Wednesday, June 3, 2026

NVIDIA’s Physical AI Bet Widens

NVIDIA’s Physical AI Bet Widens

Today’s Overview

Good morning, NVIDIA is pushing hard on physical AI with Cosmos 3 and a massive open-weights Nemotron release, while the White House is choosing a lighter-touch path for frontier model reviews. Perplexity also has a sharp idea for making search feel less like a query box and more like programmable infrastructure. Let’s dive in.

Top Stories

Trump Makes Frontier AI Reviews Voluntary

Trump signed an executive order asking AI labs to voluntarily give the federal government access to certain frontier models for a 30-day security review before release. The order backs away from an earlier 90-day requirement, rules out mandatory licensing or permitting for new models, and tells the DOJ to prioritize AI-powered computer intrusion cases.

  • The order directs agencies to create a classified benchmarking process for deciding which advanced models qualify as covered frontier models.
  • The voluntary access framework must include protections for confidentiality, cybersecurity, insider risk, and IP when developers share covered models with the government.
  • A new AI cybersecurity clearinghouse is meant to coordinate vulnerability discovery and patching with industry and critical infrastructure operators.

NVIDIA Debuts Nemotron 3 Ultra

NVIDIA released Nemotron 3 Ultra, a 550B-parameter open-weights model with 55B active parameters. NVIDIA says it is the most intelligent open-weights model from the U.S., with NVFP4 quantization planned for higher inference performance.

  • Artificial Analysis evaluated the model using BF16 weights before the planned NVFP4 release path.
  • Nemotron 3 Ultra is described as having about 90% sparsity which makes its 550B total parameter count much larger than its active footprint.
  • Its reported U.S. open-weights lead still trails the broader open frontier, with Kimi K2.6 at 54 on the same intelligence index.

NVIDIA Launches Cosmos 3 for Physical AI

NVIDIA launched Cosmos 3, an open physical AI foundation model with native vision reasoning and multimodal generation. The model is built to give developers a pretrained base for robotics, autonomous vehicles, and vision AI systems that need less data and lower training costs.

  • The model is positioned for three roles: vision language model, world model, and action backbone for training and evaluating physical AI systems.
  • NVIDIA says Cosmos 3 ranks first among open models across physical AI benchmark leaderboards covering world generation, action policy, and vision understanding.
  • The lineup includes Cosmos 3 Super and Nano now, with Cosmos 3 Edge planned for real-time inference at the edge.

Research & Analysis

Perplexity Reframes Search as Code Generation

Perplexity introduced Search as Code, an SDK-based architecture that lets models directly orchestrate search pipelines instead of relying on a fixed search interface. The approach is designed for complex agentic search tasks where models need to plan, retrieve, filter, verify, and adapt as they go.

  • In a CVE vendor-advisory case study, SaC reached 100% accuracy while cutting token use from 288.7K to 42.9K versus the non-SaC baseline.
  • The system can generate programs that run thousands of discrete operations through the search SDK instead of making a narrow series of calls.
  • On WANDR, Perplexity says SaC beat the next-best evaluated system by 2.5x while the benchmark remained difficult even for SaC.

JetBrains Details Mellum 2 for Coding AI

JetBrains introduced Mellum 2, an open-weight 12B-parameter Mixture-of-Experts language model built for software engineering. The model targets code generation and editing, debugging, multi-step reasoning, tool use, function calling, agentic coding, and conversational programming assistance.

  • Mellum 2 uses 2.5B active parameters per token despite its 12B total parameter count.
  • Its architecture includes 64 experts with 8 active plus grouped-query attention, sliding window attention, and multi-token prediction.
  • The training pipeline used about 10.6 trillion tokens and extended the base model to a 128K context window.

Trending AI Tools

  • Paste MCP & AI Tools An infinite clipboard for Claude, Codex, and other AI tools.

  • Bonsai Image 4B Compact local image-generation models, including 1-bit and ternary variants that can run on an iPhone.

Quick Hits

  • MiniMax M3 will release model weights and a technical report within 10 days, with a 1M-token context window and API pricing listed up to 512,000 input tokens.

  • Microsoft Build brought seven new MAI models, the Scout always-on agent, Project Solara, Surface RTX Spark Dev Box, and the Majorana 2 quantum chip.

  • Inherent Labs came out of stealth with $50M to build Faraday, an AI science platform pairing researchers with self-improving agents.

  • Martin Scorsese and Black Forest Labs signals high-profile Hollywood interest in AI for preproduction, with FLUX used for storyboarding rather than generated actors, sets, or footage.

  • OpenAI on Amazon Bedrock shows how to build production workflows with OpenAI models on Bedrock using the Responses API.

  • Anthropic IPO filing confidentially submitted a draft S-1 for a proposed IPO, with pricing and share counts still unset.

Keep reading for free

Enter your email. If you're already subscribed, we'll send a sign-in code. If not, you'll subscribe in the next step.

Free access. Subscribe once, then use the same email on future issues.

Free to read. Subscription just unlocks the full issue.