Relevance 10/10Importance 9/10
John Langford writes about the dissolution of Microsoft Research's AI Frontiers sub-org after it was folded into Microsoft AI, and his own departure along with the rest of AIF's leadership. He pins it on a "scaling-only" philosophy from new leadership that left no room for fundamental research or any path from research to shipped product. His argument: returns to real research are large and measurable, and abandoning innovation for pure scale is short-sighted.
Relevance 10/10Importance 8/10
MaxxBench tests whether coding agents can adapt when requirements shift, rather than how well they've memorized a popular eval. The team took 89 DeepSWE tasks and applied tiny mutations — renaming `Chrome.xml` to `chrome.xml`, switching timestamps to UTC — then measured whether models follow the new spec or talk themselves back into the remembered answer. It's a direct shot at "benchmaxxing," where a benchmark becomes part of the training loop and the score stops meaning anything.
Relevance 7/10Importance 8/10
Independent researcher Derik Kauffman found an undisclosed, multilingual table inside iOS that suppresses autocomplete and swipe-type suggestions — a conservative estimate puts it near 100,000 words and phrases, excluding Chinese characters. Examples include "dictator Trump" and "sleepy Joe" in English, "the occupied territories" in Hebrew, and "the Hamas movement" in Arabic. Apple says the keyboard is designed to avoid offering inappropriate or offensive suggestions; the researcher counters that the application is wildly inconsistent.
Relevance 6/10Importance 7/10
Coldframe is a low-interaction honeypot that impersonates Danfoss AK-SM 800A and Dixell XWEB EVO refrigeration controllers over HTTP, HTTPS, and optional Modbus TCP. It goes deep on fingerprint fidelity — header order, page titles, certificate fields, TLS versions — and lets attackers in on default creds so researchers can watch what happens after the initial probe, including CVE-2025-41450 attempts. The whole thing is a single Python file with stdlib-only dependencies.
Relevance 5/10Importance 6/10
The OpenStick is a hackable 4G modem dongle repurposed as a server: four ARM Cortex-A53 cores at 1 GHz, 512 MB of RAM with about 450 MB usable, 4 GB of eMMC, and no fan. The author stacks AdGuard Home for DNS blocking, Beszel for monitoring, Fusion for RSS, and Shiori for bookmarks — all in roughly 200 MB of memory. It's a solid reminder of how much self-hosting fits in a ten-dollar silent box.
Relevance 5/10Importance 5/10
The Ox64 is an eight-dollar RISC-V single-board computer with three different RISC-V cores that runs full Linux with memory management. At the time of writing, ethernet, wifi, and SD card were all non-functional. Getting it to boot took six separate workarounds, including UART-only programming, Wayland incompatibilities, swapping the flashing tool, and pressing an RP2040 into service as a serial bridge.
Relevance 4/10Importance 6/10
A fan team statically recompiled the 2004 PC game to WebAssembly so it runs natively in Chrome and Firefox. You get up to 120 fps in widescreen, free roam, online races with crews and friends, and support for controllers, wheels, and touch. The technique builds on the static-recompilation work pioneered in the N64 and Xbox scenes, applied here to x86 code.
Relevance 3/10Importance 5/10
A 2008 list that's aged beautifully: Abrash's Zen of Assembly Language, Noble's Scientific Forth, Reiter's Fractals, Visualization, and J, Lammers' Programmers At Work, and Brodie's Thinking Forth. The thread running through all five is real, non-contrived problems solved in unfashionable languages. Thinking Forth gets singled out for teaching problem decomposition without leaning on OOP or design patterns.
Relevance 3/10Importance 4/10
Retro Song Bird is a hardware device that accepts any Game Boy or Game Boy Color cartridge and plays its soundtrack using register-level sound emulation rather than recorded audio. It'll also archive save files and back up ROMs, and if you insert a Game Boy Camera it turns into a USB webcam. No price listed on the page yet.
Relevance 2/10Importance 5/10
Researchers analyzed two of the surviving uranium cubes from Heisenberg's B8 reactor alongside wartime technical reports and correspondence, publishing in PNAS Nexus. B8 would have needed roughly twice the uranium and more than twice the heavy water it actually had to go critical — a far larger gap than Heisenberg's postwar "slight increase in size" claim implied. Of about 1,100 cubes used across the German experiments, only 14 are accounted for today.
HN Briefing PM
