Mistral Large 4 is a state-of-the-art, open-weight, general-purpose multimodal model with a granular Mixture-of-Experts architecture. It features 49B active parameters and 1.05T total parameters, and a 1.6B vision encoder.
Mistral is launching a public preview of its ML4 model, a 1 trillion-parameter natively multimodal AI that achieves exceptional performance across various tasks. The model will be available for testing and feedback until the end of the month when weights are released.
openTPU is a learning project that brings auto-arch-tournament lessons to AI accelerators, testing how far agents can design hardware and build chips. It measures performance on various models with real weights on an Inspur YPCB-00338 card.
Polars 2.0 introduces significant features, including first-class SQL support and performance improvements that surpass DataFusion and DuckDB in TPC-H and TPC-DS1 benchmarks. The new version also includes out-of-core capabilities, improved scaling with more cores, and enhanced strictness for error handling.
The Nobel Prize in Physics 2026 was awarded to Francis Halzen for his contributions to the IceCube Neutrino Observatory. The award is announced live on nobelprize.org from October 5-12.
The author recounts the development of Smalltalk, a programming language that combined object-oriented design with interactive computing. The language was influenced by various ideas from the 1960s, including Sketchpad, Simula, and LISP.
The Tapo library has been updated to v0.11.1, adding support for TPAP protocol and allowing devices to communicate without the Third-Party Compatibility switch enabled. This change improves security by reducing the risk of password exposure during login attempts.
A micro benchmark should run for about 300ms to provide a human-perceptible range, allowing for easy comparison and iteration. This duration strikes a balance between precision and speed, making it suitable for intuition-based decision-making rather than precise measurement.
Meta's Muse AI product has faced numerous security and privacy issues since launch, including a zero-day flaw that allowed spying on Mac users. The company prioritized speed over security, leading to concerns about the potential for massive data breaches.
Gleam v1.19.0 has been released with significant improvements, including a rewritten Erlang code generator that outputs Erlang abstract forms, resulting in faster compilation times and improved accuracy of line numbers in BEAM crash reports. The new compiler also includes various bug fixes and enhancements for JavaScript support, type checking, and language server functionality.
Geothermal energy harnesses heat from the Earth's interior to produce electricity or provide domestic heating. The temperature increases with depth, making high-gradient areas prime locations for geothermal projects.
We’re checking your connection to prevent automated abuse
Researchers have developed a new algorithm to solve 3SUM and All-Pairs Shortest Paths problems in polynomial time, refuting long-standing hypotheses. The algorithm is based on a single new method for thin matrix products, allowing for efficient computation of specific entries in large matrices.
Beam is a large language model that combines pretraining and reinforcement learning to achieve strong performance in coding, agentic, and reasoning tasks. It was trained on 23.8 trillion high-quality tokens from the web and proprietary licensed datasets using a novel scaling recipe for reinforcement learning.
Deploy Parseable on a cloud for unified telemetry data, with features like SQL querying, dashboards, alerts and object storage-native data management. It offers full control and ownership of the data in open formats.
Darkplug allows precise control of an enlarger via a smart plug and iPhone. It offers advanced printing workflows with a red-on-black interface for darkroom use.
Fervo has reached a key commercial milestone by synchronizing its 100-megawatt power plant with the grid, ahead of schedule. The company aims to complete future blocks in as little as 18 months, appealing to data center operators seeking generating capacity.
Researchers have developed a zeroth-order optimization algorithm called Dust that is competitive with backpropagation at pretraining transformer language models. The algorithm perturbs activations independently at every token, allowing for parallel evaluation and efficient computation.
The route is computed using 1m lidar elevation and street graph data, balancing distance and climbing. The slider offers various options, excluding stairways on bikes while allowing them on foot.
The author has returned to using Node.js after a long period of using Deno, finding it more modern and stable. They've switched to PNPM for package management and are exploring other built-in Node APIs to improve performance.
You don't have permission to access "http://www.tdk.com/en/news_center/press/20261002_01.html" on this server.
The author uses a music-based accountability system to motivate themselves and their partner, finding it more effective than traditional bets due to its focus on joy and positive impact. This approach waters the carrot rather than magnifying the stick, providing a sense of purpose beyond just avoiding negative consequences.
This week in #curl we get to experience not one, but two compiler bugs. All research by @vsz. Two arm64-specific compiler optimization bugs, in gcc-15/16 and Rust respectively: https://github.com/curl/curl/issues/23237
Researchers used AI to design a new magnet with properties between ferromagnetic and antiferromagnetic materials, which could improve computer memory. The material has zero net spin moment but sorts electrons by energy level, allowing for faster switching and higher storage performance.
The article compares 12 Zigbee temperature and humidity sensors from various brands, including Aqara, Sonoff, Tuya, Thirdreality, Zemismart, and Moes. The sensors were tested for accuracy and reporting interval using a calibrated reference sensor in a controlled environment.
This book provides a thorough introduction to competitive programming, focusing on algorithms and mathematical thinking. It is suitable for students interested in IOI or ICPC, as well as anyone looking to learn about competitive programming.
The author enjoys building things, particularly software, which gives him satisfaction and joy. However, he prefers traditional coding over AI-assisted coding due to the lack of personal fulfillment and respect for his work.
The C++ standard proposal P4313R1 aims to add bitmask operations for enums using a C++26 annotation, but it has uncovered issues with integral promotion and boolean conversion. The code can produce unexpected results due to the language's behavior when performing bitwise operations on smaller-than-int types, which are promoted up to int before being combined.
The example.com domain has undergone several changes over the years, with updates aimed at reducing bandwidth demands and improving its utility. The latest redesign introduces multi-language support using JavaScript, showing content in a new language every 5 seconds.