tl;dr: Bluesky replaces its "Follow" button with a logo watermark in iOS screenshots by exploiting `UITextField`'s `isSecureTextEntry` property—rendering the button inside a secure text field's layer, which iOS blanks out during screenshots, revealing the logo beneath. The same trick is used by Signal and Telegram for private content, so Apple is unlikely to patch it. Community reception in the PR thread was largely negative before it was locked.

tl;dr: Summary not available

tl;dr: In 1996, id Software released a Quake shareware CD-ROM that also contained encrypted versions of their full game catalog, unlockable via a phone-in payment system powered by TestDrive Corp's software. The scheme was cracked in 39 days by hacker group GNOMON, who discovered the "unlock code" contained no server-side secret—the local FLOW.EXE could generate valid serials itself from the challenge code, relying purely on security through obscurity. The CD was riddled with additional flaws, including unencrypted plaintext copies of "encrypted" databases and a typo that broke legitimate Final DOOM unlocks, leaving id with 150,000 unsold CDs.

tl;dr: DuckDB v2.0 ("Cyanoptera"), shipping this fall, adds client/server mode via the Quack extension and CONNECT statement (with query pushdown to Postgres/MySQL), first-class VARIANT type, SQL triggers, and asynchronous I/O for much faster object-store reads. Under the hood it introduces a new PEG-based SQL parser, a new default storage format, drops ICU in favor of a native 45KB timezone/collation implementation, and expands the stable C API so extensions built once keep working across versions. Recursive CTE queries benchmark ~40× faster than v1.5.

tl;dr: Wiz's autonomous AI security agent discovered a script injection vulnerability in a Snowflake GitHub Actions workflow that allowed anyone to execute arbitrary commands by opening an issue with a crafted title, ultimately exfiltrating a Jira token with access to Snowflake's engineering and security projects. The vulnerable code was introduced via a PR that removed a safer pattern; GitHub Advanced Security's scan and Copilot Autofix (a co-author on the PR) both failed to flag it. Snowflake patched within a day, but the incident highlights how AI can both introduce and rapidly discover vulnerabilities, collapsing exploitation windows to days.

tl;dr: UC Berkeley researchers used adaptive optics laser ophthalmoscopy to selectively stimulate only M cone cells in the retina, producing "olo"—a highly saturated blue-green color outside the normal human visible gamut, seen so far by only five people. The closest sRGB approximation is #00FFCC, though some scientists dispute whether it qualifies as a genuinely new color. The team is exploring applications for treating color blindness and potentially enabling tetrachromatic vision.

tl;dr: Israel paid $900,000 to Piro, Inc. to create the "Hanover Institute for Public Policy," a fake think tank publishing over 100 AI-generated reports on Israel/Palestine designed to be cited by LLMs like ChatGPT, Gemini, and Claude—a practice known as "LLM poisoning." The reports mimic credible think-tank output with footnotes, neutral tone, and citations (often to IDF and Israeli government sources), and GPTZero flagged 11 of 12 sampled articles as AI-written. This is part of a broader Israeli influence operation that includes a separate $46.5 million contract with Brad Parscale for similar chatbot-targeting websites.

tl;dr: Researchers built a multi-vendor GPU compilation framework directly into rustc and LLVM, using Rust's ownership and aliasing guarantees to safely manage GPU data transfers without vendor lock-in or unsafe raw pointers. The approach uses a two-pass compilation pipeline to handle ABI mismatches between host and device targets. Benchmarks on RAJAPerf show the generated GPU kernels achieve performance competitive with hand-optimized CUDA and HIP C++ baselines.

tl;dr: OpenAI's new GPT-5.6 lineup (Sol, Terra, Luna) shows major gains in vision tasks, with Sol jumping from 13.8 to 46.2 mAP@50 on object detection and improving counting across the board, though OCR performance stayed flat or slightly regressed. Sol becomes unstable on images near 2000x2000px and costs ~2.5 cents/image with ~10s latency, making Gemini 3.5 Flash still the better price/performance choice for high-volume detection and counting workloads.

tl;dr: A few months after switching from Gmail to Fastmail, the author reports it's gone smoothly, with subdomain addressing (auto-sorting mail to folders based on unique addresses given per account) replacing Gmail's automatic categorization. Using a custom domain enables easy future migrations, though freshly registered domains get temporarily greylisted by recipients like Gmail. Masked email for disposable addresses and multi-domain support have been unexpected wins.

tl;dr: The author proposes "AI;DR" (AI; Didn't Read) as a personal policy: if someone can't be bothered to review and edit their AI-generated output before sending it, they won't bother reading it. While fully AI-generated content is acceptable in some contexts (like customer support), unedited AI slop in Slack messages, newsletters, or social posts signals the sender doesn't value the recipient's time. The credit for the term goes to a tweet by @seclilc.

tl;dr: Sun Clock is a free, ad-free 24-hour web clock that visualizes the sun's position, sunrise/sunset, solar noon, golden hour, twilight, and moon phase/position for your current location. It rotates direction based on hemisphere (clockwise in the North, counter-clockwise in the South) to match the sun's actual movement, and works as an installable Progressive Web App with offline support. Location data stays local in the browser, and the source is MIT-licensed.

tl;dr: A practical guide (hosted at NoToAI.org) for disabling AI features across common apps and platforms, including Adobe, Android/Gemini, Apple Intelligence, Chrome, Edge, Firefox, DuckDuckGo, Google Workspace, Slack, WhatsApp, Windows 11/Copilot, Yahoo Mail, and Zoom. Each section provides step-by-step menu paths, relevant flags, or extensions, along with recommendations for alternative browsers (Zen, Helium, LibreWolf) and debloat tools like O&O ShutUp 10 and Win11DeBloat. The author invites additions and edits from readers.

tl;dr: Summary not available.

tl;dr: Anthropic's Dario Amodei argues that framing AI regulation as a binary between regulatory capture and open distribution is false, noting that well-designed rules can constrain frontier labs while advantaging smaller competitors—citing SB53 and SB1047's revenue exemptions as examples. He contends AI structurally concentrates power due to scaling laws, so open-weights alone can't solve it, and expresses support for the Trump administration's reported approach of pre-deployment testing for frontier models plus a FINRA-like oversight entity.

tl;dr: Summary not available.

tl;dr: Sheldon Axler's "Linear Algebra Done Right" (4th edition) is now available as a free Open Access PDF under a CC BY-NC license, with translations in Chinese, Farsi, Greek, and Portuguese. The textbook's distinctive approach defers determinants to the end and focuses on understanding linear operators on finite-dimensional vector spaces through cleaner, determinant-free proofs. The new edition adds 250+ exercises, 70+ examples, and several new topics.

tl;dr: Qwen 3.8 27B is an Apache 2 licensed vision-capable model that runs impressively well from a 17GB file on consumer hardware, handling coding agents, bounding boxes, and tool calls competently. However, its default "xhigh" reasoning setting causes absurd overthinking—spending 21 minutes and 22K reasoning tokens on a simple SVG prompt—so users should dial reasoning down to low. The main drawback is speed (15-30 tokens/sec), though Multi-Token Prediction support via llama.cpp offers a ~72% boost.

tl;dr: Anthropic is rolling out text watermarking on all Claude models globally to comply with the EU AI Act, using a SynthID-style technique that biases token selection toward "green-listed" words detectable only with Anthropic's secret key. The author argues this deliberately degrades word choice—the essence of writing—for a scheme that harms honest users (who risk false accusations when using Claude for proofreading) while being trivially defeated by bad actors via paraphrasing tools. He contends Anthropic's claims of "imperceptible" quality impact are unfalsifiable marketing, and that applying an EU regulation worldwide reflects either technical incompetence or corporate cowardice.

tl;dr: Summary not available