Holo3.1 AI agents and Jensen Huang on layoffs

Holo3.1 launches a family of quantized computer‑use agents that run quickly on‑device, supporting desktop, web and mobile environments and integrating with diverse agent frameworks. The release adds FP8, Q4 GGUF and NVFP4 checkpoints plus compact 0.8B‑9B models for low‑cost, private deployment, boosting robustness and performance across platforms.

The authors demonstrate that current large language models routinely overestimate their competence, lacking the ability to decide when to solve or defer a task. By framing capability self‑assessment as a policy‑learning problem and applying reinforcement learning, they markedly improve LLMs’ self‑awareness without sacrificing original performance, with strong out‑of‑distribution generalization.

The Consilium Protocol, built on Byzantine Fault Tolerance, structures collaboration among multiple language models by treating their disagreements as epistemic signals. By assigning engineered cognitive personas, the system achieves cost‑effective, bias‑aware reasoning, matching frontier model performance at a fraction of the price. The paper releases the protocol under an MIT license for independent verification.

CAST introduces answer‑free self‑distillation for GRPO, using a stop‑gradient self‑teacher and bidirectional advantage flipping to provide dense token‑level feedback despite sparse outcome rewards. Experiments on math reasoning show improved RLVR training while keeping verifier‑grounded objectives.

In a May 2026 interview, Nvidia CEO Jensen Huang rebuked CEOs who cite AI as a reason for firing staff, calling the narrative “lazy.” He argued AI will generate far more jobs than it displaces, citing a five‑layer AI industry model and decades‑long tech history. The remarks highlight a growing debate on AI’s impact on employment.

Perplexity unveiled Search as Code (SaC), an SDK that lets AI models directly program search pipelines for task‑specific retrieval. By exposing search primitives as code, agents can orchestrate thousands of queries within minutes, turning search from a monolithic service into a flexible, programmable component for complex workflows.

OpenAI broke ground on a 1GW data center campus called The Barn in Saline, Michigan, pledging that infrastructure costs won’t be passed to local ratepayers and that a closed‑loop cooling system will protect water resources. The project promises over 2,500 union construction jobs, $10 million for community facilities, and up to $45 million in AI training credits for Michigan students.

The U.S. Department of Commerce issued new guidance requiring export licenses for advanced AI chips sold to any entity headquartered in China, even if located abroad, effectively closing a loophole that let Nvidia's Blackwell and AMD's AI processors flow to Chinese firms overseas. The move targets hundreds of thousands of chips allegedly exported without oversight.

OpenAI is calling for an international institute dedicated to youth AI safety, to set standards, share evidence, and provide guidance. Discussed at the G7 summit, the plan builds on the Youth AI Safety Institute and educator collaborations to ensure age‑appropriate, safe AI access for youth worldwide.

The Hugging Face open‑source team has revived PapersWithCode, now offering a searchable list of CVPR 2026 conference papers with linked code, data, and demos. This enables quick tracking of state‑of‑the‑art methods across computer‑vision tasks.

Open-source Khala Music AI, previously limited to NVIDIA GPUs, has been ported to run on Apple Silicon via vanilla PyTorch and Metal Performance Shaders. The Mac can now generate high‑quality songs locally, with examples and model weights provided in the author's blog.

Mistral unveiled its Search Toolkit in public preview, an open-source framework that consolidates data ingestion, retrieval, and evaluation under a single interface for production AI pipelines. The modular system aims to cut integration time, support diverse enterprise data sources, and enable rigorous retrieval evaluation, accelerating RAG and enterprise search deployments.