AEF-1 Standard for Third-Party Evaluators Emerges, Cosigned by xAI, OpenAI, and AnthropicSame Patient, Different Orders: Clinical LLM Agents Show Low Action-Level Reliability Across Repeated RunsZGCM-1: A Fully Open 7B Model for Math and Agentic SearchOne Formal Framework Unifies Iterative Policy Improvement and Recursive Self-ImprovementOrchSLM Probes Small Language Model Orchestration as Cloud LLMs Hit Latency, Privacy, and Cost LimitsModeling Application Behavior Cuts Token Costs for Web AgentsCan LLM Agents Manage Long-Horizon Physical Tasks? Study Explores Self-Adaptive Physical AIVercel's Agentic Audits Now Tailored by Site TypeDelphi Ships 100 Times a Day with Python Backend on VercelFormas Launches Cartesian, an AI 3D Modeling Tool for DesignersNVIDIA OpenShell Applies Formal Methods to Control AI Agents, Shares Lessons LearnedAI Is Breaking Our Proxies for ExpertiseAnthropic Co-Founder Tells BBC AI 'Kill Switch' May Need to Be MandatoryCan Game-Trained AI Skills Transfer to Real Work? It Depends on Training DesignGoogle Releases Gemini 3.8 Live Speech-to-Speech ModelsGoogle DeepMind Introduces Gemini 3.8 Live and 3.8 Live Extended ThinkingGoogle Turns ATLAS AI & Economy Data Into an Open Interactive ExperienceGoogle Aims to Bring AI to Every Language, Moving Beyond Text TranslationGoogle Outlines How AI Is Accelerating Science and Improving LivesGoogle Highlights AI Applications for Societal ImpactAEF-1 Standard for Third-Party Evaluators Emerges, Cosigned by xAI, OpenAI, and Anthropic

A new third-party AI evaluation standard called AEF-1 has emerged, cosigned by xAI, OpenAI, and Anthropic. The move signals a shift toward unified model benchmarking after years of fragmented, vendor-specific evaluation. For developers and enterprises, it could provide a cross-comparable reference when selecting models, rather than relying solely on self-reported vendor metrics. Detailed metrics and implementation rules have not yet been fully disclosed.

A new study reveals reliability issues in clinical LLM agents: given identical inputs, an agent may produce the same diagnostic verdict while filing materially different test orders, prescriptions, and referrals on each run. Current clinical agent benchmarks typically score one run per task, failing to capture this action-level inconsistency. The research calls for multi-run, action-level evaluation standards, warning that agent safety in real clinical settings cannot be guaranteed otherwise.

Researchers introduced ZGCM-1, a fully open 7B dense foundation model trained from scratch with extreme data, system, and algorithmic efficiency. The core premise: compact models cannot passively memorize the open web, but can be overtrained for specific capabilities. ZGCM-1 targets math and agentic search tasks and is fully open-source. For resource-constrained research teams, it offers a reproducible and efficient foundation model option.

A new paper proposes a generalized agent iteration framework that formally describes both iterative policy improvement and recursive self-improvement (RSI) under one system. The authors note that RSI is claimed at many scales today but lacks a unified formal description, causing the concept to blur across phenomenon, mechanism, and prospect. The framework aims to give autonomous evolving intelligence research a shared language. It is foundational theory work with no immediate engineering payoff.

A new paper studies the dynamics of small language model (SLM) orchestration. It notes that while LLMs are highly capable, their reliance on cloud infrastructure poses fundamental challenges for agentic pipelines, including latency, privacy, connectivity, and cost. The research focuses on using small-model orchestration to mitigate these issues. This is a methods-level exploration; concrete performance gains and applicable scenarios require reading the full paper. It is relevant to teams deploying agents locally or at the edge.

A new paper proposes token-efficient task execution for web agents via application behavior modeling. It notes that strong AI agent performance across tasks has driven massive investment in agentic infrastructure, but token processing costs are rising fast. The method aims to reduce unnecessary token consumption in web application automation. Specific savings ratios and benchmarks require the full paper. It has cost-optimization value for teams deploying web agents at scale.

A new arXiv paper explores whether LLM agents can autonomously manage long-horizon physical tasks. The research notes that physical tasks require agents to continuously observe environments, take consequential actions, and adapt strategies dynamically—posing higher demands on current perception-decision loops. The paper proposes a ‘self-adaptive physical AI’ direction, exploring how LLM agents can complete multi-step physical operations without human intervention. The work is currently methodological, with no specific benchmark data or comparison results disclosed yet.

Vercel’s Agentic reports now let users view audit checks through four site types: Docs & content, Business, App, and Commerce. Each view surfaces different standards—the Commerce view highlights payment and checkout protocols like x402, UCP, and ACP, while the App view focuses on API discovery, authentication, and error handling. The update helps developers and site operators quickly find relevant issues based on their business type instead of sifting through unrelated checks.

Delphi runs its Python backend on Vercel with 10 engineers and no dedicated infrastructure role. Everyone ships code, including product and design, achieving 100+ production deploys a day behind feature flags. Delphi builds digital minds by capturing what someone has written, recorded, and taught so anyone can tap that expertise on demand. The case demonstrates how a small team can achieve high-frequency deployment on Vercel.

Formas has launched Cartesian, an AI-powered 3D modeling tool aimed at design workflows. The tool drew attention on Hacker News with 86 points and 72 comments, signaling strong interest in AI-assisted 3D modeling among designers. The submission only links to the product page; specific capabilities, supported formats, and workflow details are not outlined. It may be worth trying for 3D designers looking for new tools.

NVIDIA’s OpenShell Research team published development notes summarizing lessons from applying formal methods to control AI agents. The post centers on building an agent policy prover and explores how mathematically verifiable techniques can constrain agent behavior. It drew 31 points and 11 comments on Hacker News, a moderate level of discussion. For researchers focused on AI safety and controllability, it offers an industry-side practical record.

Sean Goedecke published a blog post on how AI is breaking the proxies we use to judge expertise, such as credentials, certificates, writing style, or code style. When AI can easily mimic these signals, the mechanisms that rely on them for hiring or credibility assessment start to fail. The post drew 83 points and 72 comments on Hacker News, generating substantial discussion. It has direct implications for hiring, evaluation, and knowledge dissemination.

Anthropic co-founder Jack Clark told the BBC that AI ‘kill switch’ mechanisms may need to be mandatory rather than voluntary. He argued that industry self-regulation is insufficient as AI capabilities rapidly advance, and governments should set hard standards. The stance aligns with Anthropic’s long-held position on ‘responsible scaling.’ The discussion on Hacker News drew 41 points and 100 comments. If regulation follows, AI companies would need to embed externally triggerable shutdown mechanisms in their products.

Good Start Labs trained an AI on a railroad game, and one version improved at financial research tasks. The key difference was the training design, not the game content itself. The experiment suggests that whether AI skills learned in games transfer to real-world work depends on how the training is structured, rather than surface similarities between the game and the target task. For researchers studying AI generalization, this offers a practical lever: transfer outcomes can be shaped through training design.

Google DeepMind introduced Gemini 3.8 Live and 3.8 Live Extended Thinking. This is Google’s latest release in the real-time voice interaction space, with the Extended Thinking variant targeting voice scenarios that require deeper reasoning. Specific technical metrics and release dates are not provided in the summary; consult the official blog for full parameters and benchmark results.

Google has turned millions of global data points from its AI & Economy ATLAS project into an interactive, open-access experience for the public. ATLAS was previously used mainly to study AI’s economic impact; opening it up means researchers and general users can directly query and analyze the data. The post does not specify data dimensions, update frequency, or access entry points. For those studying AI’s economic impact or doing policy analysis, it is a new public data source.

Google AI Blog says it is moving beyond traditional text translation to build models that understand the world’s living languages as they are actually expressed. The shift means models would not just convert one language into another but grasp the culture and context behind them. No specific model names, language counts, or launch dates were given, making this a directional statement. If delivered, it would meaningfully improve AI tool usability for non-English speakers.

Google AI Blog published a post outlining its focus areas for using AI to accelerate scientific research and improve lives, targeting domains where advanced technology can drive breakthroughs. The article stresses that the true measure of AI is who it helps, citing examples of current real-world impact. It is largely a vision statement without new technical details or product launches.

Google published a collection of case studies showing how experts and local leaders are using AI breakthroughs to advance societal impact, aiming to ensure everyone can share in AI’s opportunities. The collection covers multiple real-world application scenarios; specific project details and participants require consulting the original source.