Windows 11 adds AI agent that runs in background with access to personal folders

Microsoft is introducing an experimental "Agent Workspace" feature in Windows 11, allowing AI agents to operate continuously in the background. These agents run in a contained session with their own runtime and user account but require access to personal folders like Desktop and Documents to perform tasks. The OS explicitly warns that enabling this feature poses potential security, privacy, and performance risks.

Read more

Grok 4.1

xAI has released Grok 4.1, a new model focused on improving creative, emotional, and collaborative interactions. It was trained using large-scale reinforcement learning with a novel method that leverages agentic reasoning models as reward models to optimize for non-verifiable signals like style and personality. The model now ranks #1 on the LMArena Text Leaderboard and demonstrates significant improvements on benchmarks like EQ-Bench and FActScore, showing enhanced emotional intelligence and reduced hallucination rates.

Read more

How to turn off Copilot and protect your data from Microsoft's AI

Microsoft is deeply integrating its OpenAI GPT-powered Copilot across Windows 11 and Microsoft 365, but full removal is restricted to enterprise users. Personal users can only limit its visibility and disable some data collection features, such as model training and personalization. This integration introduces significant privacy risks, including data usage for ad profiling and AI training, and creates new security vulnerabilities. Recent exploits like EchoLeak (CVE-2025-32711) and the CoPhish phishing technique demonstrate how Copilot's system-level access can be leveraged as an attack vector for data exfiltration.

Read more

CoreWeave, the AI industry's ticking time bomb

An article from The Verge characterizes AI compute provider CoreWeave as a "ticking time bomb." Despite its impressive public metrics, including major clients like Microsoft and OpenAI, massive revenue growth, and a successful IPO, the piece suggests a closer look reveals a deeply troubling situation.

Read more

Peter Thiel has sold his stake in Nvidia

Peter Thiel's investment fund sold its entire Nvidia stake in Q3 amid concerns of an AI bubble and mixed institutional sentiment. The move aligns with significant insider selling and short positions from other notable investors, even as sell-side analysts remain bullish. The article also highlights the growing physical constraints on AI expansion, such as power grid bottlenecks caused by the massive energy demands of new data centers.

Read more

Attacker Moves Second: Adaptive Attacks Bypass Defenses Against LLM Jailbreaks

Current evaluation methods for LLM defenses are flawed because they use static, non-adaptive attacks. The authors argue that defenses must be evaluated against adaptive attackers who use scaled optimization techniques like gradient descent and reinforcement learning to specifically target and bypass the defense mechanism. By applying this methodology, they successfully broke 12 recent defenses with over 90% success, demonstrating the need for more robust adversarial benchmarking to make credible claims of robustness.

ArXiv

Does AI-Assisted Coding Deliver? A Study of Cursor's Impact on Software Projects

A study using a difference-in-differences design on GitHub projects estimated the causal effect of the LLM agent Cursor. The findings show that Cursor adoption leads to a significant but transient increase in development velocity. This is accompanied by a persistent increase in static analysis warnings and code complexity, which was found to be a major factor causing a long-term velocity slowdown.

ArXiv

LeJEPA

This work presents a comprehensive theory for Joint-Embedding Predictive Architectures (JEPAs), identifying the isotropic Gaussian as the optimal embedding distribution to minimize downstream prediction risk. The authors introduce LeJEPA, a lean and theoretically grounded training objective that combines the standard predictive loss with a novel Sketched Isotropic Gaussian Regularization (SIGReg) to enforce this distribution. This approach eliminates common heuristics like stop-gradients and teacher-student networks, offering a simple, scalable, and stable self-supervised pre-training method that achieves 79% on ImageNet-1K with a ViT-H/14 using linear evaluation.

ArXiv

TabPFN-2.5: Advancing the State of the Art in Tabular Foundation Models

TabPFN-2.5 is a new tabular foundation model that scales to datasets with up to 50k data points and 2k features, a 20x increase over its predecessor. It achieves state-of-the-art performance on the TabArena benchmark, matching complex AutoGluon ensembles and significantly outperforming default XGBoost. For production, a new distillation engine converts the model into a compact, low-latency MLP or tree ensemble while preserving most of its accuracy.

ArXiv

AA-Omniscience: Evaluating Cross-Domain Knowledge Reliability in Language Models

The AA-Omniscience benchmark measures LLM factual recall and knowledge calibration using an "Omniscience Index" that penalizes hallucinations and rewards abstention. Results on its 6,000 questions reveal persistent factuality weaknesses in frontier models, with only three scoring above zero; Claude 4.1 Opus led with a score of 4.8. Since performance varies significantly by domain, the findings suggest selecting models based on use-case specific knowledge rather than general capabilities.

ArXiv

Show HN: ESPectre – Motion detection based on Wi-Fi spectre analysis

ESPectre is an open-source project for motion detection using Wi-Fi Channel State Information (CSI) on a low-cost ESP32-S3. It currently uses a mathematical approach, extracting 10 features from the CSI data via signal processing to work without ML training. The extracted features are designed to serve as a foundation for future ML models for more advanced tasks like activity recognition and people counting.

Repo

Show HN: Model-agnostic cognitive architecture for LLMs

The Persistent Mind Model (PMM) is a deterministic, event-sourced cognitive architecture that provides AI agents with a persistent, model-agnostic identity. It uses an immutable ledger to record all interactions and internal states, from which the agent's complete mental state is reconstructed before each LLM call. This makes identity independent of the underlying model, allowing for state continuity across sessions and LLM swaps without fine-tuning or conventional RAG. The LLM interprets the reconstructed context and outputs control lines that the PMM runtime uses to deterministically update the ledger.

Repo

Show HN: Epub2md – Turn ePub books into Markdown folders for LLM agents

epub2md is a Python CLI tool that converts EPUB files into clean, chapter-separated Markdown files. This is useful for preprocessing books into a structured text corpus for LLM training or RAG systems. The tool is installable via pip and has an external dependency on pandoc.

Repo

A first-principles model for replacing income tax in an AI-driven economy

PUT-Monolith-v2 is a compact, machine-readable specification of the Public Usage Tax (PUT) system, designed for ingestion by LLMs. It functions as a logic-complete ruleset that can be loaded as a system prompt or reasoning module to enable consistent, non-contradictory analysis of the tax architecture. The artifact encodes the system's core invariants, guardrails, and dynamic rules to provide a stable foundation for AI-driven reasoning without political framing.

Repo

Seekdb，unified search database for AI(relational, vector and full text)

OceanBase seekdb is an AI-native database that unifies vector, text, and structured data to enable hybrid search. It supports in-database AI workflows through SQL functions for embedding, text generation, and reranking, simplifying RAG and other LLM applications. The MySQL-compatible engine integrates with popular AI frameworks like LangChain and LlamaIndex and can be deployed in embedded, single-node, or server modes.

Repo