AI Daily — June 1, 2026
2026-06-01
TODAY'S NEWS
huggingface.co
Welcome NVIDIA Cosmos 3: The First Open Omni-model for Physical AI Reasoning and Action
NVIDIA has released Cosmos 3, an open omni-model designed for physical AI reasoning and action, targeting robotics and embodied AI applications. The model is notable for being the first open model in this class that combines perception, reasoning, and action generation in a unified architecture. Weights and model details are hosted on Hugging Face, making it accessible for researchers building physical AI systems.
rss.arxiv.org
Structured interactions improve distributed coordination beyond model scaling in a real-world multi-robot system
A controlled experiment with 10 physical robots across 60 runs found that switching from fully connected to modular hierarchical communication topology improved normalized task performance by 47 points, while doubling neural network hidden size yielded at most 9 points under matched hardware budgets. The study uses a transport-and-mapping task and mixed-effects model comparisons to confirm that interaction structure dominates model scale as a performance driver. This is a rare real-world empirical result challenging the default assumption that bigger models are the primary lever for multi-robot coordination.
LAST WEEK'S TOP STORIES
Anthropic raises $65B Series H at ~$965B valuation
Anthropic closed a $65B funding round at a ~$965B valuation, cementing its position as one of the most highly valued private companies ever and signaling continued massive capital concentration in frontier AI ahead of a likely IPO.
Claude Opus 4.8 becomes new default, adds dynamic multi-agent workflows
Anthropic made Claude Opus 4.8 the default model with high-effort reasoning on by default and introduced dynamic workflows that let Claude orchestrate dozens to hundreds of subagents, meaningfully expanding the scope of agentic task automation.
ITBench-AA: Frontier Models Score Below 50% on the First Benchmark for Agentic Enterprise IT Tasks
IBM Research and Artificial Analysis released the first benchmark for agentic AI on real enterprise IT operations tasks, with all frontier models scoring below 50%, establishing a rigorous public bar that exposes a major capability gap for production enterprise automation.
Mechanistic origins of catastrophic forgetting: why RL preserves circuits better than SFT
Researchers provided a mechanistic, circuit-level explanation for why RL fine-tuning causes less catastrophic forgetting than SFT on Qwen2.5-3B-Instruct, offering practitioners a principled basis for choosing fine-tuning strategies when preserving general capabilities is critical.
Why LLMs Fail at Causal Discovery and How Interventional Agents Escape
A theoretical proof shows that standard LLM training regimes (SFT, DPO, ICL) are fundamentally incapable of reliable causal graph discovery, with interventional agents proposed as the only viable path forward — a foundational result for scientific reasoning and autonomous research agent design.