Good morning. It’s Wednesday, March 19th.

On this day in tech history: in 1997, Toshiba released the first consumer DVD players in the United States.

In today’s email:

• NVIDIA’s GTC 2025 Conference

• 5 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

_{In partnership with TAU}

Breakthrough in AI: TAU Language

With its breakthrough software specification language, Tau makes huge progress in software synthesis and mechanized deductive reasoning.

Tau automatically synthesizes correct programs directly from logical requirements​, ensuring provable correctness by design.

This overcomes past limitations in expressiveness and algorithms. Tau’s logic is highly expressive yet fully decidable, leveraging novel algorithms to handle reasoning efficiently.​

The result is a better approach to software development that ensures software reliability, with provably correct programs that can adapt to requirements.

Join the Community Telegram, Tau Community Forum and explore the Tau Language on GitHub to learn more.

Nvidia Shows Off AI Superchips, Humanoid Robots, and Self-Driving Tech

At GTC 2025, NVIDIA CEO Jensen Huang made several AI and robotics reveals:

Headlining the drop is Blackwell Ultra, a silicon sledgehammer that delivers 1.5 times the speed of its predecessor. Paired with a Grace CPU, it powers the GB300 superchip, a dual-GPU system processing 1,000 tokens per second on DeepSeek’s R1 model—ten times faster than Hopper. Nvidia also previewed its next-gen chips, Vera Rubin and Feynman, set for 2026 and 2028.

cuOpt, Nvidia’s AI-powered optimization engine, is now open-source, making real-time decision-making accessible to developers. GPU-accelerated and 20 times faster than traditional methods, it integrates with Python and Nvidia’s RAPIDS framework, allowing businesses to optimize logistics and resource allocation without enterprise costs.

GR00T N1 makes humanoids less robotic: Isaac GR00T N1 pushes humanoid robotics forward, combining reflex-driven actions with deeper reasoning. With GR00T Blueprint generating synthetic training data and Newton refining movement physics, companies like 1X Technologies, Boston Dynamics, and Disney are already integrating it into their robots.

Personal AI supercomputers hit the desktop: Nvidia is bringing raw AI compute within arm’s reach. DGX Spark, a compact inferencing box, runs on the GB10 Grace Blackwell Superchip, while the larger DGX Station packs the GB300 Blackwell Ultra Superchip with 20 petaflops of performance. Asus, Dell, and Lenovo will launch versions this year.

Halos sets the bar for AV safety: Nvidia’s Halos framework stress-tests self-driving AI through simulation and real-world validation, with GM already integrating it into future vehicles. The AI Systems Inspection Lab—the first ANSI-accredited AV safety program—will oversee compliance.

AI watermarking and open-source data fuel transparency: Nvidia is embedding Google DeepMind’s SynthID watermarking into Cosmos-generated videos. It also released a 15-terabyte open-source dataset with 320,000 robotic trajectories and AV data from 1,000 cities, accelerating AI safety research.

Watch Jensen Huang's Full NVIDIA keynote here.

• GPUs go biological: brain-on-a-chip merges biology with AI compute

• Google brings a 'canvas' feature to Gemini, plus Audio Overview

• Google plans to release new 'open' AI models for drug discovery

• Google introduces interactive Mindmaps in NotebookLM for AI-driven learning

• AI-powered filmmaking takes shape with Stability AI’s virtual camera model

• Mistral’s open model outsmarts GPT-4o mini with fewer parameters

• xAI acquires Hotshot, absorbs video gen tech into Colossus

• Perplexity’s latest ad roasts Google’s ‘glue on pizza’ AI fail

• AI search is starting to kill Google’s ‘ten blue links’

• OpenAI to start testing ChatGPT connectors for Google Drive and Slack

• OpenAI exec spins off into found materials science startup

• Mac Studio with M3 Ultra runs massive DeepSeek R1 AI model locally, no GPU cluster required

• Google’s Gemini AI is really good at watermark removal

• Roblox’s new AI model can generate 3D objects

• Smart glasses will be future of computing, Meta executives say

• Paint update begins rolling out to Windows Insiders

• Amazon to introduce premium Alexa-powered hardware

• Nvidia expected to reveal details of latest AI chip at conference ​​

• Tencent expands AI push with open-source 3D generation tools

• Nearly half of U.S. adults believe LLMs are smarter than they are

• EXAONE 32B beats DeepSeek R1 671B, smaller model tops OpenAI o1 mini

• Anthropic teases Claude 3.7 Sonnet MAX with "Maximum Intelligence" upgrade

• Google acquires cybersecurity firm Wiz for $32 billion

• Adobe's new AI agents can make personal websites for your customers

• Meta announced that its Llama open-source models officially reached 1B downloads this month

• AI coding startup Cognition AI behind Devin hits $4 billion valuation

• Unitree G1 humanoid robot is the world’s first sideflipping robot

• Watch: Deep Robotics' auto-edge perception 

📄 DropletVideo: A Dataset and Approach to Explore Integral Spatio-Temporal Consistent Video Generation

📄 Being-0: A Humanoid Robotic Agent with Vision-Language Models and Modular Skills ​​

📄 SmolDocling: An ultra-compact vision-language model for end-to-end multi-modal document conversion  

📄 RWKV-7 "Goose" with Expressive Dynamic State Evolution

📄 Impossible Videos 

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on 𝕏!