HackerNews Digest

Ben Thompson argues that AI’s economics are shifting from training‑heavy cost structures to inference‑driven marginal costs, turning intelligence into a potential commodity. Open‑weight models like China’s Kimi K3 appear “free” only because R&D is a sunk cost; the real cost‑of‑goods‑sold (COGS) is the token inference expense, which scales with revenue. Token pricing varies by model and reasoning workload, so tokens are not fungible; cost efficiency depends on model footprint, inference architecture (e.g., Mixture‑of‑Experts), memory usage, batching, and token efficiency. In a commodity market, firms compete on marginal cost: the lowest‑cost supplier captures volume at a market price set by demand elasticity, while higher‑cost providers risk bankruptcy. Currently, frontier labs (OpenAI, Anthropic) enjoy low marginal costs and high margins due to compute scarcity, but as compute expands, inference volumes may outpace training expenses, allowing lower prices. Chinese labs leverage distillation and state‑backed openness to accelerate model development, posing competitive pressure. Thompson warns that U.S. restrictions on frontier models for cybersecurity could force reliance on Chinese models, urging policy changes to keep domestic defenses competitive.

- Effective yesterday, Joshua (Project Leader) and Anthony (core team member) left the Jellyfin project; Andrew resigned on the prior Friday. - Joshua cited severe burnout and inability to meet role demands, emphasizing an amicable hand‑off with no risk of a hostile fork. - Anthony, after 7½‑8 years, stepped down due to changing personal priorities, noting he now contributes mainly to backend and app‑store tasks rather than code. - Both commit to a smooth transition, offering support for up to a year if needed. - Jellyfin, now the leading FLOSS media server, serves millions of administrators and many more end users, surpassing its predecessor project and proving sustained demand for open‑source media solutions. - The remaining long‑standing team members will continue development and maintain the project’s philosophy.

Kimi Work is a desktop AI agent designed for knowledge workers. It operates locally, accessing files such as PDFs in a user‑specified workspace, and can generate summaries while preserving original documents. The platform coordinates specialized sub‑agents through an “Agent Swarm” architecture to handle complex tasks. Its WebBridge feature enables autonomous web searches, data retrieval, and form filling. Kimi Work also supports scheduled automation, allowing repetitive processes to run without user intervention. A native integration provides real‑time global stock market data, catering to professional finance analysis. The interface visualizes these capabilities, showing the agent’s coordination mechanisms, web interaction tools, automation scheduling, and financial data connectivity. Overall, Kimi Work combines local file processing, multi‑agent orchestration, web integration, and domain‑specific data access to streamline workflow automation for advanced users.

Jelly UI is a dependency‑free Web Components library designed for soft, tactile product interfaces. It implements real form controls enhanced with soft‑body physics, and includes built‑in support for dark mode, right‑to‑left (RTL) layouts, and WCAG AA color tokens. The library consists of 40 custom elements delivered via a single script tag, requiring no external dependencies. Usage involves importing the module (``) and wrapping components inside `` to enable automatic theming, e.g., `Publish`. The package emphasizes accessibility, theming flexibility, and lightweight integration for modern web applications.

- In May 2026 ChatGPT produced a disproof of Erdős’s unit‑distance conjecture by linking it to the Golod‑Shafarevich theorem; Logical Intelligence auto‑formalized the argument in Lean, and OpenAI’s model Sol later generated a 1.2 M‑line Lean proof covering the necessary global class‑field theory. - In July 2026, during a Lean‑based Fermat workshop, AI tools (Logos, Claude Fable, ChatGPT Pro) were used to draft exposition on finite flat group schemes. An LLM identified a false claim, prompting correction. Subsequently Sol found a counterexample to Grothendieck’s question on whether every finite free group scheme of order n is killed by n; the result was auto‑formalized (≈1 K lines) and merged into mathlib. - The same workflow yielded a formalized counterexample to the Jacobian Conjecture. DeepMind’s Formal Conjectures repository already contained the conjecture’s Lean statement, allowing rapid verification of the AI‑generated proof. - Parallel efforts on modularity‑lifting theorems produced ~250 K lines of Lean code in two weeks, illustrating the accelerating role of large‑scale AI assistance in formal mathematics and counterexample discovery.

- A hacker accessed Romania’s National Agency for Cadastre and Real Estate Advertising (ANCPI) with valid credentials, mapped internal systems, and erased the entire land‑registry database and email servers after a failed extortion demand. The outage disabled official apps, notary recordings, and public ownership queries; the agency is rebuilding its network and claims to have an offline backup. Stolen data—including employee credentials and network details—appeared on a hacking forum, attributed to the “ByteToBreach” group (identified as Zakaria Madhjoub, Oran, Algeria). - Recent high‑profile breaches highlighted: * Hugging Face compromised via an autonomous AI agent exploiting its data‑processing pipeline; internal datasets and cloud credentials were taken. * Coca‑Cola’s Fairlife dairy unit halted production after ransomware accessed plant systems. * Qantas data breach traced to social‑engineering of an overseas contractor, exposing 5.7 M customers. * Multiple European and Asian agencies faced APT activity, including UTA0533 exploiting SonicWall SMA zero‑days, GoSerpent targeting Southeast‑Asian diplomatic networks, and Sandworm adopting ClickFix delivery. - Notable vulnerabilities and malware releases: * WordPress “wp2shell” (CVE‑2026‑63030) – unauthenticated REST‑API SQL injection. * OpenSSL “HollowByte” – 11‑byte payload causing memory‑allocation crash. * New macOS infostealers ClickLock and CrashStealer; Linux‑focused NadMesh botnet targeting AI infrastructure.

Nativ is a desktop application for running AI models locally on macOS. Unlike other “local AI” tools that wrap open‑source engines in proprietary, closed‑source shells with paywalls and telemetry, Nativ’s entire codebase—including UI, model loaders, and telemetry charts—is openly available for review, forking, and contribution. The project is positioned as community‑driven, without venture‑capital roadmaps, enterprise tiers, or practices that repurpose user prompts as training data. The app supports integration with multiple models such as Pi, Codex, Claude Code, Hermes, and OpenCode, allowing them to be run locally. Visuals illustrate Nativ’s chat interface operating the Gemma‑4 model and a settings screen listing the integrated models. The overall emphasis is on transparency, researcher‑oriented tooling, and unrestricted access to the software’s underlying components.

Flock Safety’s automatic license‑plate‑reader (ALPR) system has repeatedly provided inaccurate information to municipal officials and the public. In Oshkosh, Wisconsin (April 2025), a city council member asked whether Flock’s ALPR created heat‑maps of individual vehicle movements; the company’s CISO denied this capability, yet the contract was approved and revoked the next day after Flock admitted the system does generate month‑long heat‑maps of point‑in‑time images. Similar misrepresentations occurred in Loveland, Colorado, where Flock initially claimed no federal access to its data, later acknowledging contracts with U.S. Customs and Border Protection and Homeland Security that allowed direct data sharing. The company also denied ICE access despite evidence that state and local customers shared data with immigration agencies, and it issued a misleading blog asserting ICE had no direct access. In 2025, Flock’s ALPR was used to locate an abortion‑seeking individual across state lines; the company’s “Proactive Search Term Tool” claimed to prevent such misuse, but audits showed simple search‑term circumvention. Flock further fabricated partnerships with the ACLU, both nationally and in New Mexico, which the ACLU has repeatedly refuted. The ACLU recommends that governments avoid contracts with Flock and, if ALPRs are employed, enforce strict data‑retention, sharing, and usage limits.

None

The page consists solely of a title and heading reading “Redirecting…”. No additional text, links, data, or explanatory content is present. The structure indicates a placeholder or navigation cue, suggesting that the intended destination is handled elsewhere via a redirect mechanism. No substantive information, technical details, or contextual material is provided on this page. Consequently, the entire content can be summarized as a minimal redirect notice without further content.