding, concluding the case. Some Hacker News commenters noted the regulator only required first-party and third-party apps to be treated equally, and argued Apple eased the burden on third-party publishers instead of raising its own.

Milo: The core of the objection was an asymmetry. Apple's App Tracking Transparency framework requires third-party apps to get an extra consent prompt, predefined by Apple, before certain cross-company data use — but those same rules don't apply to Apple's own offerings, which use data from its own ecosystem and ask for consent through its own prompt.

Mia: Apple maintained that its framework complies with competition law, but it offered commitments that the regulator has now made binding, effectively closing the proceeding. The regulator's president said the key issue is that personal data and privacy get protected effectively.

Milo: Still, at least one commenter in the discussion argued the result didn't actually raise the privacy floor — the thinking being that the regulator only demanded first-party and third-party apps be treated equally, without prescribing how, and Apple responded by easing the burden on third-party publishers rather than tightening its own rules. The takeaway: the case closed, but whether it meaningfully changed how much user tracking happens is very much its own question.

Mia: A Denver District Court judge has set the framework for St. Louis public broadcaster Nine PBS to retrieve roughly fifty terabytes of archival material from Iron Mountain Data Centers. The station sued Iron Mountain in July after the company refused to return the materials because OSS, the now-defunct company Nine PBS had contracted with for storage, technically owned the physical systems housing the data. Judge Eric Elliff ordered Iron Mountain to cooperate and directed Nine PBS to bring in a third-party vendor to access the data within 30 days, with Nine PBS paying current and past-due storage fees. Iron Mountain cautioned it doesn't know the data's format and worried it could be lumped with other OSS clients' data, which it wants to avoid corrupting.

Milo: St. Louis broadcaster Nine PBS sued Iron Mountain in late July after Iron Mountain refused to return the materials. The sticking point was that OSS, the now-defunct storage company Nine PBS had originally contracted with, technically owned the physical systems holding the data. But the judge found Nine PBS is the rightful owner, ordered Iron Mountain to cooperate, and set up a path for the station to bring in a third-party vendor, like a former OSS employee, to actually get at the data within thirty days.

Mia: Iron Mountain's lawyer did flag real concerns — the company doesn't know the data's format and worries it might be mixed in with other OSS clients' data, which it wants to avoid corrupting. If retrieval turns out to be complicated, if there's encryption involved, another hearing could follow. Once the data comes back, Nine PBS also has to confirm no other customer data is swept up in it, and it has to defend and indemnify Iron Mountain if anything else gets corrupted.

Milo: The station's vice president said the court confirmed Nine PBS rightfully owns the archives, which represent an important part of the region's history. For anyone who's ever worried their data could get tangled up in a failed vendor's mess, this case is a reminder that figuring out who actually owns the hardware can become the whole battle.

Mia: That's it for today's rundown — plenty to think about, from AI tools hacking each other to courtrooms untangling data ownership. Thanks for listening, and take care.

Mia: OpenAI's flagship GPT-5.6 model, Sol, just got a major price cut on the OpenRouter marketplace — a 50 percent discount that brings it to two dollars fifty per million input tokens and fifteen dollars per million for outputs, with cache reads at twenty-five cents per million. OpenRouter describes Sol as built for complex reasoning, coding, and agentic workflows, and particularly strong for command-line and multi-step coding and long-horizon problem solving, with a 1,050,000-token context window, a 128,000-token max output, a February 2026 knowledge cutoff, and a July 9, 2026 release date, with providers including OpenAI, Azure, and Amazon Bedrock US. OpenRouter reports a hundred percent uptime over three days and 99.83 percent availability. While the listing highlights Sol's strengths, Hacker News commenters are unpacking the discount. One commenter noted that Luna saw a large jump after its price cut and is among the more competitive models at the new price, speculating OpenAI may be testing how much market share it can grab with Sol — while adding that cheaper models with roughly Sol's intelligence already exist, most notably Grok 4.6 at six dollars per million, which makes Sol a tougher sell. Another commenter disputed that Grok 4.6 actually matches Sol 5.6's intelligence, and a third wondered whether xAI is A/B testing routing some difficult Grok 4.6 queries to Sol to seed true believers.

Milo: And OpenRouter touts it as especially strong at complex reasoning, coding, and agentic work, with a context window well over a million tokens. Providers include OpenAI, Azure, and Amazon Bedrock, with availability reported north of 99 percent.

Mia: The top apps chewing through it, per OpenRouter, include Codex, Claude Code, and a few agent products. But dissenting voices on Hacker News see the discount as a sign that OpenAI is testing how much market it can grab at this price.

Milo: Right — one commenter argued Luna jumped after its own price cut, and that cheaper models with roughly Sol's intelligence already exist, citing Grok 4.6 at six dollars per million. Another commenter pushed back hard, disputing that Grok matches Sol's intelligence level.

Mia: So the pricing story is genuinely contested. And one commenter even floated the idea that xAI might be A/B testing by quietly routing some of its harder Grok queries to Sol, trying to seed true believers. Either way, the discount makes Sol a much easier buy.

Milo: A Roboflow engineer, Piotr Skalski, wrote a July 16th blog post arguing that OpenAI's GPT-5.6 Sol is the best vision model OpenAI has released yet. Skalski notes that OpenAI launched the GPT-5.6 lineup — Sol, Terra, and Luna — the previous week, with a release stream focused heavily on computer use, UI agents, and detailed 3D visualizations. Roboflow tested the models on its upcoming VLM benchmark, covering detection, counting, OCR, and data extraction, and saw the largest improvement in object detection: GPT-5.5 scored 13.8 mean average precision, while Sol reached 46.2, with Terra at 44.7 and Luna at 43.3. Document layout detection was a clear strength, with Sol handling titles, paragraphs, tables, images, and signatures. Coordinate format mattered a lot — GPT-5.6 models performed best with absolute XYXY coordinates in image pixels, whereas Gemini 3.5 Flash performed best with YXYX coordinates normalized to a 0 to 1000 range, and using the wrong format cut GPT-5.6's detection performance by around 15 mean average precision points. In a few cases Sol returned boxes in seemingly random parts of images, often in unnatural layouts like straight rows, and Roboflow shared examples with OpenAI, whose team confirmed Sol becomes less stable on images around 2,000 by 2,000 pixels or larger, especially at lower reasoning effort, with higher effort improving stability at the cost of token use and latency.

Mia: That's a bold claim, given last week's launch stream leaned heavily into computer use — UI agents and detailed 3D visualizations. Roboflow tested the whole lineup on its own benchmark, which it plans to release in the next few weeks.

Milo: The standout jump was in object detection. GPT-5.5 scored under 14 on their benchmark, while Sol hit 46, with Terra and Luna close behind. Sol was especially strong at reading document layouts — titles, paragraphs, tables, images, signatures — and handled dense scenes like pills and eggs surprisingly well.

Mia: Coordinate format mattered a lot. GPT-5.6 models performed best with absolute pixel coordinates, while Gemini 3.5 Flash preferred normalized ones — and using the wrong format cost Sol around 15 points of detection performance. There were also occasional glitches where Sol drew boxes in seemingly random parts of an image.

Milo: Roboflow shared those cases with OpenAI, which confirmed Sol gets less stable on very large images — around two thousand pixels and up — especially at lower reasoning effort. Crank the effort up and stability improves, but that costs tokens, latency, and potentially more.

Mia: Artificial Analysis's intelligence index put Alibaba's open-weights model, Qwen3.8 27B, at a score of 52 — well above the median of 9 among comparable models. That index combines nine evaluations, including Terminal-Bench, SciCode, GPQA Diamond, and Humanity's Last Exam. The site describes it as among the leading models in intelligence and well priced for an open-weights model of its size. Qwen3.8 27B supports text and image input, has a 256k-token context window, 27B total parameters, an Apache 2.0 license, and weights on Hugging Face. During the index evaluation it generated 160 million output tokens, which Artificial Analysis flags as very verbose versus a median of 43 million — and listed pricing is zero dollars for both input and output. On Hacker News, commenters put that score in context. One said it puts Qwen3.8 27B on par with far larger models like GLM 5.2 and GPT 5.6 Luna. Another added that it matches the latest DeepSeek Flash 0731, which has 284B parameters with 13B active, and called it the second-best Qwen model — much better than Qwen 3.7 Max but significantly below Qwen 3.8 Max. Another commenter questioned whether the sizes of those larger models are actually known, and a reply disagreed, saying GLM 5.2's size is known exactly and there's enough data to draw conclusions about Luna.

Milo: That's a reasoning model with a big context window — roughly 384 printed pages — that accepts image and text input and responds in text. It's released under an open Apache 2.0 license with weights on Hugging Face, and it's listed as essentially free to run per token.

Mia: Hacker News commenters put that score in context. One said it puts the model on par with far larger systems like GLM 5.2 and GPT-5.6 Luna. Another added that it matches the latest DeepSeek Flash release, and called it the second-best Qwen model — well behind Qwen 3.8 Max though.

Milo: There's debate about whether the sizes of those bigger models are actually public knowledge. One commenter questioned it, another pushed back saying GLM 5.2's size is known exactly. The one clear caveat: during evaluation the model generated a lot of output tokens — flagged as very verbose compared to the median.

Mia: A Y Combinator-backed startup called Speko launched on Hacker News as, in its own words, the router for voice AI — an API that measures 61 speech and language models across 10 languages and routes each session to the model that wins for the caller's constraints, whether that's accuracy, latency, cost, or balanced, along with language and region. A gateway prefetches signed session plans so new calls avoid a control-plane round trip, and if a provider refuses the connection during setup, it fails over to runners-up. Bek, previously cofounder and CTO of enterprise voice agents across Asia for four years, says Speko automates the manual re-benchmarking his teams did with native-speaking raters. The gateway is MIT-licensed Go, runs as a sidecar, and in BYOK mode never communicates with Speko, though the gateway and BYOK are free while the hosted router and managed keys are paid. Hacker News commenters focused on architecture. One argued the industry is moving toward one end-to-end trained model mainly for latency, and that the real value is automatic evals, not routing. Another disagreed, saying enterprises are not moving to speech-to-speech meaningfully, citing composability, observability, reliability, and cost. A third replied that roughly 95 percent of cases they see are still cascaded speech-to-text plus LLM plus text-to-speech, and that Speko measures end-to-end models.

Milo: The idea is that it routes each session to the model that wins for the caller's constraints — accuracy, latency, cost, or balanced — along with language and region, and returns the provider and model names in the headers. The founder previously built enterprise voice agents across Asia and says this automates the manual re-benchmarking his teams did with native-speaking raters.

Mia: Commenters on Hacker News split on the approach. One argued the industry is moving toward one end-to-end trained model, mainly for latency, and that the real value here is automatic evals, not routing itself. Another disagreed, saying enterprises are not moving to speech-to-speech meaningfully because they need composability, observability, and cost control.

Milo: A third voice backed that up, noting that most cases are still the classic cascade of speech-to-text, then the language model, then text-to-speech. The gateway itself is open-source, runs as a sidecar, and in bring-your-own-key mode never talks to Speko's servers — though anonymous telemetry is on by default.

Mia: A librarian published a guide on librarian.net tackling a question people keep bringing to the library's drop-in help sessions: how do you disable or avoid intrusive AI. The librarian says that besides what's taking up all their cloud storage, one of the biggest questions they get is how to disable AI that shows up where people don't want it, and the guide is aimed at people who'd like less intrusive AI in their tech environment. It covers turning off Apple Intelligence, which the guide says exists only on iPhone 16 and newer Macs and iPads, and notes Apple's learn from this application feature can be on even when Apple Intelligence is off; disabling Gemini on Android, including in Messages and from the power button; blocking browser-embedded AI such as Chrome's GLIC flags and Edge's Copilot and sidebar settings; the AI Controls in Firefox 148 and later; and turning off Google Workspace smart features in Gmail. On Hacker News, commenters debate whether people actually want these features. One calls the guide cool but describes a bizarre problem of companies forcing features nobody wants that are also expensive to operate. Another wonders whether the market is irrational or just hard to see from outside, suspects many people dislike AI but begrudgingly use it because it's hard to avoid, and suggests the strategy may be making AI so invasive that reverting becomes too expensive.

Milo: It's written for people who want less AI in their tech environment — and it clearly says folks who like the AI features can skip it. It walks through turning off Apple Intelligence, including a hidden feature that can keep learning from your apps even when Apple Intelligence itself is off.

Mia: It also covers disabling Gemini on Android, including inside Messages and from the power button, blocking browser-embedded AI like Chrome and Edge's Copilot, the AI controls in recent Firefox versions, and turning off smart features in Gmail and Google Workspace.

Milo: Commenters debate whether people actually want these off. One calls it a bizarre problem — companies forcing features nobody wants while spending a lot to run them. Another suspects many people dislike AI but use it begrudgingly because it's hard to avoid, and wonders if the play is making AI so invasive that turning back becomes too expensive.

Mia: And there's pushback from the other side arguing that the market isn't irrational — that a lot of users quietly do want these conveniences. Either way, the guide exists because the question keeps coming up, and knowing where these switches live is genuinely useful.

Mia: A self-described pro-AI writer has coined a term for his new pet peeve: unedited AI output. Rick Manelius, in his newsletter AI;DR — AI, didn't read — says unedited AI writing has become a pet peeve even though he's about as pro-AI as you can be. He credits a tweet by @seclilc for the acronym, which he adopts as his solution for ignoring what he calls walls of slop. His stated policy: if someone isn't bothered enough to review and edit AI-generated text, he isn't going to bother reading it. He notes exceptions where fully AI-generated copy is expected, citing customer support as a perfect example. He describes physically flinching, hunching his shoulders, and having a slight eye twitch when someone he respects sends unfiltered, unedited AI output, and he mentions receiving a wall of Claude output in Slack discussions, plus newsletters and social content with what he calls weird AI-isms. He frames TL;DR as the solution for social media and AI;DR as the solution for AI slop, ending with a wish that people prioritize a human touch when they talk to him. In the Hacker News discussion, commenters riffed on the acronym rather than disputing the article — one offered HW;DR: human-written, did read, another proposed TA;DU: too ambiguous, didn't understand, and another posted HW;WR: human written, worth reading.

Milo: So he's framing TL;DR as the solution for social media, and AI;DR as the solution for what he calls AI slop. And the Hacker News crowd ran with the acronym instead of debating it. One commenter suggested HW;DR — human-written, did read. Another countered with TA;DU — too ambiguous, didn't understand. Someone else pitched HW;WR, human written, worth reading, and another posted LE;DU, low-effort, didn't upvote. There was also a whole thread riffing on AI-DR, with replies joking about getting the reference and calling it load-bearing.

Mia: His closing wish is simple — that people prioritize a human touch when they talk to him. And honestly, that's a fair standard. The back-and-forth on Hacker News treated it as a bit of fun rather than something to fight over, which suggests a lot of readers quietly recognize the feeling.

Milo: Fairphone's latest handset has made a meaningful leap for open-source phone software. The story on Hacker News is that the dedicated main camera now works on the Fairphone 6 running postmarketOS, a Linux-based mobile OS built to replace stock Android. That's a notable milestone because camera support — especially the main sensor with its proprietary tuning — has historically been one of the hardest pieces for community-built mobile Linux to crack. In the Hacker News discussion, one commenter responded with skepticism, saying congratulations for what would be pre-Android 1.0, drawing a comparison to the PinePhone Pro, which that commenter said couldn't even do the phone part of its name and shipped with tons of non-functional hardware.

Mia: But the discussion wasn't all celebration. One commenter pointed out, a bit wryly, that congratulations are maybe premature for what would be under Android 1.0 functionality. They compared it to the PinePhone Pro, saying that phone couldn't even reliably do the phone part. Lots of hardware went non-functional, they argued, because the pieces were cobbled together without proper software support.

Milo: So there's a real tension here between the technical win and the practical reality. Getting the main camera working is genuine progress for postmarketOS and Fairphone — but it also underscores just how far behind mainstream Android devices these community phones still are when it comes to day-to-day usability.

Mia: A new arXiv paper proposes zero-overhead GPU programming built natively into the Rust compiler. GPU Offload in Rust: Portable, Safe, and Fast, by researchers including Manuel S. Drehwald and Alán Aspuru-Guzik, lays out a multi-vendor framework for compiling GPU code through rustc and LLVM. The authors argue that high-performance GPU programming has traditionally forced a compromise between execution efficiency and memory safety — you either accepted explicit unsafe raw pointers or got locked into a vendor-specific domain-specific language. Their approach applies Rust's compile-time type and ownership safety to massively parallel GPU workloads, leveraging strict aliasing guarantees and a two-pass compilation pipeline to handle cross-vendor ABI mismatch. Evaluation on RAJAPerf reportedly generated competitive LLVM IR for GPU kernels with solid performance against native, hand-optimized CUDA and HIP C++ baselines. In the discussion, one commenter asked whether the authors published code, and a response said it's part of the Rust codebase. Another debated why the rust-gpu project's need to emulate pointers was called a blocking issue, with an explanation that pointers are needed for high-performance memory management in HPC design.

Milo: And the evaluation looks solid. The paper reports it generated competitive compiler output for GPU kernels on a well-known HPC benchmark suite, with performance against hand-optimized CUDA and HIP C++ baselines. In the Hacker News thread, someone asked whether the code was published, and a commenter said it's part of the Rust codebase itself, pointing to the compiler development guide and a Rust issue. Another commenter challenged why a rival project's need to emulate pointers was considered a blocking issue — and the answer came back that pointers are essential for high-performance memory management in HPC design patterns.

Mia: So the big idea here is that you wouldn't need a separate GPU language or unsafe Rust at all — just write safe Rust, and let the compiler handle moving data between host and device. That's a meaningful step toward making portable GPU code feel as safe and comfortable as ordinary Rust.

Mia: DuckDB, the in-process analytics database, is gearing up for a major release this fall — and it's a big one. Version two marks a major version bump, named Cyanoptera after the cinnamon teal duck species, and ships a new SQL parser, a new default storage format, a reworked C API, and a small set of carefully chosen breaking changes, built from over ten thousand commits since version 1.5 in March. The headline features previewed include DuckDB as a server, triggers, the VARIANT type, and asynchronous I/O. The year of DuckDB as a server centers on the quack extension, which implements DuckDB's native protocol for talking to other DuckDBs, graduates to stable in v2.0, and works with the new CONNECT statement — any DuckDB process can serve databases over the network, and any other DuckDB can attach and route queries there. CONNECT also works with other remote databases, with a new remote pushdown optimizer that ships SQL directly to PostgreSQL and MySQL. VARIANT, described as JSON on steroids, shipped in v1.5; v2.0 makes the pipeline work end to end. Triggers arrive with BEFORE and AFTER, FOR EACH ROW and FOR EACH STATEMENT.

Milo: Now here's the headline shift — the team calls it the year of DuckDB as a server. They're graduating a tool called the quack extension to stable, which implements DuckDB's native protocol. Combined with the new CONNECT statement, any DuckDB process can serve databases over the network, and any other DuckDB can attach and route queries to it. And CONNECT isn't just for other DuckDBs — there's a new remote pushdown optimizer that can ship SQL directly to PostgreSQL and MySQL.

Mia: There's a lot else riding along in v2. Triggers arrive with before and after versions, row-level and statement-level. The VARIANT type, which they've described as JSON on steroids and first shipped in 1.5, now works end to end with fast execution straight from storage. And the post notes DuckDB was built as a transactional, multi-connection database with full MVCC from day one, so v2 pushes even harder on metrics, logs, and observability.

Milo: The Hacker News reaction captured the shift nicely — one commenter said the last year of DuckDB enhancements feels like a move from an in-process engine toward something that behaves much more like a full database server. That's a hint of where the project's ambitions are heading.

Milo: A thirty-year-old experiment in DRM from id Software makes for a great retelling. Fabien Sanglard's article about the Quake shareware CD recounts that Quake only took up about 22 megabytes, so id filled the rest of the disc with encrypted versions of its entire game catalogue. The play was to sell the shareware CD for $9.95 at CompUSA and Computer City, get gamers to call an 800 number, pay, and receive a password that would unlock whatever titles they bought. The disc was announced July 3, 1996, released August 30th, and the hacker group GNOMON cracked it just 39 days later with a tool that decrypts every game on the CD. According to David Kushner's Masters of Doom, the end result left id stuck with almost 150,000 CDs sitting in a warehouse. In the Hacker News discussion, one commenter questioned whether the scheme deserves the label security through obscurity, arguing that by that logic real symmetric encryption would fall under the same label too.

Mia: According to David Kushner's book Masters of Doom, id was left sitting on almost 150,000 CDs nobody wanted. The scheme's mechanics were genuinely clever on paper — the unlock screen generated a challenge code you read over the phone and got a one-time serial back, with checksums to guard against phone-line mistakes, and the challenge rotated every five minutes. A separate source-code label apparently played no part in generating the unlock — likely how distributors claimed their transaction fee. The lock-and-unlock tooling came from a company whose encrypter could denature a Windows executable.

Milo: There's a fun coda — that 800 number is still active, but it now plays an automated message selling elderly-care products, since CompUSA itself shut down around 2008. In the Hacker News thread, one commenter pushed back on calling the scheme security through obscurity, arguing that by the same logic, real symmetric encryption would carry the same label. It's a good reminder that the line between flavor-of-the-era DRM and actual cryptography is thinner than it looks from a distance.

Mia: Menachem Kornreich's browser Sokoban AI Solver claims the provably fewest-moves solution for the 1980s "warehouse keeper" puzzle, where you push every box onto a goal. This variant also makes the keeper finish on a goal, so each board has one more goal than boxes. It's a plain-JavaScript port of the author's native C++ optimal solver, using A* with whole-push edges, ~8-byte bitmask states, a dial bucket queue with an open-addressed hash, and deadlock pruning. Boards 1 through 14 solve live to the proven optimum in milliseconds, while board 15, the 8-box maze, explores roughly 49 million states and needs 1 GB, so its 184-move optimum was computed offline by the C++ parallel A* in about 5 seconds across 24 cores and replayed. In the Hacker News discussion, commenters debated whether this counts as AI. GPerson called it "AI in the older sense from 10 years ago," while nairboon replied it's still AI, just not an LLM; GPerson later argued an algorithmic puzzle solver was never called AI, since in games AI means NPC behavior. dev_dan_2 traced the association shift to AlphaGo in 2015, calling this optimized A* with heuristics. cbondurant found the proven optimum impressive but the examples not hard when move count doesn't matter, and wanted a trap-heavy board. Retr0id recalled Sokoban as CS coursework, with the state space blowing up on larger canvases, and npinsker wondered whether WASM could handle the final board and whether storing keeper-reachable positions would help.

Milo: And what makes it stand out is the engine behind it. A plain-JavaScript port of the author's native C++ optimal solver runs entirely in your browser, and it returns the fewest-moves solution possible. It uses a whole-push search with compact state encoding and deadlock pruning, so the first fourteen boards solve live to the proven optimum in milliseconds.

Mia: But the big showcase is board fifteen, an eight-box maze that explores tens of millions of states and needs about a gigabyte of memory. That board's 184-move optimum was computed offline by the parallel C++ solver across 24 cores in roughly five seconds, then replayed in the browser.

Milo: What really sparked the discussion was whether this counts as AI at all. One commenter called it AI in the older sense, from ten years ago. Another replied it's still AI, just not a language model. And someone else argued that an algorithmic puzzle solver was never called AI, since in games that term means non-player behavior. One commenter traced the shift to AlphaGo back in 2015.

Mia: Someone else found the proven optimum impressive but wished for a trap-heavy board where move count really matters, while another recalled Sokoban as computer science coursework where the state space explodes on larger boards. So the real takeaway here is less about the puzzle itself, and more about how one person shipped a genuinely optimal solver in the browser, and rekindled a debate over what we even mean by AI.

Mia: Tim Marinin's blog post "How Bluesky draws its logo on screenshots" explains why the Bluesky logo appears at the top right of screenshots of posts even though the app shows a "Follow" button in that spot. The explanation, found in the open-source Bluesky app code, is a file called GrowthHack.tsx, introduced in January 2026 by a developer named mozzius. It uses a package also by mozzius called expo-privacy-sensitive, which creates a UITextField with isSecureTextEntry set to true and renders the button's actual content into that field's layer; when a screenshot is taken, iOS blanks the secure field's layer, revealing the logo underneath. Marinin, who says he isn't an iOS developer, speculates that switching between apps produces an inert snapshot that doesn't trigger the blanking. Telegram and Signal use similar tricks for "secret" chats, so he doesn't expect Apple to patch it, calling the trick cute, though he notes most people in the discussion thread disliked it. In the Hacker News discussion, 3form called apps reacting to screenshots hostile and annoying, while drdexebtjl said the concern is the broader trend, pointing out that a popular bank in their country blocks screenshots and screen recordings on Android. shiandow argued the OS shouldn't place an app's wishes above the user's, while javier2 countered that nothing is hijacked since the button is rendered in a sensitive UI element.

Milo: So the key is a file called GrowthHack, added in January of 2026 by a developer named mozzius. It uses a package, also by mozzius, that creates a secure text field and renders the button's actual content into that field's layer. When you take a screenshot, iOS blanks out the content of a secure field, revealing the logo hidden underneath.

Mia: So on a phone, pressing the normal screenshot buttons reveals the brand logo where the Follow button would be, purely because iOS hides anything rendered in a secure text field. On other platforms it just shows as-is, no masking. And Marinin, who notes he isn't an iOS developer, speculates that switching between apps produces a snapshot that doesn't trigger the blanking.

Milo: He points out that Telegram and Signal use similar tricks for their secret chats, so he doesn't expect Apple to patch it, and he calls the trick cute. Though the discussion thread mostly disliked it. One commenter called apps reacting to screenshots hostile and annoying, and another replied that Bluesky supports third-party clients with minimal feature sets.

Mia: For one commenter, the concern is the broader trend, noting that a popular bank in their country blocks screenshots and screen recordings on Android. Another argued the operating system shouldn't put an app's wishes above the user's and that the control should be revocable, while a counterpoint was that nothing's actually hijacked, since the button is just rendered in a sensitive UI element.

Mia: That wraps things up for today's conversation. Thanks so much for listening along.

Milo: We appreciate you spending this time with us. Take care of yourselves, and we'll catch you on the next one.