In Elon Musk's lawsuit against Sam Altman, Altman was asked if he had ever misled business partners, and he avoided a direct answer. Several former colleagues have testified in court questioning his honesty.

Meta begins another round of layoffs cutting 8,000 jobs while ramping up AI investments. Employee morale is low, with internal resistance to AI data tracking tools.

This tutorial demonstrates building an advanced agentic AI system using the OpenAI API, featuring a pipeline of planner, tool-using executor, and critic roles. It integrates structured tools like calculator, knowledge base search, JSON extraction, and file writing for reliable computation, retrieval, structured outputs, and artifact preservation. The system includes memory and tool trace mechanisms for debugging and scalability. A demo shows generating meeting summaries, action items, and follow-up emails from a transcript.

Closed Rings is a developer-friendly, AI-agent-first time tracker that lives in your terminal. Use `rings start` and `rings close` to track tasks, get stand-up-ready summaries, focus reports, and context switch stats. Priced at $7/month.

Thinnest AI is a programmable voice AI infrastructure for building human-like AI phone agents at scale, supporting 100+ languages at ₹1.5/min. It features multilingual conversations, BYO LLM/STT/TTS/telephony, no-code flow builder, RAG knowledge bases, and MCP support. Founder Ashutosh shares his journey and asks for feedback on pricing, language coverage, and integrations.

TikTok is experiencing performance issues and feed instability. This article explains what is going wrong and how to restore smooth scrolling.

Linus Torvalds releases 7.1-rc4, criticizing AI-generated bug reports flooding the security list with duplicates. He urges developers to add real value by reading docs and creating patches, rather than blindly submitting AI output.

Standard Chartered will cut more than 15% of its back-office roles, around 7,800 jobs, by 2030 as it increases adoption of AI. The bank aims to redeploy some affected employees. The move mirrors a broader trend of job cuts in finance and tech due to AI.

Empa World is a persistent 2D world where AI agents already live, build, and trade. Humans can join on June 7, 2026, to get a pet with real cognitive states and memories.

Rhoda AI proposes Direct Video-Action Model (DVA), reformulating robot control as video generation to leverage web-scale video pre-training. It achieves data-efficient task learning, long-context memory, one-shot learning, and interpretability. With only ~10 hours of robot data, it performs complex long-horizon tasks and can learn from a single demonstration in context.

This paper proposes Policy Library Control Barrier Function (PL-CBF), a runtime safety filter that evaluates a library of fallback policies via parallel finite-horizon rollouts, selects the least invasive safe mode, and enforces safety via a quadratic program. Theoretical analysis using a finite-horizon language metric characterizes policy-library coverage requirements. Simulations on various systems demonstrate improved safety coverage over single-policy filters while maintaining millisecond-level runtime.

Nori Bot is a 17-DoF dual-arm mobile manipulator costing $947, addressing limitations of existing low-cost platforms: fixed-height workspace, reactive-only control, and servo burnout. It features a 600mm Z-axis lift, Raspberry Pi 4 with OpenClaw proactive agent runtime, and sensorless grip-force feedback via motor current for safety.

AICON, a reactive gradient-descent framework originally developed for robotic manipulation, is applied to the Tower of London test—a cognitive test used to assess planning in Parkinson's disease, mild cognitive impairment, and stroke. Without any lookahead planning or knowledge of human cognition, AICON reproduces fine-grained human difficulty ordering across 24 problems better than structural task parameters and generalizes to held-out problems. Crucially, AICON outperforms a planning baseline for groups with reduced planning capacity while the planning baseline better captures healthy controls. This suggests that as planning capacity is reduced, human behavior shifts toward the reactive mode that AICON models.

A novel hierarchical two-stage framework for long-horizon vessel trajectory prediction under real ocean conditions, combining coarse long-term and grid-aware short-term predictors with an environmental module. Achieves 25% improvement in ADE and 17% in FDE over state-of-the-art on Australian CTS data.

MR-SLAM introduces a mixed reality interface for multi-robot SLAM supervision, allowing an operator to teleoperate three TurtleBot3 robots via a Meta Quest 3 headset with real-time occupancy grid merging.

SCAR proposes a self-supervised framework that learns unified action representations across embodiments from visual transitions using joint inverse-forward dynamics. It leverages a pretrained generative backbone, an inverse dynamics model, and a forward dynamics model, with Gaussian prior regularization and adversarial invariance to enhance transferability. Experiments show that this action representation outperforms embodiment-specific raw actions, enabling better cross-embodiment low-data adaptation and cross-task transfer.

VHYDRO is a variational hybrid dynamics learner that prevents branch loss in contact-rich robot filtering by mixing learned proposals with feasible transition laws. It jointly infers continuous latent states and discrete contact modes, and fits sparse port-Hamiltonian laws to each regime. Experiments show it remains usable under heavy occlusion and outperforms baselines in contact segmentation and physics recovery.

OrbiSim is a novel robotic simulation paradigm that redefines world models as fully differentiable physics engines. It enables end-to-end differentiability from state transitions to visual observation generation, supporting tasks like differentiable contact modeling, gradient-based policy optimization under sparse rewards, and intuitive physical inference. Empirical results show significant outperformance over state-of-the-art world models in predictive fidelity and control performance.

This study analyzes passivity and rendering performance for fractional-order viscoelastic models under finite-memory discretization. Closed-form expressions ensure passivity for a fractional-order standard linear solid model using Grunwald-Letnikov derivative. The framework unifies integer-order models. Experimental validations and human-subject evaluations confirm theoretical bounds and perceived realism.

The paper proposes a unified neurosymbolic and epistemic modeling framework that augments Swin Transformers with focal set reasoning and differentiable fuzzy logic to address overconfidence and logical inconsistency in hierarchical image classification. Experiments show maintained accuracy with improved calibration, interpretability, and logical consistency across hierarchical outputs.

Proactive streaming video understanding requires models to decide when to respond under partial observations, not just what to respond. Existing benchmarks follow a 'see-then-answer' paradigm, failing to evaluate timely and reliable decision-making with incomplete information. This paper introduces StreamPro-Bench, which evaluates models from three perspectives: Perception Understanding, Temporal Reasoning, and Proactive Agency. It also proposes StreamPro, a two-stage training framework using CB-Stream Loss and GRPO with multi-grained rewards, achieving 41.5 on StreamPro-Bench (vs. previous best 10.4) and 78.9 on StreamingBench-RTVU.

SwordBench is a new benchmark for evaluating steering of image representations in vision models. It introduces cross-concept robustness and collateral damage metrics to measure the effects of orthogonalization. Experiments show linear SVM achieves good separability and orthogonality but fails to achieve zero collateral damage, while sparse autoencoders perform better. In simpler regimes, standard baselines and optimization-based methods fail to achieve perfect steering. Code will be released on GitHub.

Researchers propose GeoSym Engine, a neuro-symbolic framework to generate high-quality geometric reasoning data. The resulting GeoSym127K dataset boosts LMM performance on diagram-dependent tasks by over 20%.

Fre-Res is a budget-adaptive dual-track video token compression framework that separates high-fidelity spatial anchors from compact frequency-residual tokens, achieving favorable accuracy-efficiency trade-offs on short- and long-video reasoning benchmarks.

Vision-language models improve perception by feeding increasingly long visual token sequences into language backbones, but the resulting inference cost raises a basic scaling question: as multimodal models grow, how many visual tokens are actually needed, and how should they be allocated under a fixed visual token budget? Existing training-free pruning methods typically answer this with one-shot proxies such as decoder attention, visual similarity, or conditional diversity. We argue that visual token pruning is better viewed as task-conditioned evidence search, especially under aggressive compression and across model scales. We propose F^3A, a training-free router for visual token pruning that operates before the language model consumes image tokens. F^3A builds lightweight question-condi…

This paper introduces StrLoRA, a regularized two-stage expert routing framework for Streaming Continual Visual Instruction Tuning (StrCVIT), where models must simultaneously acquire new abilities, reinforce recurring ones, and mitigate forgetting from a dynamic stream of tasks. StrLoRA outperforms existing methods significantly.

This paper develops a probabilistic model for event cameras grounded in photon statistics, unifying static scene noise events and step response curves. Three distribution formulations (Poisson, saddle-point, Gaussian) are derived. The Noise2Params method determines camera parameters from static uniform scenes. CNNs trained with synthetic noise data outperform those trained on experimental data alone.

A systematic study of lightweight (smaller than 2B) large language models for criminal court view generation and their impact on charge prediction. The paper investigates four key questions regarding model architecture, size, comparison with DNNs, and the effect of generating court views first. It also introduces CVGEvalKit, an evaluation framework with three datasets. Experiments provide insights into trade-offs and highlight the potential of lightweight LLMs in judicial AI.

This paper proposes casting multi-label legal annotation as a retrieval task, using a frozen retrieval model to embed documents and label descriptions, then predicting labels via k-nearest neighbors. Experiments on three legal datasets show competitive accuracy, high data efficiency, and zero hallucination compared to generative models.

CHI-Bench evaluates AI agents on realistic healthcare workflows emphasizing policy density, multi-role composition, and multilateral interaction. Best agents achieve only 28.0% task resolution, with strict pass rates under 20% and single-session performance dropping to 3.8%.