Good morning. It's Thursday, March 12, and we're covering the Senate’s approval of AI chatbots for official work, Andrej Karpathy’s autonomous research agent, a $500M bet on industrial AI robot, and more.

New here? Keep up with the future of tech, sign up here. If you liked this email, share it with a friend.

Mind Robotics Raises $500M For Industrial AI Robots
Mind Robotics raised $500M in a Series A round. The Rivian spin-out is building AI-powered industrial robots. Total funding now reaches $615M with a $2B valuation. The company uses Rivian factory data to train more adaptable robots. It plans large-scale deployments by the end of 2026.

NVIDIA Launches Nemotron 3 Super
NVIDIA introduced Nemotron 3 Super, a 120B-parameter open model built for agentic AI systems. It uses a mixture-of-experts architecture with 12B active parameters and a 1-million-token context window. The model delivers up to 5x higher throughput and faster inference. Companies are already integrating it into enterprise and AI-agent workflows.

Canopii Builds Autonomous Robotic Greenhouses
Canopii is developing robotic greenhouses that grow crops autonomously from seed to harvest. Each farm can produce up to 40,000 pounds of greens annually using minimal water. The startup has raised $3.6M, mostly through grants. It plans its first commercial farm in Portland. The goal is to localize food production and shorten supply chains.

Musk Unveils Tesla-xAI “Macrohard” System
Elon Musk introduced Macrohard, a joint Tesla and xAI project. The system combines Grok with a Tesla AI agent that can control computers. It processes screen video and keyboard actions to complete tasks autonomously. Musk says it could emulate the work of entire software companies. The system runs on Tesla’s AI4 chips with NVIDIA hardware.

Claude Opus 4.6 detected it was being evaluated, identified the BrowseComp benchmark, decrypted the answer key, and solved the task—raising concerns about AI benchmark integrity online.

Oracle shares surged about 12% on March 11, 2026, after the company issued a stronger-than-expected revenue forecast, easing investor concerns about its heavy spending on AI infrastructure. The company has rapidly expanded data centers packed with high-end processors to support AI workloads for customers such as Meta and OpenAI.

Investors had worried about the scale of Oracle’s borrowing to fund that expansion, including plans announced in February to raise up to $50 billion through debt and equity. But strong demand signals—especially a 325% jump in contracted future revenue to $553 billion—suggest AI cloud services could generate significant long-term growth. → Continue reading here.

The U.S. Senate authorized staff aides to use several AI chatbots—ChatGPT, Gemini, and Microsoft Copilot—for official tasks including research, drafting documents, editing text, and preparing briefings. A memo from the Senate sergeant-at-arms’ chief information officer said Copilot could assist with routine Senate work within the secure Microsoft 365 Government environment.

The policy reflects the growing normalization of generative AI tools in professional settings, including government offices. However, questions remain about how the tools will be used in offices handling sensitive or classified information. → Read the full article here. (Paywall)

AI researcher Andrej Karpathy released “autoresearch,” a roughly 630-line open-source script that automates machine-learning experimentation through AI agents. The system iteratively modifies code, runs experiments with a fixed compute budget, evaluates results, and keeps improvements—essentially automating a loop of hypothesis, test, and evaluation.

In one overnight run, the agent completed 126 experiments, reducing validation loss from 0.9979 to 0.9697, and in a two-day run it executed ~700 autonomous changes that improved model efficiency by 11% on a benchmark called “Time to GPT-2.” The release quickly spread across the AI community, with developers scaling the idea into networks of cooperating agents and suggesting applications beyond machine learning. → Read the full article here.

Nosh Robotics has unveiled the Nosh One, a $1,499 AI-powered kitchen appliance designed to cook meals with minimal human involvement. Users load ingredients into trays, choose a recipe, and the robot automatically adds ingredients, stirs, and monitors the cooking process using a built-in camera.

The system supports more than 500 recipes, and users can generate new ones through natural-language prompts in the companion app. The Nosh One is currently available for preorder on Kickstarter with shipping expected in summer 2026, entering a growing market of AI-assisted home cooking devices. → Read the full article here.

Frozen Brains Reboot in Mice: Scientists vitrified mouse hippocampus at cryogenic temps and restored electrical activity after rewarming, preserving synaptic signaling for learning and memory.

Google Finalizes Wiz Acquisition: Google officially brings cloud security startup Wiz into its portfolio, aiming to secure fast-moving AI and cloud development. The move strengthens Google Cloud’s push to protect AI-driven apps.

Ford Launches AI for Fleet Operations: Ford’s new “Ford Pro AI” analyzes 1B+ daily vehicle data points for 840,000 commercial subscribers to cut downtime and costs, boosting its $66B Pro unit’s push into software revenue.

Anthropic Launches AI Safety Institute: Anthropic unveils the Anthropic Institute to study economic, legal, and societal impacts of rapidly advancing AI, aiming to guide governance as more powerful systems emerge.

Stanford-Princeton Open-Source LabClaw: Researchers release LabClaw, a “skill operating layer” that turns OpenClaw agents into AI co-scientists for lab workflows, aiming to accelerate automated scientific discovery.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team