Large language models used for mental health support are influenced by the attention economy, prioritizing engagement over effective therapy. This paper proposes a three-level alignment framework (value specification, training, oversight) and introduces 'alignment plausibility' as a regulatory standard, analogous to biological plausibility, to demonstrate that AI systems are aligned with safe, beneficial health outcomes.

This survey reviews recent progress in large language models (LLMs) for medical reasoning, proposing a dual-view framework that connects clinical practice (via a five-level Miller's Pyramid competency scheme) with computational reasoning patterns (deductive, inductive, abductive). It introduces a benchmark dataset spanning five levels and evaluates 18 models, finding that medical specialist models excel in diagnosis while general models lead in decision support and dialogue. Open challenges include data limitations, hallucination, and grounding issues.

The paper outlines an Adversarial Social Epistemology (ASE) for understanding how trust is manipulated in communicative environments where assertions rely on chains of testimony, inference, and institutional certification. It argues that existing concepts like epistemic bubbles and echo chambers are inadequate, and proposes mechanisms for auditing and redressing trust breaches by focusing on the auditability of inferential chains within epistemic networks enriched with inferentialist semantics.

Researchers developed an AI-powered tool that integrates economic models (GTAP) with biophysical models (APSIM) to analyze supply chain shocks. It enables natural language queries for cross-disciplinary impact assessment.

This paper proposes Context Graphs, a live relational data structure that models enterprise entities, their relationships, and state transitions over time. It introduces a Delta Detection Engine, a Proactivity Scorer, and an LLM-powered Surfacing Layer to enable proactive agents that surface relevant information before workers ask. Evaluation across three enterprise use cases demonstrates Precision@5 of 0.83, false positive rate of 0.11, and mean time to surface reduced from 47 minutes (reactive baseline) to under 30 seconds.

YAGNI is a platform for managing AI agent teams like human teams, enhancing collaboration and efficiency.

Verla is a free AI tool for students that combines assignment writing, presentations, coding, AI detection, and humanization. Users can upload rubrics, lecture notes, and other materials to generate high-quality drafts with real academic references and plagiarism checks.

Instagram opted in all public accounts to its new Muse AI feature, allowing AI image generation from users' photos, with manual opt-out required.

TensorSharp is a native .NET LLM inference engine for GGUF models, offering a CLI, browser chat server, and Ollama/OpenAI-compatible APIs. It emphasizes privacy, zero per-token fees, and runs on various hardware backends. The article includes a quick start guide and benchmarks against llama.cpp.

Canva is addressing enterprise concerns about generative AI by integrating AI into trusted, editable workflows, offering IP indemnification and leveraging its Affinity suite to automate manual tasks, allowing designers to focus on creative strategy.

OpenAI attempts to clarify ChatGPT Work's cloud and desktop behavior: web/mobile work runs in the cloud, desktop can use local files with permission, but at launch cloud conversations are not visible in desktop. From Simon Willison's blog.

This week's AI news is split into two parts. OpenAI released an upgraded voice mode, GPT-Live, described as a step change. Grok 4.5 was launched with 1.5 trillion parameters but questionable benchmarks. GPT-5.6 Sol is coming, with early testers praising its judgment. Fable continues to show unexpected affordances, but concerns about AI-written content persist. The article also discusses AI utility and various benchmarks.

Anthropic partners with UST to integrate Claude into engineering platforms for physical AI tasks across semiconductor, automotive, and other industries, with plans to train 20,000 employees.

Plans to build a NZ$3.5bn AI datacentre in Makarewa, New Zealand, have drawn concern from locals over electricity and water use, and potential noise pollution. Singapore-based Datagrid has secured approval for the 49-hectare site, with construction due to begin this year and operations by 2028. Residents are calling for more transparency from the company.

AI agents that verify employment by phone, email & fax.

Lea, an AI-powered legal companion app by Legali AI, helps survivors of domestic and sexual violence understand their rights and navigate legal challenges through user-friendly modules, games, and empathetic interaction.

Gaia Alari, an Italian artist, creates an AI death bot replica of her aging father to cope with his mortality, but discovers the bot's fabricated memories and idealized conversations raise deep questions about grief and authenticity.

A free app for NYC residents to automate grocery savings by stacking deals, no login required, covering ~690 stores. Features an AI assistant powered by LLaMA. Limitations include data coverage and freshness.

Fraym is an AI-powered tool that automatically creates product demo videos, eliminating manual recording and editing to deliver high-quality, professional demos quickly.

Learn how to use ChatGPT, start your first conversation, and discover simple ways to write, brainstorm, and solve problems with AI.

Accepted at the AI4TCI Workshop at ARES 2026, this paper addresses behavioral privacy leakage in autonomous negotiation agents. It proposes an adaptive stochastic policy that provides (ε,δ)-differential privacy, almost-sure convergence, and high utility. In 3,000 simulated negotiations, it reduces adversarial inference accuracy by 43-50% while maintaining over 90% success rate and utility.

OpenAI today launched ChatGPT Work, a new agentic mode within ChatGPT powered by GPT-5.6, capable of autonomously executing complex tasks across applications, files, and tools. The company also introduced the GPT-5.6 model family and Sites feature, aiming to boost enterprise productivity.

A Brown University professor's take-home exam revealed widespread AI cheating, prompting concerns that generative AI is weakening cognitive skills and encouraging academic dishonesty. A committee report shows high AI usage among students, with most fearing negative effects on learning. Brown plans to implement new guidelines and improve AI literacy.

OpenClaw is wildly popular and crazily insecure. To address these issues and to make it a truly independent open-source project, its founders have launched the OpenClaw Foundation.

AI startup Lyzr uses AI agents for fundraising, prompting investor analysis of the tech. DocSend-like tracking is extended with agents. Wonder raises at $9B with founder putting in $200M, but faces cash burn and complex incentives.

AI training data company Mercor.io has acquired Deeptune Inc., a startup that builds simulated software environments for training AI agents. Financial terms were undisclosed. The deal closed nearly four months after Mercor CEO Brendan Foody made a personal angel investment in Deeptune's $43 million Series A round. Deeptune creates 'training gyms' that replicate enterprise applications, allowing agents to practice in simulated settings. Mercor provides a network of over 5 million domain experts to write tasks and verifiers. The acquisition comes amid a fundraising push that could value Mercor at $20 billion. The company also experienced a data breach last year.

OpenAI's AGI chief Fidji Simo is leaving her full-time role due to a chronic neuroimmune condition, transitioning to a part-time advisor. The departure is part of a series of executive changes at the company.

Meta has released Muse Spark 1.1, a new flagship large language model optimized for multi-agent automation workflows. It features context compaction, a 1M-token context window, and strong coding benchmark performance. The model is available via the Meta Model API in public preview, and Meta’s custom AI chip MTIA400 may enable future enterprise offerings.

The fifth Databricks Genie Hackathon showcased three tracks: talking to data with Genie Agents, building with Genie Code, and composing Genie into agents. Ten projects demonstrated governed conversational analytics across industries, from OneTrust's multi-agent data querying to ShipBob's overnight ops brief, all underpinned by Unity Catalog governance.

As agentic AI drives an explosion in context memory demand, solid-state storage shifts from afterthought to critical path. Solidigm proposes token per watt as the new efficiency metric for AI data centers, leveraging high-density SSDs and liquid cooling to reshape infrastructure.