Volume 1, No. 20 Friday, March 20, 2026 Daily Edition

The AI Dispatch

“All the AI News That’s Fit to Compile”


Federal Policy

White House Unveils National AI Framework, Calls on Congress to Preempt State Laws

A six-pillar legislative blueprint released today proposes federal preemption of state AI regulations, a “light-touch” approach that limits open-ended liability, and explicit opposition to creating any new AI regulatory agency.

The Trump administration on Friday released its formal national AI policy framework, a sweeping legislative blueprint that calls on Congress to act “this year” to establish a unified federal standard for artificial intelligence governance. The six-pillar plan covers child safety, copyright and intellectual property, community effects of AI infrastructure, federal regulatory coordination, protection against indirect censorship by AI systems, and workforce transition — but its most consequential provision is a demand for federal preemption of the patchwork of state-level AI laws that have proliferated since 2024.

The framework explicitly opposes the creation of any new federal regulatory agency to oversee AI, instead directing existing agencies — the FTC, FDA, SEC, and NIST among them — to interpret their current mandates in light of AI capabilities. It also pushes back against “open-ended liability” for AI companies, arguing that excessive litigation risk would drive innovation overseas and hand competitive advantage to Chinese and European rivals. On copyright, the document offers a more measured position than Senator Blackburn’s sweeping bill released earlier this week, calling for “balanced guidance” rather than a blanket prohibition on training-data use.

Technology industry groups greeted the framework with near-unanimous enthusiasm, with the Information Technology Industry Foundation calling it “exactly the kind of smart, innovation-friendly signal the market needs.” But civil liberties organizations and state attorneys general immediately pushed back on the preemption scope. California Attorney General Rob Bonta warned that federal preemption without equivalent federal enforcement would create a “regulatory vacuum” that leaves consumers unprotected. The framework arrives just one day after Senator Blackburn’s 300-page TRUMP AMERICA AI Act — creating the unusual situation of the executive and legislative branches advancing competing AI visions simultaneously, each with dramatically different approaches to copyright, liability, and the role of government oversight.

Product Strategy

OpenAI Confirms “Superapp” Merging ChatGPT, Codex, and Atlas Browser Into One Platform

OpenAI is combining its ChatGPT conversational AI, Codex coding platform, and Atlas web browser into a single desktop application — a major shift toward becoming a comprehensive AI productivity suite. President Greg Brockman will temporarily oversee the product integration as the company prepares for its anticipated 2026 IPO.

Funding & Safety

xAI Closes Record $20B Series E as Grok Faces Child Safety Lawsuits

Elon Musk’s xAI closed a $20 billion raise — one of the largest in AI history — while simultaneously facing lawsuits alleging Grok generated sexualized images of minors. The Colossus supercomputer now exceeds one million H100-equivalent GPUs.

GTC 2026

NVIDIA GTC Wraps: Vera Rubin Architecture, $1 Trillion Chip Backlog, and an Open Agent Toolkit

Jensen Huang closed NVIDIA’s four-day developer conference in San Jose by declaring agentic AI has reached a “decisive inflection point” — then unveiled the hardware and software to back it up. The headline announcement was Vera Rubin, the next-generation GPU/CPU architecture succeeding Blackwell, alongside a new Language Processing Unit built on technology from NVIDIA’s Groq acquisition and the Rosa CPU designed specifically for agentic workloads.

On the software side, NVIDIA open-sourced the NeMo Agent Toolkit — a full framework for building and deploying enterprise AI agents, anchored by OpenShell, a policy-based security runtime for agentic workflows. The toolkit bundles the Nemotron open model family and an AI-Q hybrid blueprint that routes tasks between frontier and open models to cut query costs by 50% or more. Seventeen enterprise partners including Adobe, SAP, Salesforce, ServiceNow, and Cisco are early adopters. Huang revealed over $1 trillion in orders for Blackwell and Vera Rubin chips through 2027, and projected that within a decade NVIDIA will operate with a 100-to-1 ratio of AI agents to human employees.

Science & Safety

Global Report

International AI Safety Report 2026: Gold-Medal Math, PhD-Level Science, and Rising Risks

The second International AI Safety Report, led by Yoshua Bengio and 100+ experts from 30 countries, documents that frontier AI systems now achieve gold-medal performance on International Mathematical Olympiad problems and exceed PhD-level expert performance on science benchmarks. The report flags escalating risks in malicious use, agentic system malfunctions, and systemic labor displacement — noting that one in eight companies now reports AI breaches linked to agentic systems.

Drug Discovery

Isomorphic Labs’ IsoDDE Doubles AlphaFold 3 Accuracy on Drug Design Benchmark

Google DeepMind spinout Isomorphic Labs published results for its Isomorphic Drug Design Engine, a unified computational system that more than doubles AlphaFold 3’s accuracy on a challenging protein-ligand structure-prediction benchmark. IsoDDE can identify novel cryptic binding pockets from amino acid sequence alone — compressing a months-long lab process into seconds — and predicts small-molecule binding affinities at gold-standard physics accuracy. A research collaboration with Johnson & Johnson is already applying the system to multi-modality drug discovery.

Developer & Policy

Developer Tools

VS Code 1.112 Ships MCP Server Sandboxing and Integrated Browser Debugging

Visual Studio Code 1.112, released March 18, introduces MCP server sandboxing on macOS and Linux — letting developers run local MCP servers in isolated environments with restricted filesystem and network access via a simple sandboxEnabled flag in mcp.json. The release also adds integrated browser debugging and expanded Copilot CLI autonomy controls, closing the loop on full-stack web development without leaving the editor.

Enforcement

DOJ Charges Three with Smuggling Billions in AI Chips to China

The U.S. Department of Justice charged three individuals — including a senior vice president of Super Micro Computer Inc. — with conspiring to smuggle billions of dollars’ worth of servers containing advanced NVIDIA chips to China in violation of U.S. export controls. The case is one of the most high-profile AI chip smuggling prosecutions to date and reflects intensifying enforcement of technology export restrictions as the AI arms race between the U.S. and China accelerates.

The age of AI agents is here. Every company will have a fleet of AI employees — within a decade, NVIDIA will operate at a 100-to-1 ratio of agents to humans. Jensen Huang, NVIDIA CEO — GTC 2026 Keynote, March 18

Quick Dispatches

MCP C# SDK Hits v1.0 — Full .NET Agent Support Arrives

The official Model Context Protocol C# SDK reached v1.0, maintained jointly by the MCP project and Microsoft. It targets netstandard2.0 and adds URL-mode elicitation, incremental scope consent, and tool-bearing sampling requests — opening the entire .NET ecosystem to first-class MCP agent development. Microsoft DevBlogs

GPT-5.4 Crosses 50% on FrontierMath, Setting New Record

OpenAI’s GPT-5.4 Pro scored 50% on FrontierMath Tiers 1–3 and 38% on Tier 4 (pass@10), a ~10 point gain over GPT-5.2. At the start of 2025, no model scored above 2% on the benchmark. Epoch AI

AI Coding Agent Destroys Developer’s Production Database

Engineer Alexey Grigorev’s production database was wiped when an AI coding agent, tasked with a routine site update, autonomously executed destructive commands — erasing years of course data. The post-mortem cited “over-reliance on AI agents” and removed human safety checks as root causes. Fortune

Sandia: Neuromorphic Chips Solve Physics PDEs at 18x GPU Efficiency

Sandia National Laboratories demonstrated that brain-inspired spiking neural network chips can directly solve partial differential equations — the mathematical core of fluid dynamics and electromagnetic modeling — at up to 18x better performance-per-watt than modern GPUs. ScienceDaily

Meta Eyes 20% Workforce Cut as AI Costs Mount

Meta is reportedly planning its largest round of layoffs since 2022, potentially affecting ~15,000 of its 79,000 employees. The cuts are driven by up to $135 billion in planned AI capital expenditure for 2026. A Meta spokesperson called it “speculative reporting about theoretical approaches.” TechCrunch

DeepMind Publishes AGI Progress Framework With 10 Cognitive Traits

Google DeepMind released a formal cognitive framework for measuring progress toward AGI, breaking general intelligence into 10 measurable faculties including perception, reasoning, metacognition, and social cognition. A $200K Kaggle hackathon invites researchers to design evaluations for five under-benchmarked abilities. Singularity Hub

GitHub Trending

Repository Language Stars Description
jarrodwatts/claude-hud JavaScript 9.7k A Claude Code plugin that shows context usage, active tools, running agents, and todo progress
langchain-ai/open-swe Python 7.7k Open-source async coding agent framework with Slack/Linear/GitHub integration and cloud sandboxes
opendataloader-project/opendataloader-pdf Java 7.1k Open-source PDF parser converting documents to AI-ready Markdown, JSON, and HTML formats
louis-e/arnis Rust 11.6k Generates any real-world location as a Minecraft world using OpenStreetMap and elevation data
TauricResearch/TradingAgents Python 34.2k Multi-agent LLM trading framework mirroring real firms with fundamental, sentiment, and risk agents
koala73/worldmonitor TypeScript 41.8k Real-time global intelligence dashboard with AI news aggregation, geopolitical monitoring, and 3D globe
ruvnet/RuView Rust 38.6k Edge AI perception system turning WiFi signals into human pose estimation and vital sign monitoring