Mistral states that input and output data may be used to train its models, but users may opt out at any time. Opt-out defaults vary by product: Vibe is not opted out by default, while Vibe Enterprise is opted out by default and managed at the admin level; opt-out for Mistral Studio and API data training is available via the Admin panel with separate toggles. Documents uploaded within Vibe are treated as input data for training purposes.
Trending
A web tool lets you upload a file to check for signs that Claude helped create or edit it by reading a C2PA content credential embedded in the file. If a credential points to Claude, the tool reports it; a missing signal does not prove Claude was not involved, since signals can be removed or unsupported by some formats. The service runs in your browser, supports many formats up to 100 MB, and your file stays on your device while only the embedded credential is read.
The writer describes initial skepticism about smart telescopes due to cost and the live-observation experience, then buys a used Seestar S50 and finds it surprisingly capable for imaging. It provides auto-tracking and automatic target acquisition, and it allows access to raw exposures that can be processed with standard tools, yielding more detail than the app’s built-in processing. While acknowledging limits from a 50 mm aperture, the author notes it’s portable and practical for travel and for beginners.
New brain-imaging research shows aging changes how memories are recalled. Among older adults, stronger similarity between learning and recall brain patterns in the hippocampus is linked to cross-category errors—confusing a face paired with an object with a scene—rather than to better recall. The study, with 61 participants aged 18 to 74, found that hippocampal shrinkage or baseline organization only partially explain the pattern, and attention differences did not account for the errors; the authors say the memory system shifts from precise replay to broader misbinding, though the exact cause remains open.
Gemini 3.8 Flash is the next version in the Gemini 3 model family, building on Gemini 3.7 Flash with improvements for software engineering and knowledge workflows. It supports multimodal inputs (text, images, audio, and video) with a 1-million-token context and can produce up to 64,000 tokens of output. The model's evaluation suite covers coding, knowledge work, and multimodal tasks, and safety assessments note ongoing limitations such as potential hallucinations and occasional slowness, with frontier safety reviews indicating no new major capabilities versus 3.7 Flash.
The author announces they are ending their Elixir exploratory writing, a log started 3.5 years ago to explore Elixir and the Erlang VM. They say they will no longer publicly share exploratory programming work, citing concerns about machine-generated code and impersonation in the community, and will instead pursue other subjects and non-experimental Elixir projects.
Quasar 438B is Multiverse Computing’s flagship 400B+-parameter reasoning model for enterprise-scale tasks, available through the CompactifAI API. In the Artificial Analysis Intelligence Index v4.1.1, it scores 43, the highest among European models, and it produces a 500-token response in 15.3 seconds, faster than most peers. It outperforms several competitors on long-context reasoning and terminal tasks but trails Claude Opus 5 on the index; it supports English and Spanish for enterprise use.
The article argues that, when accounting for energy, labor, and production costs, landfilling can be more environmentally sensible than recycling for many materials. It notes that the benefits of recycling vary by material—metals show large energy savings, while plastics and other items often incur high processing costs and contamination reduces effectiveness. The piece cites energy comparisons for bags, textiles, and cups to illustrate a nuanced view rather than a blanket endorsement of recycling.
A study released Sept. 2, 2026 found that three sites generated 215,128 machine-generated 'best software' pages across 380 categories to ground AI recommendations, yielding 7,534 citations across 2,055 domains. Of these citations, 59.8% point to domains ranked worse than #100,000 on the Tranco top-1M and 23.4% to domains not in the top million, with many unranked domains being newer. The sites—worldmetrics.org, gitnux.org, and wifitalents.com—share Cloudflare hosting and a common page template with 'Facts & Grounding Page' titles, producing the bulk of the grounding content.
The piece argues that properly managed landfill can be cheaper and less energy-intensive than recycling for many common wastes, challenging the idea that recycling is always the better option. It lays out calculations showing energy costs vary by material—for example, plastics and cotton often require substantial energy, while metal recycling offers large energy savings—along with issues like contamination that can undermine recycling efficiency.
LWN said subscription prices will increase by about 20% starting September 15, with monthly rates listed as: Starving hacker $6, Professional hacker $11, Project leader $19, and Maniacal supporter $55; group subscriptions rise by the same amount. Existing subscriptions will remain valid through their current expiration dates, and active monthly subscriptions before the change will be charged at the old rate for six months. The note cites inflation and rising costs since the 2022 change and notes LWN will mark its 30th year in January.
The Nederlandsche Bank moved about 86 tons of gold from New York and Ottawa to London between March and August 2026, giving London 32.1% of the Dutch gold reserves. The total reserve is about 612.4 tons, valued at around €72.2 billion, with New York and Ottawa now holding 18.5% each and Zeist 30.8%, as the bank says the move aims to improve crisis readiness and diversify risk.
The article shows how to use jq to format JSON copied to the clipboard, with macOS and Linux examples using pbpaste/pbcopy or xclip. It notes a problem where non-JSON clipboard content can cause jq to exit with an error and overwrite the clipboard. A revised alias provides error handling by attempting jq and, if it fails, restoring the original clipboard content before writing back.
Uber announced a set of organizational changes that will reduce its workforce by about 10%. The company says the moves aim to simplify the organization, cut management layers (including a 20% reduction in roles seven or more layers from the CEO) and reduce micro-teams by about 50%, and to implement a hub-based location strategy with most remote staff moving to office hubs. It is consolidating teams, such as combining the three Delivery Ops units and merging Core Services Engineering with Science, to improve decision-making and capital allocation, and stresses the changes reflect organizational priorities rather than individual performance.
The piece explains that Microsoft Entra Passkeys provide phishing-resistant, passwordless authentication based on the WebAuthn standard, with the private key stored on the user’s device and the public key in Entra ID. It covers device-bound and synced passkey models, configuration steps in Entra, end-user enrollment, and how to enforce passkey usage through authentication strengths and Conditional Access policies, noting upcoming broader support in March 2026.
Humanoid robots have shown rapid progress with demonstrations like robotic bartending, synchronized dances, and even a fast 100-meter run, but experts say such feats do not translate into reliable, broad-capability workers. The biggest hurdles are manipulation of real-world objects, generalization across varied tasks, and the ability to carry out long, multi-step projects, with many performances relying on teleoperation rather than autonomous control. Experts warn that solving these challenges for safe, affordable, real-world deployment will take many years, if not decades.
AISLE reported six curl CVEs after OpenAI Codex Security and Anthropic Mythos announced zero results. The six CVEs, tied to curl 8.22.0, involve OpenSSL and wolfSSL components and were fixed in that release, with AISLE credited Stanislav Fort as reporter. The timeline noted a zero-result from frontier AI systems on August 24, followed by AISLE’s findings and a rise in pending CVEs through August 28.
HomeOS is a self-hosted family operating system designed for a kitchen touchscreen. It centralizes household tasks, shopping lists, notes, prayer times, and birthdays on a single shared display, and runs on a Linux mini PC with Docker-based deployment. The project uses a React/TypeScript frontend, a Python FastAPI backend, SQLite for storage, and is released under the MIT license, with Quick Start and Documentation available.
Submarine fiber-optic cables carry most international internet traffic and lie on the seabed with repeaters powered by copper cores every 50 to 80 kilometres to sustain the signal. A handful of firms manufacture and lay these cables after route surveys and permitting, with transoceanic links costing hundreds of millions and a typical design life of 20–25 years, and repair is done by dedicated ships when faults occur. Ownership of the cables is distributed among many co-owners, including major telecoms and tech companies such as Google, Orange, BT, Sparkle, Vodafone, and Meta.
Justin Andersun argues that retreating into a private 'Cave' of constant training can create a misleading sense of progress, and that real-world success requires moving beyond that private grind. He recounts years wrestling at Kenowa Hills High School, including weight cuts and the brutal discipline of the mat, and concludes that technique, conditioning, and heart matter, with victory pursued openly for the benefit of the team. The piece encourages taking risks in the real world—shipping products, sharing stories, and embracing competition.
WebLLM is a high-performance in-browser language model inference engine that runs entirely in web browsers with WebGPU acceleration. It provides OpenAI API compatibility, supports streaming and JSON-mode generation, and offers multiple model families such as Llama, Phi, Gemma, Mistral, and Qwen, installable via npm, Yarn, or CDN.
A study audited Perplexity’s citations across 310 questions about 210 tech companies, examining 1,826 figure-related citations, and found that 34.7% pointed to pages that wouldn’t open or did not contain the cited numbers; 14.4% of 872 numeric claims failed when tested per marker. Across 2,915 cited URLs, 84.2% of sonar’s answers cited at least one URL ordinary readers could not open and 10.6% cited at least one that was dead, with many failures tied to gated pages or pages that do not state the claim. Perplexity’s sonar and sonar-pro yielded similar pass rates on numeric pairs (about 65–66%), while GPT-4.1 with a web plugin used fewer inline markers and relied more on the company’s own domain, making sentence-by-sentence auditing more difficult.
An author notes that 9000 RPM equals 150 revolutions per second, a rate faster than typical screen refreshes. He created an online animation showing four engine layouts—inline-4, 60-degree V6, boxer-6, and flat-plane V8—and explains how, at standard 60fps, the crankshaft appears as blur, while slowing to 1/50x lets a single revolution be seen step by step. The post promotes the PWAs tool collection and invites readers to try the animation.
Rohan, a 26-year-old Londoner, has experienced back pain since age 13. He spent three years building a wearable—described as a tank-top device—to track everyday movement in an effort to understand his pain. He says back-pain care lacks diagnostics and has shared project details and contact information.
Bridson’s 2007 algorithm partitions the space into a grid with cell size r^d to ensure at most one point per cell and builds samples by selecting points from an active list and drawing points from an annulus around each, up to k=30 attempts, with grid-based collision checks for efficiency. The article describes improvements including a 2D parental optimization to constrain sampling angles and a parameter c that adjusts the radius distribution to balance density and randomness, with empirical guidance such as c = -1.4 - 17/k. It also mentions applications like stippling and GPU-accelerated parallel implementations such as PixelPie and Poisson Cam.