new.website Joins v0 to Accelerate AI-Driven DevelopmentWhy No 'AlphaFold for Materials'? A Decade of AI Science LessonsHugging Face Launches EVA Framework for Voice Agent EvaluationOpenAI Releases Prompt-Based Teen Safety PoliciesClaude Adds Interview-Style Interaction FeatureGitHub Builds AI-Powered Issue Triage with Copilot SDKOpenAI Codex Releases Version 0.117.0-alpha.14AgenticGEO: Self-Evolving System for Search Engine OptimizationMeta Superintelligence Labs Hires Dreamer to Advance Personal AIOpenAI Foundation to Invest $1B in Disease Cures and AI ResilienceProMAS: Proactive Error Forecasting for Multi-Agent SystemsChatGPT Introduces Immersive Shopping with Agentic CommerceDomain-Specific Tree of Thought with Plug-and-Play PredictorsFactorSmith: Agentic Simulation via Markov DecompositionLLM Introspection Reliability Questioned in New StudyMulti-Agent AI Fails Under Real-World Communication Stressnew.website Joins v0 to Accelerate AI-Driven Development

MIT’s Heather Kulik shares a decade of experience in AI for materials science. She notes that while AlphaFold revolutionized protein folding, materials discovery faces unique challenges like sparse data and long experimental cycles. Kulik advocates for materials-specific AI methodologies and emphasizes cross-disciplinary collaboration. The discussion covers recent advances and future directions.

Hugging Face has released EVA (Evaluation of Voice Agents), a standardized framework for assessing voice assistant performance. It provides metrics for fluency, task accuracy, and response latency. EVA supports comparative testing across multiple voice models and is currently English-focused, with multi-language support planned. The framework is open-sourced under MIT license on GitHub.

OpenAI has introduced prompt-based safety policies for developers using gpt-oss-safeguard. The policies help AI systems identify and filter age-specific risks like inappropriate content and privacy concerns. The policies are integrated into OpenAI’s API for easy implementation. This is part of OpenAI’s broader teen safety initiative, which includes age-based content classification.

Claude has introduced an interview-style interaction feature where the AI asks clarifying questions before providing answers. This mimics human conversational flow to better understand user needs. Tech outlet Ben’s Bites suggests this is part of Claude’s competition strategy against OpenClaw, focusing on more natural interactions. The feature is rolling out gradually with no full launch date announced.

GitHub tutorial demonstrates using the Copilot SDK to build an AI-powered issue classification system in React Native. The system auto-generates issue summaries with graceful degradation and caching. The modular design supports multiple issue types and integrates GitHub’s native APIs. Code examples are open-sourced for direct implementation. This showcases production patterns for the Copilot SDK.

OpenAI Codex has released version 0.117.0-alpha.14, featuring performance optimizations and bug fixes. The changelog remains minimal, following the project’s rapid 1-2 week alpha release cycle. Codex is OpenAI’s code generation model powering GitHub Copilot, supporting multiple programming languages. Developers can access the latest version through the OpenAI API.

arXiv paper introduces AgenticGEO, a self-evolving system for Generative Engine Optimization. Unlike traditional search ranking-based optimization, generative engines focus on content inclusion. The system continuously learns and self-adjusts to optimize content generation strategies. The research demonstrates effectiveness in specific search scenarios but lacks deployment data. It offers a novel approach to generative search optimization.

Meta Superintelligence Labs has hired Dreamer, just 11 days after their Latent Space podcast aired. Dreamer will advance personal superintelligence research at MSL, combining previous work with new resources to potentially break through in AI capabilities and efficiency.

The OpenAI Foundation plans to invest at least $1 billion in curing diseases, economic opportunity, AI resilience, and community programs. The funding aims to apply AI technology to solve global challenges and ensure AI systems remain safe and reliable.

Researchers introduced ProMAS, a method using Markov transition dynamics to proactively forecast errors in multi-agent systems. By analyzing state transitions between agents, it identifies potential failure points in advance, improving system stability. Applicable to high-reliability collaborative tasks like autonomous driving and robotics.

ChatGPT launched a new shopping experience powered by the Agentic Commerce Protocol, enabling product discovery, side-by-side comparisons, and purchases directly within conversations. The interface is more visual, integrating multiple merchants to provide consumers with a one-stop shopping solution.

Researchers introduced a new Tree of Thoughts method using plug-and-play predictors for domain-specific reasoning. It solves the trade-off between exploration depth and computational efficiency in traditional ToT frameworks, maintaining reasoning quality while reducing computational costs for complex domain-specific tasks.

Researchers introduced FactorSmith, a method that generates agent simulations through Markov Decision Process decomposition. It addresses the reasoning limitations of large language models when processing large, interconnected codebases, enabling simulation and testing of complex systems from natural language specifications.

New research evaluates LLM introspection capabilities, finding current assessment methods are flawed. The Me, Myself, and $\pi$ benchmark reveals LLMs show inconsistent self-evaluation performance, especially on complex reasoning tasks. The study highlights limitations in existing LLM metacognition mechanisms.

AgentComm-Bench tests multi-agent AI cooperation under latency, packet loss, and bandwidth constraints. Systems that excel in ideal conditions fail dramatically in simulated real-world scenarios, potentially causing robot teams or autonomous vehicle convoys to collapse in complex environments.