The author runs a self-hosted stack on a $10 OpenStick (512 MB RAM, about 450 MB usable; four 1 GHz Cortex-A53 cores; 4 GB eMMC; 150 Mbit/s Wi-Fi; 4G and GPS) after a somewhat involved installation that leaves it running Debian with systemd and Podman. Services include CastSponsorSkip, AdGuard Home, Zeroclaw, Cloudflare Tunnel, Beszel, Dozzle, Gatus, Fusion, Shiori and Diun. The author's AdGuard Home fork handles large blocklists efficiently, while the official version is fine only with smaller blocklists (about 200k entries). Zeroclaw is given access to the OpenStick only, nothing beyond it, and is used mainly to monitor logs and disk space via Telegram. Cloudflare Tunnel exposes services without opening ports, at the cost of Cloudflare being able to see traffic. Beszel uses about 20 MB total (10 MB agent, 10 MB server) and can send Telegram alerts.

Independent researcher Derik Kauffman found a hard-coded “grey list” of more than 100,000 words and phrases that Apple’s iPhone predictive text will not suggest and that often fail when swipe-typed, sometimes producing altered results such as “United Iceland” for “United Ireland.” The list includes drug and sexual references, “iMessage isn’t working,” and politically charged terms such as “Trayvon Martin,” “Kashmir dispute,” “Muslim extremist,” “male chauvinist pig,” the Hebrew phrase “the occupied territories” and the Arabic phrase “the Hamas movement.” Apple says the keyboard is designed to avoid inappropriate or offensive predictive suggestions, and users can still type words out or override suggestions.

MaxxBench tests how much coding performance depends on familiarity with a benchmark by changing the instructions or starting code of 89 DeepSWE medium-duration feature-implementation tasks, revising the reference solution and tests accordingly; the reference patches grow by a median of 17 lines (3.5%). The revised requirements must be met while the original feature keeps working, for example lowercase report filenames (chrome.xml instead of Chrome.xml) and UTC timestamps with a trailing Z. Across 15 configurations, the spread between highest and lowest scores widens from 20.9 points on DeepSWE to 48.3 on MaxxBench; Claude Opus 5.5 scores 76.4% versus 74.2%, while Gemini 3.8 Flash falls from 73.8% to 53.4%. Trajectories show models guessing hidden tests and reverting to original DeepSWE behavior.

The author reports that the Ox64, a RISC-V board with three different cores, can run Linux for about $8, but notes that at the time of writing ethernet, wifi and the SD card did not work. Programming it is not trivial: it cannot be flashed over USB like an Arduino and must be programmed via UART TX/RX pins. The Dev Cube tool crashes on launch under Ubuntu on Wayland, requiring Xorg. A CH340G dongle failed the handshake, so the author used an RP2040 (KB2040) loaded with picoprobe as a serial bridge. DevCube still could not flash the board, so bflb-mcu-tool was used after editing bflb_interface_uart.py to strip DTR and RTS handling. The author then flashed low_load_bl808_m0.bin and received serial output, but had not yet run Linux.

Coldframe is a single-file Python honeypot that impersonates a Danfoss AK-SM 800A or Copeland/Dixell XWEB EVO refrigeration supervisor, logging connection attempts, tried passwords and recognized public exploits. The Danfoss persona serves HTTP and HTTPS, but HTTPS is off by default and requires CERT_FILE and KEY_FILE. An optional Modbus TCP service is also off by default, since real AK-SM 800A units use Modbus only with their own case controllers. Vendor favicons and the Danfoss firmware index.html are not bundled; serving them requires FAVICON_FILE and INDEX_FILE pointing to copies the deployer may use, for private deployments. Manual container deployment needs rootful podman or Docker, because rootless port forwarding rewrites source addresses. The honeypot never runs commands, reads files or forwards traffic; exploit payloads are logged and dropped.

Ultraviolet light is the dominant lethal factor for microbes in space, but a thin dust film, a crevice or other microbes can shield survivors. In a three-year ISS experiment, D. radiodurans colonies were revivable and in better shape than an inside control; a 1 mm pellet was estimated to survive direct interplanetary exposure for 2–8 years, and potentially decades if shielded from sunlight. Cosmic radiation then limits survival: roughly 50,000–100,000 years in a basketball-sized rock, with 50,000 years the upper limit for small rocks and 500,000 years for larger ones. Permafrost bacteria 400,000 and 600,000 years old showed viability evidence, but 740,000- and 1,000,000-year-old cores did not. Claims of reviving microbes from ancient halite have been undermined by contamination.