> AI_ENGINEER.NEWS

OpenAI, US government, and Persona built a surveillance system with biometric databases, watchlists, and direct reports to authorities.

Firefox 148 introduces an AI kill switch, enhanced privacy, security features, and improved web platform capabilities.

AI rebuilt Next.js in a week, forming vinext—a faster, smaller Vite-based alternative deployable to Cloudflare Workers.

Steerling-8B is an interpretable AI model that traces generated tokens back to input, concepts, and training data, enabling precise control.

Moonshine models surpass WhisperLargeV3 in live speech accuracy with lower latency, optimized for real-time edge voice applications.

Mercury 2 is the fastest reasoning LLM, using diffusion for parallel response generation, enabling real-time AI in latency-sensitive applications.

OpenAI lower its 2030 spending goal from $1.4T to $600B, targeting $280B revenue amid expansion concerns and major strategic investments.

enveil secures project secrets by storing them encrypted locally, injecting real values into environment variables at runtime, avoiding plaintext leaks.

Hugging Face Skills define AI/ML tasks, enabling interoperability across major coding agents like Codex, Claude, Gemini, and Cursor.

Emdash is an open-source, provider-agnostic environment to run multiple coding agents locally or remotely over SSH.

Hegseth pressures Anthropic to drop AI safeguards by Friday; motive and implications unclear.

Speculative AI doomsday scenario causes US market plunge, highlighting potential economic upheaval, mass unemployment, and societal unrest.

xAI and Pentagon collaborate to integrate Grok AI into classified systems for enhanced security and intelligence operations.

Sam Altman suggests human life is as wasteful as AI data centers, emphasizing efficiency concerns and resource use.

NIST seeks public input on AI security; access to federal site data via APIs is restricted, with CAPTCHA verification for human users.

AI Functions (3.1) enhances prompt-driven AI code with runtime execution, native objects, validation, and feedback loops for reliable, dynamic AI-powered software.

Proposal to add "AI generated" as a flag to help identify, moderate, and educate about AI-produced content and reduce spam.

Microsoft execs warn AI may reduce entry-level coding jobs, urging mentorship and training to preserve future skills.

Turning off ChatGPT's memory avoids context rot, bias, and confusion, offering more control and unbiased results.

Software stocks rebounded after Anthropic announced new enterprise AI partnerships, easing fears of AI disrupting existing workflows.

Firefox 148 offers AI controls and kill switch, allowing users to disable AI features like translation, chatbots, and link previews.

Qwen 3.5 is a 35B multimodal model with enhanced reasoning, multilingual support, and scalable inference, available via APIs and open-source frameworks.

Frontier LLMs show potential in CAD design, with Claude Opus 4.6 achieving near-perfect grip tests, but often produce imperfect, non-usable solutions.

Anthropic updates its Responsible Scaling Policy to Version 3.0.

Cursor cloud agents now control their own virtual machines, enabling autonomous code building, testing, and merging across multiple platforms.

Cybercriminals used AI tools to compromise 600+ firewalls globally, automating attacks and attack planning with minimal skills.

Hegseth threatens to cut Anthropic from Pentagon supply chain amid a showdown with its CEO.

AI skepticism functions as a moral class boundary, reinforcing hierarchy through notions of authenticity, effort, and purity amidst technological change.

AI productivity gains are delayed; true efficiency improvements haven't yet materialized.

Open-source GPT-based model predicts golf outcomes; trained automatically, beats GPT-5, adaptable to other domains.

AgentBudget enforces real-time dollar budgets on AI sessions, preventing overspending via cost tracking, limits, and automatic halts.

Bareclaw routes Telegram messages to Claude code on Mac, enabling self-hosted AI interactions via phone, with minimal code and security.

Autonomous browser agents streamline healthcare prior authorization, scaling 7x, ensuring security, reliability, and human-AI collaboration.

Implements multi-LLM routing to manage Claude's code quota limits effectively.

Claude enhances government AI with secure, compliant, high-performance models for national security, regulatory review, and official archives.

Noodles visualizes AI-generated codebases with interactive diagrams, making code flow easier to understand without reading every line.

Researchers trained an open-weight VAE for image-video compression, revealing that better reconstruction isn't necessarily better for diffusion-based generation.

Hegseth threatens to blacklist Anthropic over safety standards and 'woke AI' stance, risking a $200M defense contract and government use.

Qwen 3.5 Medium Model Series announcement; details on signing up via Google or Apple.

AI developers often build the wrong solutions due to shared misconceptions and misaligned priorities.

Global regulators warn AI image tools must follow privacy laws to prevent harms like non-consensual and harmful content.

Even with metacognition, AI societies in simulated capitalism naturally form bubbles through herding and knowledge transfer.

A terminal game simulating AI-driven survival amid rapid AI growth, emphasizing choices, trade-offs, and replayability.