Good morning. It's Tuesday, March 17, and we're covering AI humanoid battlefield robots, Wall Street’s AI pricing confusion, how Pokémon Go data is training robots to navigate the real world, and more.

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

NVIDIA Unveils NemoClaw for AI Agent Security
NVIDIA introduced NemoClaw, an open-source stack that adds privacy and security controls to the OpenClaw agent platform. Developed with OpenClaw creator Peter Steinberger, the platform adds enterprise-grade controls for how AI agents access data, invoke tools, and interact with systems. The goal is to address one of the biggest barriers to deploying autonomous AI agents: security vulnerabilities such as prompt injection and uncontrolled system access.

NVIDIA Targets $1 Trillion From Blackwell and Rubin
At GTC, NVIDIA CEO Jensen Huang projected at least $1 trillion in orders for Blackwell and Vera Rubin chips through 2027. The forecast reflects surging demand from hyperscalers and enterprises racing to build AI infrastructure, with Blackwell already ramping production and Rubin slated as its successor. Huang framed the opportunity as part of a broader shift toward “AI factories,” where data centers are optimized to generate intelligence at scale.

Pokémon Go Players Train Massive AI Map
Niantic says images captured through Pokémon Go and its augmented-reality apps have produced a dataset of more than 30 billion real-world photos and scans. The company is using the data to train AI systems that help delivery robots visually navigate city streets. Instead of relying solely on GPS, the robots can recognize landmarks and precise locations. The project highlights how AR gaming data can power real-world robotics.

Britannica Sues OpenAI Over AI Training
Encyclopaedia Britannica and its subsidiary Merriam-Webster sued OpenAI in Manhattan federal court, alleging the company used their reference content to train ChatGPT without permission. The lawsuit claims OpenAI copied nearly 100,000 articles and that ChatGPT can produce near-verbatim summaries of Britannica material. Britannica says the summaries divert traffic from its sites. OpenAI says its models use publicly available data under fair use.

Overview of modern AI types and tools—ChatGPT, Claude, Gemini, Grok, open-source models, plus image, video, coding, and audio AI capabilities.

San Francisco startup Foundation is developing the Phantom MK-1, a humanoid robot designed to operate alongside soldiers and potentially carry human weapons in combat. The company has secured about $24 million in research contracts with the U.S. Army, Navy, and Air Force and has sent two units to Ukraine for reconnaissance testing as the war increasingly relies on autonomous systems.

Supporters argue humanoid robots could reduce troop casualties and operate in dangerous environments where humans cannot. Critics warn the technology may lower the threshold for war, raise accountability questions if machines cause civilian harm, and accelerate a global arms race in autonomous weapons. → Read the full article here.

An analysis from The Economist finds that financial markets remain uncertain about how artificial intelligence will reshape corporate winners and losers. Stocks tied to AI disruption and AI benefits have both declined recently, even as broader markets hover near record highs—suggesting investors lack a clear model for pricing the technology’s impact.

Company narratives are shifting rapidly: firms once seen as vulnerable are rebounding while others touted as AI winners are falling sharply. History suggests this confusion is typical during major technological transitions, when markets swing between hype and skepticism before the economic effects become clearer. → Read the full article here. (Paywall)

Moonshot AI announced Attention Residuals (AttnRes), a new neural network design that replaces traditional fixed residual connections with attention-based aggregation across earlier layers. Instead of simply adding outputs from previous layers, the method allows models to selectively attend to prior representations, dynamically retrieving useful information during computation.

The company says Block AttnRes matches the loss of a baseline trained with 1.25× more compute and serves as a practical drop-in replacement with marginal overhead. AttnRes was tested within the Kimi Linear architecture (48B parameters total, 3B activated), where it reportedly produced consistent downstream performance gains. → Read the full article here.

ByteDance Halts Seedance 2.0 Expansion: The TikTok owner paused the global rollout of its AI video generator after Disney and Paramount Skydance sent cease-and-desist letters over alleged copyrighted training data.

Meta Secures $27B AI Cloud Deal: Nebius signed a five-year AI infrastructure agreement with Meta worth up to $27 billion. The deal includes $12 billion in dedicated compute capacity built on NVIDIA’s Vera Rubin platform.

Google Scraps Crowdsourced Health AI: Google removed its “What People Suggest” search feature, which used AI to summarize health tips from online discussions, as scrutiny grows over inaccurate AI-generated medical information.

Resumes Filtered by AI: Most job applications never reach a human because applicant-tracking systems screen candidates first, pushing workers to showcase AI fluency, digital portfolios, and algorithm-friendly résumés.

xAI Recruits Wall Street to Train Grok (Paywall): Elon Musk’s startup is hiring bankers, traders, and credit analysts to teach Grok financial modeling and investment strategy, targeting AI tools for professional investors.

Alibaba Creates AI Token Hub: The company is launching “Alibaba Token Hub,” a new unit unifying Qwen models, apps, and AI products under CEO Eddie Wu to accelerate enterprise adoption and monetize AI services.

Use this prompt after you've already formed an opinion or made a plan. Paste your draft decision, strategy, or assessment and add:

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team