Factory CEO Matan Grinberg accused VC advisor Chris Degnan of spying for competitor Cognition, claiming Degnan shared confidential information while attending board meetings. Degnan disputed the allegations, saying he resigned rather than being fired and that he hadn't shared confidential information, while simultaneously joining Cognition as chief revenue officer.
This paper introduces agentic meta-reasoning, an inference-time framework where a controller makes explicit decisions about task execution while workers perform computations. Tested on multiple benchmarks including program reconstruction, the approach achieves significant improvements over direct control baselines, with gains of 3.6–4.2 points across frontier models and better performance at larger compute budgets.
Kids are using NPR podcast comments on Spotify as an unmoderated chat room to evade social media restrictions. NPR staff initially mistook the comments for bot activity before recognizing the behavior as children communicating with each other using coded usernames and casual language patterns.
Apple patched CVE-2026-86950, a unit conversion error in CoreGraphics triggered by malicious PDFs with crafted fonts, causing out-of-bounds memory writes. The vulnerability was exploited in the wild targeting specific individuals and may have a zero-click WhatsApp attack vector. Meta researchers discovered the bug through patch analysis and found WhatsApp added strict PDF validation to detect malformed fonts.
The Hutter Prize challenges competitors to compress the first 1GB of Wikipedia (enwik9) better than previous records, with prizes up to €500,000. The competition aims to encourage development of intelligent compression algorithms as a pathway toward artificial general intelligence, based on the principle that superior compression relates to intelligence.
California's Court of Appeals upheld dismissal of a former Premise Data CEO's lawsuit against journalist Jack Poulson, who reported on the CEO's sealed arrest for domestic violence. The court affirmed that the First Amendment protects lawfully obtained truthful publication and rejected the CEO's appeal, citing California's anti-SLAPP statute and Shield Law.
A Hacker News post describes a creative audio/log player interface concept for consuming content about OpenAI DevDay 2026, presented as a fictional narrative format with scene-setting and transcript features.
Testbox is a fast, zero-dependency test runner written in Lua that supports unit testing and API testing. It can be installed via Cargo or downloaded as a compiled executable, and includes batteries for executing background processes and making HTTP requests.
Turbopuffer, a serverless vector database, is transitioning from a vector-index-centric architecture to a generalized search database with a new storage engine. The company is moving the ANN vector index from primary to secondary status to support diverse query plans at greater scale, evolving beyond its original specialized use case.
CSS, the styling language behind web design, deserves serious recognition as a powerful programming tool, not dismissal as merely decorative. Despite its widespread use and sophisticated modern capabilities, CSS faces cultural disrespect partly due to gender bias and historical gatekeeping debates over whether it qualifies as a "real" programming language.
A service for verifying cold-chain shipment data using signed logger CSVs with offline verification capabilities. Users can purchase credits via card, bank transfer, or BSV cryptocurrency to access BSVKey services that cryptographically sign results for independent verification.
RexaDB is a modern database studio that supports 100+ databases via JDBC, offering features like multi-tab SQL editing, interactive ER diagrams, fast data browsing with filtering, and AI-powered SQL generation in plain English. Built to provide a seamless experience across query writing, schema visualization, and data inspection.
Growth.engineer is a collection of open-source, vetted workflows for AI agents that automate go-to-market tasks like lead generation, content creation, customer analysis, and sales outreach by integrating tools like Claude, HubSpot, Slack, and Notion.
An open source project must choose between accepting or rejecting AI-generated code contributions, as middle-ground policies are unsustainable and effectively favor AI adoption. While accepting AI may increase contributions, it risks alienating community members and creates dependence on proprietary tools; rejecting AI avoids these drawbacks and won't drive away users, only contributors unwilling to work without LLMs.
Benchmarking study comparing Java performance across OpenJDK versions 8-27 and alternative JVM implementations including GraalVM, Eclipse Temurin, and IBM Semeru on AMD Ryzen Threadripper hardware running Fedora Workstation.
A JavaScript/TypeScript library inspired by Emacs' rx macro provides a readable DSL for building regular expressions as named trees instead of opaque strings. The approach improves maintainability by eliminating manual escaping and making complex patterns like semantic versioning validation self-documenting.
Google's DeepMind team has developed a protein watermarking system that marks AI-designed proteins without affecting their function, enabling identification of proteins created by trusted researchers. The technology, based on Google's SynthID system, addresses biosecurity concerns by allowing detection of potentially dangerous AI-generated proteins that current DNA screening tools cannot identify.
EDG C/C++ compiler is now open source, featuring a front end known for parsing compatibility with Clang, GCC, and MSVC, along with C and C++ generating back ends, template instantiation handling, and development tools. The project includes extensive documentation and supports source-to-source transformation applications.
A GitHub issue proposes adding a configurable indent level option to Black, the Python code formatter, to support 2-space indentation used at Twitter and other organizations. The requester argues that indent level is as fundamental as line length and should be configurable, similar to the existing --line-length flag.
Researchers from University of Washington, Meta, and MIT introduce Context Language Models (CLMs), which allow language models to manage their own context by treating it as an updatable file. The approach achieves state-of-the-art performance across multiple benchmarks with significantly reduced computational costs, and supports multi-agent systems and reinforcement learning optimization.