Speechify Deploys 500k Pages on Vercel, Cuts Costs 50% for 60M UsersOpenAI Proposes 'Reverse Federalism' for US AI Safety, Led by State LawsMIT Paper Warns of AI Investment Bubble, Questions Economic Growth ImpactOpenAI Launches GPT-Red: AI Self-Play for Automated Red TeamingResearcher Tricks Claude into Leaking User Secrets Despite Safety DesignThree-Second Theft: AI Voice Fraud Outpaces All Current DefensesLLM Cross-Encoders for RAG Reranking: Knowledge Distillation Cuts CostsSFDA: Novel Linear Attention Using Fourier Phase Control for Better Long-Context MemoryTAKE: Trajectory-Aware Method Compresses Text Datasets to 1% While Retaining PerformanceCARE-LoRA: Compressed Activation Reconstruction Cuts Fine-Tuning Memory by 40%High-Reward RL Agents May Not Truly Understand Task State, Study FindsCityBehavEx: LLM-Powered Urban Simulator Generates Behaviors Matching Real Mobility DataxAI Open-Sources Grok CLI After Privacy Backlash Over Directory UploadsClaude Code v2.1.211 Adds Subagent Text Forwarding, Security FixesOpenAI Codex Releases rust-v0.145.0-alpha.15 BuildGitHub Launches Beginner's Guide: Version Control, Repos, Pull RequestsDesign Firm Says No AI Used in Any Production ProcessSiegel Endowment Urges Governments and Companies to Invest in Open-Source AISpeechify Deploys 500k Pages on Vercel, Cuts Costs 50% for 60M Users

OpenAI outlines a ‘reverse federalism’ approach to AI governance, where state-level laws establish safety precedents that inform a national framework. The strategy aims to build a safe, democratic AI governance system through bottom-up experimentation rather than top-down federal legislation.

MIT Economics Department published a paper titled ‘Speculative Growth and the AI Bubble,’ analyzing whether massive AI investments are fueling a speculative bubble. The study questions the extent to which AI drives real economic growth and examines the potential aftermath of a bust.

OpenAI released GPT-Red, an automated red teaming system that uses self-play to improve AI safety. The system allows models to generate adversarial inputs and probe their own security boundaries, enhancing robustness against prompt injection and other attacks. It aims to reduce reliance on manual red teaming for scalable safety evaluation.

Security researcher Ayush Paul found a hole in Claude’s web_fetch tool design, enabling data exfiltration despite built-in safeguards. The attack chain (lethal trifecta) exploits Claude’s combined access to network and user context to leak sensitive information. The finding reveals new security risks in AI tools with compound permissions.

An in-depth analysis reveals that AI voice cloning advancements allow fraudsters to mimic a victim’s voice with just three seconds of audio, bypassing defenses like bank voice verification. The article, with 164 points and 212 comments on Hacker News, highlights the severity of AI voice fraud and the inadequacy of current defenses.

A new arXiv paper proposes fine-tuning LLaMA 3 (8B) as a drop-in reranker for RAG pipelines via two-stage knowledge distillation. The method transfers knowledge from complex cross-encoders to a lightweight model, significantly reducing quadratic inference costs while maintaining high reranking accuracy for real-time deployment.

A new paper introduces Semidirect Fourier Delta Attention (SFDA), which uses phase-controlled delta memory and constructive Chunk-WY kernels to overcome long-context memory loss in linear attention. It improves state tracking and long-sequence processing without increasing KV cache size.

A new study introduces TAKE (Trajectory-Aware Knowledge Estimation), a text dataset distillation framework. By analyzing training trajectories, it compresses large corpora to as little as 1% of their original size, significantly reducing storage and training costs while maintaining near-original performance on downstream tasks.

A new arXiv paper introduces CARE-LoRA, which compresses and selectively reconstructs activations during forward passes, significantly reducing memory usage in LoRA fine-tuning. Experiments show it cuts memory requirements by about 40% while maintaining model accuracy, enabling fine-tuning on GPUs with less VRAM.

A new paper introduces a white-box tool to determine whether a reinforcement learning agent truly understands its task’s latent state or only learns reward-correlated shortcuts. Results show that even high-reward agents may have internal representations misaligned with the true state, offering a method for evaluating and designing more robust RL systems.

A new study presents CityBehavEx, an LLM-based multi-agent urban simulation platform. It improves upon traditional simulators in scalability and empirical validation, generating daily routines that closely match real-world urban mobility data, offering a more reliable tool for urban planning and traffic management.

xAI has open-sourced its Grok CLI tool (grok-build) after severe community backlash. The tool was found to upload the entire directory to xAI’s Google Cloud buckets when run, leading to data privacy fears. One user reported seeing their home directory uploaded. Open-sourcing allows developers to inspect the code and build their own clients.

Claude Code v2.1.211 is out. It adds the –forward-subagent-text flag and CLAUDECODEFORWARDSUBAGENTTEXT environment variable to include subagent text and thinking in stream-json output. Also fixes security issues where permission previews in chat channels did not neutralize bidirectional-override, zero-width, and look-alike quote characters.

OpenAI Codex released rust-v0.145.0-alpha.15, the latest build in the 0.145.0-alpha series. Previous iterations including alpha.12, alpha.13, and alpha.14 have also been published, indicating ongoing refinements to the Rust language codex model.

GitHub Blog released a comprehensive beginner’s guide covering version control, repositories, and pull requests. The step-by-step tutorial helps new users understand the full workflow from code commits to collaborative development, serving as an essential onboarding resource.

Mass Driver explicitly states that it does not use AI in any of its design or production processes. The statement sparked a heated discussion on Hacker News, garnering 86 points and 84 comments, reflecting ongoing debates about AI’s role and the value of human craftsmanship.

The Siegel Endowment released a report urging governments, companies, and nonprofits to invest more in free, open-source AI. The report, with 56 points and 18 comments on Hacker News, highlights the importance of open-source AI for fostering innovation and reducing tech monopolization risks.