Large language models still struggle with social reasoning, according to a new benchmarking task called the Epistemic Asymmetry Schelling Task (EAST). The task reveals that even advanced models fail at epistemic tracking…
OpenAI CEO Sam Altman publicly criticized Anthropic's new social initiative, calling it satire and referencing the company's controversial guardrails on its Fable 5 model that silently downgraded users. The feud adds to …
Researchers introduced Adversarial World Modeling (AWM), a new framework that trains autonomous vehicle planners using a min-max game to handle rare, safety-critical traffic scenarios. Tested on nuPlan and InterPlan benc…
OpenAI's GPT-5.6 Sol and Anthropic's Claude Fable 5 represent a substantial advance in AI, demonstrating improved understanding of complex assignments and agentic capabilities that reduce the need for human guidance. The…
Researchers have developed Diachronic Sample Integration (DSI), a method that aggregates samples across training checkpoints to improve rare event prediction in deep generative models. In tests on synthetic data and high…
Taiwan Semiconductor Manufacturing Co. (TSMC) is expected to report its fifth consecutive record quarterly profit on Thursday, driven by surging demand for advanced AI chips from clients like Nvidia and Apple. Analysts p…
Researchers have developed a language-guided vision-AI framework for waste management that integrates visual and linguistic reasoning, achieving a BLEU score of 0.8291 and BERTScore of 0.9273. The system, built on the ne…
AgentCheck, an open-source platform for testing AI tools under real-world failure conditions, allows developers to simulate 12 types of faults before deployment. In tests across five agents, the top performer passed 105 …
Researchers introduced commit-time authorization as a new security standard for large language model agents, ensuring durable effects like database updates or sent emails are only committed if authorization conditions re…
Researchers introduced the Causal Discovery Foundation Model (CDFM), a unified framework for zero-shot structural causal inference that outperforms traditional algorithms across diverse domains. By treating unknown causa…
LaunchSafe's Mako, an autonomous web exploitation engine, uses a self-evolving AI agent to automate vulnerability targeting, achieving full coverage on public benchmarks. The system's dual-use nature raises ethical conce…
Researchers introduced RepTran, a repair method targeting feed-forward networks in Transformer models, achieving a 74.7% average repair rate on 18 fault benchmarks. The tool uses a two-pronged scoring system and differen…
A University of Lausanne study found that a machine learning pipeline assessing mental health risks in students produced stable results due to data construct correlations rather than accurate risk prediction, with model …
A developer published a technical guide for compiling llama.cpp on Debian 12/13, covering native builds for x86_64 and ARM64, as well as cross-compilation from x86_64 to ARM64. The guide includes steps for enabling OpenB…
Researchers introduced AMID, a multi-agent framework that automates medical imaging model development by refining task-specific data into high-performing models. Across 20 medical imaging tasks, AMID outperformed existin…
Neverswipe launches an AI agent that dates on behalf of users, handling introductions by matching people based on briefs rather than swiping. The service aims to replace traditional dating apps by having agents communica…
AI coding agents are optimized for generating new code, but software engineering increasingly relies on reusing existing components. This focus on generation over reuse creates technical debt, as token-based economics re…
Nvidia has slashed its list of Asian buyers for advanced AI chips by more than half, implementing a stringent new white list after tightening compliance checks to block shipments from reaching China, the Financial Times …
A new paper argues that AI safety discussions overlook the risk of large-scale suffering from misused or misaligned AI, particularly from uncontrolled AI aligned with values that do not prioritize suffering prevention. T…
Google's July 2026 System update is rolling out with improvements to the Play Store, Wear OS, and Maps, including AI-generated image labels in the EU and a smoother Google One purchase flow. The update also brings work p…