Read Tech News With Digest

NVIDIA's board announced that Jensen Huang has stepped down as president, CEO, and board member, effective immediately. CFO Colette Kress, who joined in 2013 after nearly 25 years in finance roles at Cisco and Microsoft, takes over as interim CEO while the board searches for a permanent successor. Harvey C. Jones, former CEO of Synopsys, becomes the board's first-ever chair. The company says operations and its product roadmap will continue uninterrupted.

I noticed that https://status.cloud.microsoft/ is very slow to load, and it shows a "Service Degradation" for "Microsoft 365 (Business or Enterprise)" with details that users may be unable to connect to their mailboxes in Exchange Online through multiple connection methods. The status page says they've identified connectivity errors on a portion of Exchange Online infrastructure and are investigating. Given that Teams, OneDrive, etc. are essentially SharePoint and Exchange in a trenchcoat, I suspect this is creating a mess for services beyond Exchange Online too. Edit: Looks like the status page needs a status page.

Google Antigravity now offers Gemini 4 Hydrogen, a standalone coding model scoring 51 on the Artificial Analysis Intelligence Index and 57% on Terminal-Bench 4.1. A new Orchestrator option pairs Hydrogen with Gemini 4 Carbon, passing work back and forth mid-trajectory: Carbon plans, Hydrogen edits and tests, and Carbon repairs the plan when Hydrogen stalls. On Terminal-Bench 4.1, Orchestrator matched Carbon's success rate while using 24% less quota and finishing 19% faster.

Rust 1.100.0 stabilizes the never type `!` after a decade and five failed attempts, letting it be named anywhere and making `Infallible` an alias for it. The release also ships the core `Allocator` API for custom allocators on `Vec` and `Box`, the `DropGuard` helper for running closures on every exit path, `conjure_zst` for safely creating zero-sized types, Cargo's unused-dependency warnings, and min-publish-age dependency cooldowns.

Sifting through Vera C. Rubin Observatory data, Matthew Holman and colleagues found 2026 ED138, an object 46 times farther from the sun than Earth that takes 281 years to orbit. Traced back to 1937 photographic plates, it may be 370 to 930 miles wide — big enough to be round, and possibly the sixth dwarf planet. Infrared or stellar-occultation observations are needed to confirm its size.

ZigZag is a free, open-source embroidery editor built in Rust that turns SVG artwork into machine stitches. It supports PES, DST, EXP, JEF, and VP3 formats, and can send designs directly to a Brother Skitch via Bluetooth. The visual editor offers Illustrator-style pen and shape tools, smart guides, and advanced stitch types like satin, fill, and appliqué. With built-in MCP support, AI agents can create and edit designs alongside you. Available for Mac and Windows, ZigZag eliminates subscription fees and gives you full control over your embroidery workflow.

Google's Gemini 4 Hydrogen is a speed-and-efficiency model for everyday coding and agentic work, priced flat at $1.50 per million input tokens and $7.50 per million output tokens across a 1-million-token context. It edges out Gemini 4 Argon on the Artificial Analysis Intelligence Index (51 vs. 50 at default, 54 at xhigh), triples Terminal-Bench 4.1 scores over 3.8 Flash, and powers a new Antigravity Orchestrator that switches between models mid-session.

While exploiting a PostgreSQL extension, security researcher Mehmet Ince wondered how PostgreSQL decides what a value is. That question led him to a type-confusion flaw in PostgreSQL core that had gone unnoticed for 22 years. By giving PostgreSQL a value inside a composite type, changing the type of one of its fields, and asking it to read the value back, he built one of the most reliable memory-corruption exploits he has seen, tracked as CVE-2026-6460.

openQ4 0.14.1, the Awakening Edition, lets players run The Awakening — a Quake 4 expansion Raven Software and Ritual Entertainment built but never released, recovered and published by Justin Marshall — alongside the original campaign. All 13 maps are rebuilt on openQ4's own engine and game code, patching broken alpha content like the Valkaryne finale and immobile Retches. The update also adds stable shadow mapping, Vulkan on Windows, a rebuilt interface, smarter bots, and experimental Android and VR support.

A developer pushes large language models onto hardware nobody expects to run them: a 1.9 GB Xiaomi TV box and a jailbroken 2013 iPad Air with 1 GB of RAM. Through custom kernels, 1-bit lookup tables, expert paging from flash, and weight streaming, the TV box hits 7.2 tok/s on an 8B mixture-of-experts model and 5.1 tok/s on a 16B one, while the iPad runs a 1.7B model at 24–26 tok/s. Every claim is measured on the device itself.

Relax is a live compiler and JIT runtime for Rust, written in Rust with 438k lines and zero crates.io dependencies. It runs a Cargo package straight from source, with no rustc, LLVM, or linker, compiling each function on first call and patching edits into the running process. A large 3D game shows its first frame 11 seconds after a clean checkout on an M4 MacBook, versus roughly 3 minutes with rustc, and 92 tests re-run in 2.2 seconds after a save. Optimized code still trails rustc --release.

AMD's Mike Clark, the original Zen architect, walks through what changed in Zen 6 and Zen 6c: a denser core that now carries the full 4MB-per-core L3 and the highest IPC in the family, larger queues to compensate for lower clocks, a reworked front end, two integer multipliers instead of three, extra VNNI pipes squeezed in late for inference, and a branch predictor that recovers faster without losing accuracy.

Cornell researchers analyzed four major U.S. power grids from 2020 to 2025 and found that data centers covered just 38% of the $8.25 billion in infrastructure costs tied to their growth, leaving $5.13 billion to residents. In the hardest-hit hubs, wholesale prices rose up to 20%, and the share of rural families facing a high energy burden climbed 76% faster than among urban residents. The study recommends shifting peak demand, charging data centers directly, and targeting relief to low-income households.

Kapso is an open-source engine from Leeroo that automates iterative experimentation on enterprise data. Give it an objective and success criteria, and it runs experiments in a sandbox, learns from each run via a git-based bank of lessons, and deploys the best solution to your stack. It is model- and coding-agent-agnostic, and has topped benchmarks like RelBench and OpenAI's MLE-bench. Use it to migrate legacy data estates, build lakehouse pipelines, or optimize ML models—with learning that compounds over time.

Eleven years ago, Jeff Kaufman pulled the emergency brake on Boston's Red Line, convinced a passenger's arm was caught in the door and would be severed. He later learned the man's bag—not his arm—was likely trapped, and that bystanders closer to the door saw it differently. The episode becomes a case study in how trolley-problem thought experiments fail in real life: we act under uncertainty, can't fully reason ahead, and the unilateralist's curse means one person's confident misreading can delay thousands.

Danish monastic burials show elevated nitrogen-15, indicating freshwater fish consumption. The reservoir effect can make radiocarbon dates appear anomalously old, sometimes predating the convent's construction by centuries. The authors warn that without stable isotope analysis and proper correction, medieval human bones from aquatic diets can yield misleading chronologies across Europe.

AgentRun is an open-source platform that supercharges your favorite agent harnesses—Claude Code, Codex, OpenCode, or pi—with self-improvement, durable execution, and a unified control plane. After each run, agents keep useful notes, skills, or workflows so they stop rediscovering the same APIs and docs, cutting token costs and latency over time. Runs survive server restarts, and you get approvals, spend caps, and one console for every agent. Start with a single npx command, add a model key, and run a template like KYB due diligence in minutes. Apache-2.0 licensed.

ExcalidrawZ wraps the familiar Excalidraw canvas in a native SwiftUI app for Apple platforms. It adds real document management, iCloud sync across devices, file history checkpoints, PDF and Mermaid import, Apple Pencil support, and an AI assistant that previews changes before touching the canvas. It can also run as an MCP server for compatible AI clients, and locked files stay encrypted and hidden from AI. Available on the App Store or as a direct .dmg download.

A demo called Distributed Ray Tracer uses Signed Distance Fields, with the entire scene written as code by a language model. Hatchet orchestrates parallel tiles rendered on Cloudflare Workers running Rust, and durable streams pass messages to each frontend. Built to showcase Hatchet's new serverless endpoints and durable streams, the renderer uses a Rhai and Fidget DSL you can edit yourself, and is based on a modified version of rpt.

Drawing on computability and complexity theory, this essay argues that writing correct programs is fundamentally hard, not merely a matter of better tools. The halting theorem and its corollaries—Rice's theorem, the Busy Beaver function—show that verifying arbitrary program properties is undecidable, and that the difficulty grows faster than any computable function. Even small programs, like a 1919-state Turing machine, can have behavior independent of ZFC set theory, meaning their termination cannot be proven.

Easel tested AWS Lambda Web Functions and found execution environments stay alive and bill memory for 5–7 minutes after the last request—up to 1,700 times the request duration in metered memory time. Even after terminating idle environments, AWS logged each exit as a runtime failure. Easel switched to self-operated compute, which delivered 1.5–3× faster warm requests and 3.5× faster cold starts than Web Functions.

Jane Street is trialing new onsite interview formats for campus candidates this summer, shifting away from pure code-writing toward software design, judging code quality and intent, and using LLMs to generate code. Most rounds stay traditional and LLM-free, and the firm stresses it doesn't expect candidates to have a polished agent workflow — just clear thinking about what the tools are good at and where they fall short.

Jane Street published a Socratic tutor skill for AI agents that teaches by asking questions instead of lecturing. It runs a short diagnostic to place you on a 0-5 skill map per subskill, then drills one question at a time, never handing over answers until you've failed two or three attempts. Levels are earned through repeated checks and unannounced spaced review, and flattery is explicitly banned.

Nintendo filed four DMCA notices with GitHub targeting Switch emulators Yuzu and Citron Neo, but the real focus was Picofly mod chip firmware — 144 repositories removed and replaced with takedown notices. The firmware lets users solder a chip to the Switch mainboard and bypass copy protection. It's still hosted elsewhere on GitHub, and after wiping 8,500 Yuzu mirrors in 2024, Nintendo may not be done.

Singapore says the U.S. offered little help as it investigated Alice Huang, the woman at the center of a spy plot that has put Nvidia chief executive Jensen Huang under scrutiny. Authorities held her passport from June 2025 to February over potential false declarations tied to her company, Megaspeed, then released her with a stern warning and barred her from returning. The timeline shows the passport was still held long after President Trump was briefed that she was believed to be a Chinese intelligence asset.

Anthropic has published its IPO roadshow, planning to list Class A shares on Nasdaq under the ticker ANT, with pricing expected November 18 and trading to begin November 19. Q3'26 figures show $17.0B in revenue, more than 12x year-over-year growth, over 11,900 customers spending at least $100K, and a 15% adjusted operating margin. A $23.3B net loss stems largely from a $21.9B non-cash charge tied to convertible notes and SAFEs.

Bootvine is a Rust tool that turns the boot console into a live tree of pending systemd jobs, inspired by nix-output-monitor. It also shows a timeline of the slowest units, handles shutdowns, answers LUKS passphrase prompts in place, and displays why a unit failed with its last journal lines. It requires systemd ≥ 246, libdbus, libsystemd, and a virtual terminal; a Nix flake lets you try a sample NixOS VM in QEMU without touching your system.

LibreAGI is building a swarm-based distribution network for open-source model weights, paying seeders more for scarce files: a model with only one copy earns four times the base rate. It accepts only safetensors and GGUF files under Apache-2.0, MIT, or the Qwen Community License, verifies each against its source commit, and staff-review every model. A signed public catalog records every hash and source so the network can be rebuilt if the site disappears.

A Broadcom PEX88096 switch feeding four AMD Radeon AI PRO R9700s and two RTX 3090s from a single x16 slot needs a full 32 GB Resizable BAR on each R9700 for ROCm's peer-to-peer mode. Linux 7.0's pbus_size_mem() sums child windows without leaving room for the 2 MB doorbell BARs, so only two cards fit. A 14-patch series from Ilpo Järvinen fixes the sizing, and an out-of-tree initramfs module pads each window to its alignment.