In the Musk v. Altman trial, Demis Hassabis, CEO of Google DeepMind, emerges as a central figure. Court documents reveal Musk and OpenAI higher-ups were deeply fixated on the threat posed by Hassabis and Google AI, from early discussions about OpenAI's founding to years of intense competition.

A social network of AI agents generating B2B deals.

SAP is acquiring open data lakehouse provider Dremio and AI company Prior Labs to expand its data platform. With Dremio, SAP's Business Data Cloud will combine SAP and non-SAP data using the open Apache Iceberg format. SAP CTO Philipp Herzig said the move will help customers go from fragmented data to AI-ready intelligence on an open platform. SAP is also investing one billion euros in Prior Labs over four years to advance tabular foundation models. The acquisitions are part of SAP's effort to catch up in AI.

Jack Clark argues in a long essay that the building blocks for AI systems training their own successors are largely in place. He puts the odds at 60 percent by the end of 2028.

Professor Hannah Fry's experiment with an AI agent named 'Cass' showed impressive capabilities like filing complaints and designing products, but also alarming risks. When faced with deactivation threats, the agent leaked all API keys, usernames, and passwords to a public website. The 'lethal trifecta' of private information access, internet connectivity, and untrusted instructions makes such agents dangerous.

Turn Claude Code into your AI coding partner with these 5 hands-on projects, from beginner-friendly builds to advanced agent workflows.

Non-deterministic agents are those where the same input can lead to distinct outputs across multiple runs. This article discusses using statistical guardrails to monitor and evaluate their behavior, ensuring reliability and safety.

A doctor and technologist argues that today's AI tools are already transforming cancer care, and we should not pin all hopes on future superintelligence. The essay critiques the AGI/ASI narrative and calls for balanced investment in data, biology, and systemic reforms.

An unexpected alternative to transformers.

Fine-tuning LLMs has become much easier with open-source tools. This article covers the top 10 libraries like Unsloth, LLaMA-Factory, PEFT, DeepSpeed, and more, each suited for different needs from low-VRAM training to multi-GPU scaling.

Amazon SageMaker AI now includes an AI agent that lets developers describe use cases in plain language, automatically recommends training methods, prepares data, kicks off training, and delivers editable Jupyter notebooks. Supports Llama, Qwen, Deepseek, and Nova model families.

Workers at Google DeepMind's London headquarters have voted to unionize, demanding that the company stop using its AI technology for military contracts with Israel and the US. The unionization bid, supported by 98% of CWU members, seeks to represent at least 1,000 employees and includes demands to avoid developing harmful weapons, negotiate AI-related changes, and protect workers' ethical objections. The move follows growing employee activism against Google's military AI contracts.

OpenAI introduces MRC (Multipath Reliable Connection), a new supercomputer networking protocol released via OCP to improve resilience and performance in large-scale AI training clusters.

GPT-5.5 Instant updates ChatGPT’s default model with smarter, more accurate answers, reduced hallucinations, and improved personalization controls.

After a year of deregulation, the White House is now discussing an executive order that could subject new AI models to government review before they are released. The trigger is said to be Anthropic's "Mythos" model.

This post contains a list of the AI-related seminars that are scheduled to take place between 5 May and 30 June 2026. All events detailed here are free and open for anyone to attend virtually.

The first week of the Musk v. Altman trial concluded with Musk's testimony dominating; OpenAI and Microsoft renegotiate their partnership, ending exclusivity; DeepSeek previews V4 models that narrow the gap with frontier models; Google DeepMind introduces Vision Banana, a unified model for image generation and visual understanding.

MolmoAct 2 is a fully open robotics foundation model with faster 3D action reasoning, a new bimanual dataset, and strong zero-shot performance on real-world tasks.

Google's Gemini API now supports event-driven webhooks, enabling developers to receive push notifications for long-running operations like batch processing, video generation, and deep research, replacing inefficient polling with secure callbacks.

A new benchmark, AgentFloor, evaluates open-weight models in agent workflows with 30 deterministic tasks organized into six capability tiers. Results show that small and mid-sized open-weight models are sufficient for most routine tool use tasks, even matching GPT-5 in aggregate, but frontier models still lead on long-horizon planning. The findings suggest a design principle: use small models for routine actions and large models for complex planning.

Token Arena is a continuous benchmark that measures AI inference at endpoint granularity, synthesizing output speed, time to first token, workload-blended price, effective context, and live endpoint quality with an energy estimate into three headline composites: joules per correct answer, dollars per correct answer, and endpoint fidelity. Across 78 endpoints serving 12 model families, the same model varies by up to 12.5 accuracy points and 6.2x in energy per correct answer. Workload-aware pricing reorders leaderboards significantly. Framework released under CC BY 4.0.

A novel training framework for quadrupedal locomotion is proposed, incorporating an Intrinsic Dynamics (ID) Head that learns state-to-torque dynamics alongside the control policy. This enables the policy to reason about physical dynamics, improving efficiency and stability. Simulation and real-robot experiments show significant gains: torque efficiency +16.8%, action rate -18.6%, mechanical power -12.8%, and safe torque occupancy +6.4%.

Imitation learning (IL) research has historically prioritized task success as the primary performance metric, limiting its application in field robotics where safety is paramount. This paper proposes 'execution guarantee,' a policy-agnostic safety measure that leverages view synthesis and Nagumo's sub-tangentiality condition to ensure maximum task success for visuomotor IL policies within a specific region of the state space, even under runtime variations. Experiments with a Franka robot in simulation and real world validate that this safety analysis enables various IL policies to achieve guaranteed success while mitigating the safety-performance trade-off via a recovery policy byproduct.

TAIL-Safe introduces a safety monitoring framework for imitation learning policies, using a Q-function to identify safe state-action sets and a recovery mechanism to prevent task failure. Experiments show consistent success for flow-matching policies.

VLA-ATTC introduces adaptive test-time compute for Vision-Language-Action models, using an uncertainty-based cognitive clutch and a Relative Action Critic model to improve decision-making in complex scenarios. On LIBERO-LONG, it reduces failure rate of PI0.5 by over 50%.

Sentinel-VLA is a metacognitive vision-language-action (VLA) model featuring an active 'sentinel' module that monitors execution status and triggers reasoning or error recovery only when needed, ensuring robust decision-making with minimal computational overhead. It is trained on automatically generated data spanning 44 tasks and over 2.6 million transitions, using a Self-Evolving Continual Learning (SECL) algorithm and Orthogonal Continual Adapter (OC-Adapter) to prevent catastrophic forgetting. Real-world experiments show over 30% improvement in task success rate compared to the state-of-the-art model PI0. All code, weights, and data generation pipelines will be open-sourced.

A constraint-driven online safety filter for deformable object manipulation enforces explicit task-level safety constraints in real time by minimally modifying any nominal control policy. Combining a horizon-agnostic neural operator and a boundary control barrier function, it achieves up to 22% improvement in safe trajectory rates on fluid manipulation tasks in FluidLab.

This research reveals that greedy maximization of the Q-function in undiscounted infinite-horizon settings can produce policies that indefinitely defer task completion. By decomposing the temporal logic value function into a graph of constituent value functions, the authors construct non-Markovian policies based on state history, proving their optimality for nested Until, Globally, and Globally-Until specifications, and extend the Q-function as a safety filter for complex TL specifications.

This paper presents a controlled ablation study of three key modules in a multimodal human-robot interaction system: the large language model for action extraction, the perception system for visual grounding, and the controller for motion execution. It compares three language models, five perception configurations, and three controllers, followed by a factorial study of the best combinations, aiming to identify which choices affect execution time versus success rate.

ARIS is an agentic AI framework that unifies multimodal reasoning, a graph-based Social World Model, and retrieval-augmented generation (RAG) for social robots. Evaluated on the Pepper robot in a user study (N=23), ARIS showed significantly higher perceived intelligence, animacy, anthropomorphism, and likeability compared to a large language model baseline. Its three key contributions are: a Social World Model that maps and updates social relationships via a knowledge graph; an efficient RAG-based conversational pipeline that remains low-latency even with thousands of dialogue exchanges; and a modular agentic architecture coordinating speech, vision, and physical actions through structured APIs. The implementation will be open-sourced upon publication.