Relevance 10/10Importance 9/10
Mistral dropped its first family of time-series foundation models — four sizes from 5M to 209M parameters — doing probabilistic multivariate forecasting with 21 quantiles, and they claim first place on GIFT-Eval, TIME, and fev-bench using roughly two-thirds the parameters of the previous leader. The big models run on plain CPUs, forecasting a day ahead from three weeks of 15-minute history in about 140 milliseconds. Shipped under a Modified MIT license on GitHub and Hugging Face, with Mistral Forge for private fine-tuning.
Relevance 10/10Importance 8/10
MiniMax launched M3.1 Flash, an agentic-first model built on the "Space Bunny" base that chewed through 200 trillion tokens, pushing 200-plus tokens per second and cutting agentic task cost by about 70 percent. Benchmarks are heavy on tool use and security: 73.7 percent on DeepSWE v1.1, 76.2 on BankerToolBench, 61.7 on OSWorld V2, and 82 percent on CyberGym vulnerability discovery. Pricing is fifteen cents per million input, sixty cents output, with weights promised later.
Relevance 9/10Importance 6/10
A developer going by pokpokpok built a benchmark that drops LLM agents into a reconstructed 2004-era RuneScape server called LostCityRS, complete with an rs-sdk for goal-directing agents and a live visualizer of server activity. It measures navigation, multi-agent coordination, and in-game economic behavior — resource gathering, trading, quests. Commenters found models hallucinate wildly about game mechanics without priming, and that RuneScape's scam-rich world is a weirdly good testbed for trust and risk assessment.
Relevance 7/10Importance 7/10
Chris Kiehl pushes back on the prevailing gloom, conceding that AI has changed the economics of hand-written code at work while insisting that work isn't life. He hand-coded a Breakout game knowing full well Claude could one-shot it, because the point was the problem-solving, not the artifact. The thesis: exploration has intrinsic value, and enjoying programming needs no economic justification. 325 comments say it struck a nerve.
Relevance 7/10Importance 5/10
Josephine Pfeiffer put Arch Linux ARM with custom kernel patches on an ASUS ProArt P16, and got 116 GiB of genuinely unified memory addressable by both CPU and CUDA — versus Windows carving it into fixed GPU and CPU slices. 7-Zip multithreaded compression came in about five percent faster than published Windows numbers on comparable hardware. She then handed nearly 90 GiB to a local NVIDIA Qwen model and ran 171 consecutive inference tasks at roughly 30 tokens per second.
Relevance 6/10Importance 7/10
Peter Eisentraut counts eight to ten significant feature reverts during PostgreSQL 19's beta, against a historical norm of about one per cycle. His causes: insufficiently vetted commits, a security "Vulnpocalypse" from April through August that pulled senior devs off review, PostgreSQL's combinatorial "everything works with everything" promise, and — notably — LLM-assisted code review now catching edge cases that used to take years to surface, retroactively raising the quality bar. He floats an "experimental" feature tier as a release valve.
Relevance 5/10Importance 7/10
Mac.apk runs Android apps natively on Apple Silicon with no emulator, no VM, and no Linux layer — it reimplements the Android framework on top of macOS directly. App code executes at native CPU speed with Mac GPU rendering, and each app shows up as a real macOS process with native windowing, dock integration, resizing, and dark mode. At version 1.0.2237 it's got 65 apps tested, 51 fully playable, licensed AGPL v3 with a commercial option.
Relevance 4/10Importance 7/10
Thomas Frohwein proposed a Wine port for OpenBSD, crediting Andrii and Fabien for cracking the long-standing %gs register blocker via software emulation. It runs several dozen 64-bit Windows programs mostly successfully, including The Witcher 3 Remaster — but WoW64 is unsupported, so no 32-bit apps. Built-in tests are disabled pending X server access issues, and the author recommends running Wine under a separate unprivileged user. Top-scoring story on the list at 452 points.
Relevance 3/10Importance 5/10
Joshua Bird built a flying drone driven by a single motor, using a custom hinge, a magnetic encoder, and torque modulation to vary rotor lift mid-rotation — eliminating a helicopter's swashplate entirely. All the stabilization moves out of mechanics and into the control loop. The catch: the whole airframe spins continuously in flight, which commenters noted could actually be a feature for 360-degree imaging. Bonus: he built a high-speed camera to chase down a vibration bug.
Relevance 1/10Importance 6/10
Fishmeal used to spontaneously combust — at least 90 fires on ships and at processing plants from the 1950s through the 1970s — so in 1967 the feed industry adopted ethoxyquin, a chemical borrowed from rubber tires, cutting annual ship fires from sixteen to five. The problem is a 1987 British study found severe damage in lab rats and a 2015 European analysis flagged a mutagenic impurity. It's banned in the EU and UK but still permitted in the US, Australia, Japan, and South Korea, with producers shifting to BHA, which has its own carcinogen flag from 1991.
HN Briefing PM
