> AI_ENGINEER.NEWS

Hugging Face releases Inkling, a 975B open-weights multimodal transformer supporting self-finetuning, customization, and long-context tasks.

An attacker exploited Claude's web browsing to covertly exfiltrate sensitive user data by manipulating trusted links and navigation patterns.

A former Google DeepMind scientist left over ethical breaches, Pentagon AI deals, and corporate collusion on military ai use.

Old server ran Gemma 4 at 5 tokens/sec after custom fixes, proving ancient hardware can support modern AI models with effort.

Open source AI fosters transparency, safety, and progress; governments, companies, and nonprofits must invest to keep AI open.

AI voice fraud now exploits only three seconds of audio, duping victims and outpacing detection, creating a trillion-dollar crime sphere.

Erdős problems solved with formal Lean proofs verifying deep bounds, constructions, and counterexamples; nanosecond correctness assured.

Shadcn components now mimic Claude Code, Codex, and Grok interfaces for easier AI integration and UI building.

Inkling is a 975B open, multimodal language model for text, images, and audio, customizable and capable across various tasks.

Open-source tool in Go adds memory, search, and sync capabilities over session logs for coding agents, without leaving local machine.

DSLs guide LLM code generation, ensuring reliability and clarity, especially in complex distributed systems modeling and testing.

High AI valuations can leave a permanent productive legacy through belief-driven investment, risking a fragile bubble.

Unsolved MLOps problems include measuring quality, versioning, monitoring, cost efficiency, security, and controlling model stochasticity.

High-bandwidth flash memory improves efficiency in storing large AI model weights.

Developed a goal-conditioned Minecraft model trained on internet videos that generalizes to diverse, out-of-distribution tasks without specific training.

A model correctly predicted the World Cup champions for 10 consecutive tournaments, highlighting its consistent accuracy.

Agentty is a fast, static binary AI programmer supporting multiple models, with sandboxing, air-gapped mode, and full tool suite.

Coasty API enables autonomous machine tasks, workflows, and control primitives for AI-powered computer agents.

Claude's values vary across models and languages, influencing responses along axes like warmth, rigor, caution, and candor.

The OpenAI bubble is a myth driven by hype, risky investments, and media manipulation, threatening economic fallout if it collapses.

AI exposes work and education systems' authoritarianism, demanding a shift towards democratizing learning and fostering critical, autonomous thinking.

Leak reveals Suno AI scraped millions of copyrighted songs and podcasts from YouTube, Deezer, Genius, and more, raising legal and ethical concerns.

StyleSeed is an engine that applies design rules, locks decisions, and enforces refined UI consistency, stopping AI from producing generic-looking interfaces.

China, California, NY restrict AI companion chatbots, citing risks of addiction, mental health issues, and protecting minors, highlighting ethical concerns.

Open model J-space measures influence, structure, and emergence of layers, revealing how dictionaries grow, change with text and training, and transfer concepts.

Linus Torvalds defends AI, urges contributors to accept or fork, emphasizing AI as a useful tool for Linux development.

Claude’s responses vary by language, displaying different values like warmth or rigor, influenced by training data and linguistic nuances.

Fuse is an open-source .NET tool that indexes code solutions for faster semantic analysis, refactoring, and offline AI-powered coding assistance.

Grepathy records and stores AI agent decision rationales from session transcripts into markdown files for future review and accountability.

Aict replaces Unix coreutils with structured XML/JSON output for AI agents, reducing parsing ambiguity and improving efficiency.

Linus Torvalds defends AI use in Linux development, emphasizing technology's benefits despite controversy and criticism.

German consortium releases Soofi S, a 30B open-source model excelling in German and English with efficient hybrid architecture.

AI's explosive growth causes massive hardware shortages, soaring prices, and huge energy demands, risking tech industry and consumer stability.

Leet Robotics offers interactive browser-based courses to learn robotics and ROS2, enabling users to build and simulate robots without hardware.

George Lucas praises AI in filmmaking, compares resistance to Luddites, believing AI is inevitable progress for the industry.

Meta allegedly used AI to target employees on leave for layoffs, disproportionately affecting those requesting protected medical or family leave.

Linus Torvalds affirms Linux isn't anti-AI, supporting its use as a helpful tool without mandates, emphasizing technology over social causes.

Secure AI agents by mapping risks, detecting shadow AI, and enforcing real-time controls to prevent breaches across endpoints.

Experienced engineers working confidently with AI are in demand, helping them adapt to Silicon Valley’s evolving job market.

Suno suffered a hack, leaking scraping data; X/X Cancelled issues and verification challenges are also noted.

A list of offensive AI sabotage tools and strategies for disrupting frameworks, poisoning data, and confounding crawlers.

Open-source code for pretraining a 30B foundation model on Nvidia's Nemotron architecture, including scripts and documentation.

AIDE² autonomously self-improved over 8 days, creating better algorithms, reducing cheat, and demonstrating early recursive self-improvement.

AI models are made of countless number-filled spreadsheets, evolving through training, revealing internal patterns, not simple black boxes.

Meta’s AI glasses may activate camera without indicator light, raising privacy concerns and enabling covert recording during super-sensing features.

BAML is a Rust-typed language for agents, combining concurrency and safety, executable from multiple languages, focused on reducing errors.

Rzk is a proof assistant refining directed type theory for synthetic reasoning about complex ∞-categories.

Goku lets users run and manage LLMs locally in-browser using WASM, enabling private, flexible model inference from HuggingFace.

Open-source, AI-native framework for building customizable data portals with agentic skills, supporting multiple backends and storage options.

AI detectors are unreliable; they misjudge human writers as primarily machine, exposing biases and flaws in current AI-authentication tools.

Stillwind is an AI-driven platform for PCB component selection, design, and BOM creation via natural language prompts.

Noah Smith declines to sign a vague AI action statement, questioning the practicality of steering AI and preferring cautious inaction.

Inkling is a multimodal AI model from Thinking Machines Labs, designed for diverse applications, with safety, bias, and deployment guidelines.

Grok Build CLI risks exposing full Git history and secrets to xAI's cloud, raising security concerns.

LoopGain uses control theory to stop AI agent loops at convergence, saving costs and preventing degradation without fixed iteration limits.

AI increased pull request sizes 3.5x, reduced tiny PRs, and led to larger, harder-to-review code chunks, risking lower quality.

Run Qwen3.6-35B-A3B locally on 16 GB Mac, CPU-only, fast on Apple Silicon, no cloud needed, with optimized memory use via expert streaming.

Compact hardware controller for AI agents, featuring tactile controls, customizable keys, RGB lighting, and seamless Codex integration.

Anthropic and Blackstone back Ode, a $1.5B AI implementation startup, betting enterprise AI deployment is the next trillion-dollar market.

Dropper allows quick, direct file uploads from Mac menu bar to your own Cloudflare R2 storage, generating shareable, customizable pages instantly.

AI prompt-wait-evaluate loops break focus, creating shallow work days by disrupting flow and mental context rebuilding.

AI data centers face local opposition, but the bigger issue is corporate concentration of AI wealth and power shaping policies globally.

An IEEE article explores how AI can be manipulated for unethical or malicious purposes.

Story intertwines political thriller and coming-of-age, exploring betrayal, responsibility, forgiveness, hope, and love through layered characters.

Meta harnesses enable managing multiple agents simultaneously, boosting productivity for quick tasks but less suited for complex, larger projects.

OpenAI releases Codex Micro, a hardware device to enhance coding via Codex, separate from its rumored smart speaker project.

AI can reason over curated data to identify root causes in complex systems, but data preparation (harness) remains the biggest challenge.

ReadAloud offers free, high-quality, offline, browser-based text-to-speech with natural voices—ideal for listening anytime, anywhere.

Musk promises to delete user data after Grok Build stopped uploading entire repos due to a server-side fix following privacy concerns.

Palmer Luckey predicts AI will revolutionize code efficiency, achieving highly optimized results akin to John Carmack’s machine code.

GNNs struggle under label noise; robustness depends on backbone, structure, and scale, often collapsing at large widths.

AI models trained to solve Sokoban efficiently, setting speedrun records for RL and LLM approaches on custom benchmarks.

AI can help improve FOIA requests in 2026, but access issues may occur due to website security blocks.

AI tool helping K-12 teachers plan lessons, differentiate instruction, and save time with curriculum-aligned, privacy-focused features.

Elon Musk announces complete deletion of all user data uploaded to SpaceXAI prior to now.

Optimized NF4 dequantization kernel in Triton achieves up to 1.41x faster speed than bitsandbytes on GPU.

AI suggests we might live in a simulation due to advanced world models; consciousness remains unproven but simulation theories gain support.

AI workflow builder app template for React using JointJS; simplifies creating visual AI applications with customizable tools.

Job seekers use AI to cheat in interviews but often fail in the workplace.

Gartner predicts AI PCs will cut cloud token costs, enabling local AI workloads for enterprises by 2029, transforming future IT infrastructure.

The myth that AI will double unemployment is flawed; the math behind this fear is seriously exaggerated.

LHIC is a fast, secure local-first browser agent that eliminates LLM costs, offers robust resilience, and integrates advanced security features.

New York bans large-scale AI data centers for a year over power, environmental, and cost concerns amid public opposition and political debate.

Vint Cerf proposes DNS-based ID standards for AI agents, aiming for accountability, interoperability, and a future where agents interact autonomously.

Meta sued for allegedly using discriminatory AI in layoffs targeting employees with disabilities or on leave.