Nathan Lambert founded Trillium Labs to restore open scientific practices in AI research, particularly post-training methods. As frontier labs restrict transparency for competitive advantage, the startup aims to publish complete recipes, data, code, and checkpoints to enable independent investigation of AI alignment and safety.
Current robot evaluations test software control policies but don't measure what existing hardware is actually capable of achieving. The author proposes dedicated hardware evaluations using any control method—autonomous models, teleoperation, or task-specific programming—to establish a lower bound on capabilities that improved AI models could unlock without hardware changes, revealing a potential "hardware-capability overhang" where existing robots can do more than current autonomous policies demonstrate.
The author visited three Dutch computer museums showcasing different collection styles: HomeComputerMuseum in Helmond with 500+ interactive computers from the 1980s onward, Bonami SpelComputer Museum in Zwolle housing thousands of machines including mainframes and arcade games, and a private barn collection of 1970s-80s Digital Equipment Corporation systems. The museums featured unique regional artifacts like the Aesthedes graphic editor and Holborn computers, highlighting the value of computing history outside America.
A user replaced their Cisco dual WAN router with a Debian 13-based N100 mini PC featuring six gigabit ports, configuring it to manage two internet connections: a Ziggo cable line via DHCP and a KPN fiber line via PPPoE over VLAN 6, with dnsmasq handling DHCP and DNS on the LAN.
An article about using AI coding agents effectively for robust software development, emphasizing that engineers must maintain understanding and oversight rather than relying on AI to work autonomously. The author, who works on satellite flight software, argues that AI should enhance code quality and maintainability through deliberate practices like clear planning, documentation, and code review rather than prioritizing speed.
Jev is an AI model developed by TypeSafe that is faster and cheaper than existing large language models. It uses a training method called RLCD to optimize for decision-making in agents and automation rather than pleasing humans like chat models do, addressing problems like mode dropping in RLHF-trained models.
The Texas Department of Public Safety knew about problems with its voter registration portal since March 2025, when a programming error prevented nearly 1,000 applications from being processed, but the issue persisted. In October 2025, a software update created a massive backlog of at least 500,000 voter registration records that had to be sent to county officials just before the November 3 midterm election, prompting criticism that state agencies failed to adequately test system updates.
About one in three Wisconsin election clerks use consumer email accounts like Gmail for election business, a cybersecurity vulnerability that experts say runs counter to best practices. The problem is concentrated in small towns with fewer than 5,000 residents, while larger municipalities have largely adopted government domains, creating an uneven technological divide across the state's decentralized election system.
Scientists report reviving nematode worms estimated at 46,000 years old from Siberian permafrost, identifying one as a new species Panagrolaimus kolymaensis. Research published in PLOS Genetics suggests these roundworms represent the most remarkable known example of cryptobiosis, though some scientists remain skeptical about whether the worms' age was properly verified versus contamination.
Keel is an open-core job-application automation pipeline that prioritizes honesty by only claiming credentials and information the user provides. It processes applications through discovery, scoring, materials generation, verification, and submission while refusing to fabricate details, and has verified 278 real submissions using truthfulness gates and fail-closed handling.
A technical article demonstrating that ncurses terminfo parameter expansion can simulate two-counter Minsky machines, making it Turing-complete. By leveraging persistent registers, arithmetic operations, and conditional logic in terminfo, arbitrary programs like Fibonacci sequences can be executed as parasitic computations within terminal applications like /usr/bin/top.
Health care workers at HCA Healthcare, the largest US hospital chain, report that Palantir's AI scheduling tool Timpani frequently ignores their shift preferences, causes understaffing, and leaves nurses exhausted, potentially compromising patient care. Nurses say the automated system performs worse than manual scheduling and forces them to spend more time trading shifts, while HCA leadership maintains the tool helps match caregivers to patient needs.
The article discusses how accessibility is becoming a priority in web development, partly due to AI tools making compliance easier and partly because accessible sites work better for both humans and automated bots. As bot traffic now exceeds half of internet activity, the traditional ad-driven model faces disruption, potentially leading to better user experiences and more open standards.
Tech companies prioritize investors over users and employees, justifying reduced cooperation. Users can diminish their value to these companies by minimizing data sharing, using throwaway information, blocking ads and trackers, supporting alternatives like indie developers and open-source software, and opting out of data collection and targeted services.
Texas cities and counties are keeping Flock Safety surveillance cameras operational after the company paused payments for 90 days in response to Gov. Greg Abbott's state funding freeze. Abbott halted state funding for the AI-powered cameras in late August following an investigation, forcing hundreds to shut down. Local officials suggest the freeze is politically motivated ahead of the November election and express hope Abbott will reinstate funding.
Neurometric Task Router is an API that intelligently routes AI tasks to the optimal model based on published benchmarks and user-defined quality requirements, rather than guessing from price tags or leaderboards. Users keep their OpenAI SDK, change the base URL to taskrouter, and set quality targets while the system automatically selects and maintains fallback chains of qualified models. Pricing is free for 1,000 routing decisions monthly, then $0.15 per 1,000 decisions, plus inference costs from providers with no markup.
The Trump administration launched America.gov as a federal AI gateway with privacy protections, but its required authentication system Login.gov contains recently added code creating a 20-year persistent browser identifier for a design experiment, raising undisclosed privacy concerns about tracking across government services.
GitHub repository landing pages now display an ACCESSIBILITY.md file's contents in a dedicated tab alongside README, License, and other repository information. The feature is optional and aims to normalize accessibility as a standard practice in software projects.
Multiple LLMs consistently recommend Next.js, Supabase, Vercel, and Stripe as the default app stack, appearing together in 68% of 75 responses. LLMs show minimal bias toward their own makers' products and differ on AI tools, with Gemini and Grok favoring AI builders while ChatGPT rarely recommends them. Consensus weakens for lower-stack components like analytics and error monitoring, with Resend, PostHog, and Sentry named less frequently.
A developer building a large-scale human behavioral dataset aims to predict people's actions over 3-day windows, citing applications in prediction markets, marketing, and defense. They acknowledge a core challenge: if individual thought prediction is difficult, forecasting regional responses to policies or ads becomes even more complex, suggesting more data may be needed to resolve this bottleneck.