Microsoft has released this useful tool for performing R&D with LLM-based agents.

Xiaomi's development of a SOTA audio reasoning model leverages DeepSeek's GRPO RL algorithm, achieving a 64.5% accuracy on the MMAU benchmark in just one week.

The breakthrough involves applying GRPO to the Qwen2-Audio-7B model, trained on 38,000 samples from Tsinghua University's AVQA dataset, marking a significant advancement in multimodal audio understanding.

The MMAU benchmark, introduced in 2024, tests models on complex audio tasks across speech, sound, and music, with even top models like Gemini Pro 1.5 achieving only 52.97% accuracy, highlighting the challenge Xiaomi's model addresses.

The total value of real world tokenized assets has 📈 by ≈20% over the last 30 days. Data by RWA.xyz.

Total Real-World Asset (RWA) value continues to climb—over $18B is now tokenized onchain, excluding stablecoins.

A breakthrough in brain signal analysis that combines PCA and ANFIS to hit 99.5% accuracy in cognitive pattern recognition.

It could be a game-changer for #neuroscience, #BCI tech and clinical applications.

Mistral announced Small 3.1: multimodal, multilingual, Apache 2.0

Lightweight: Runs on a single RTX 4090 or a Mac with 32GB RAM, perfect for on-device applications.

Fast-Response Conversations: Ideal for virtual assistants and other applications where quick, accurate responses are essential.

Low-Latency Function Calling: Capable of rapid function execution within automated or agentic workflows.

Specialized Fine-Tuning: Customizable for specific domains.

Advanced Reasoning Foundation: Inspires community innovation, with models like DeepHermes 24B by Nous Research built on Mistral Small 3.

ByteDance Seed, Tsinghua, and UHK dropped open-sourced a new RL algorithm for building reasoning models.

DAPO-Zero-32B, a fully open-source RL reasoning model, surpasses DeepSeek-R1-Zero-Qwen-32B, and scores 50 on AIME 2024 with 50% fewer steps.

It is trained with zero-shot RL from the Qwen-32b pre-trained model.

Everything is fully open-sourced (algorithm, code, dataset, verifier, and model).

Cool research on open-source by Harvard

$4.15B invested in open-source generates $8.8T of value for companies (aka $1 invested in open-source = $2,000 of value created).

Companies would need to spend 3.5 times more on software than they currently do if OSS did not exist.

HuggingFace and IBM introduced SmolDocling an ultra-compact VLM for end-to-end multi-modal document conversion

SmolDocling is good for enterprise use cases:
- 256M parameters - cheap and easy to run locally
- Performs better than 20x larger models
- Fast inference using VLLM – Avg of 0.35 secs per page on A100 GPU.
- Apache 2.0 license

Demo.

Biggest deal in Google/Alphabet history: Google is buying Wiz for $32B to beef up in cloud security

Wiz. is an Israeli cloud security startup headquartered in New York City. The company was founded in January 2020.

This acquisition positions Google to better compete with AWS and Azure.

Anthropic is working on voice capabilities for Claude.

The company’s chief product officer, Mike Krieger, told the Financial Timesthat Anthropic plans to launch experiences that allow users to talk to Anthropic’s AI models.

NVIDIA, Google DeepMind and Disney Research are collaborating to build an R2D2 style home droid.

Jensen giving the little guy voice and gesture commands live on stage.

Robot’s name is Blue, he is so cute.

Nvidia announced GR00T N1, the world’s first open foundation model for humanoid robots

The power of general robot brain, in the palm of your hand - with only 2B parameters, N1 learns from the most diverse physical action dataset ever compiled and punches above its weight:

- Real humanoid teleoperation data.
- Large-scale simulation data: we are open-sourcing 300K+ trajectories
- Neural trajectories: SOTA video generation models to “hallucinate” new synthetic data that features accurate physics in pixels. Using Jensen’s words, “systematically infinite data”
- Latent actions: novel algorithms to extract action tokens from in-the-wild human videos and neural generated videos.

GR00T N1 is a single end-to-end neural net, from photons to actions:

- Vision-Language Model (System 2) that interprets the physical world through vision and language instructions, enabling robots to reason about their environment and instructions, and plan the right actions.
- Diffusion Transformer (System 1) that “renders” smooth and precise motor actions at 120 Hz, executing the latent plan made by System 2.

Code.
Weights on HF.
Open Physical AI dataset release.
Blog.

Also Nvidia introduced Newton, an open-source physics engine developed by NVIDIA and Google DeepMind, is designed to accelerate robot learning and development.

Built on NVIDIA Warp, which enables robots to learn how to handle complex tasks with greater precision, Newton is compatible with learning frameworks such as MuJoCo Playground or NVIDIA Isaac Lab—an open-source, unified framework for robot learning. 

Disney Research will be one of the first to use Newton to advance its robotic character platform.

GitHub.

Anthropic is gearing up for the web search launch on Claude Web with a composer UI redesign.

This change brings all tools besides the model selector into a new dropdown, which is also getting closer to what it looks like on mobile.

Not available yet.

Report by Dune and Artemis reveals key trends and metrics that continue to highlight the immense growth and evolving role of stablecoins

The total stablecoin supply has skyrocketed to $214 billion as of February 2025, with over $35 trillion in #transactions processed over the past year.

However, compared to traditional #fiat currencies, stablecoins still represent a fraction of the overall liquidity—around 100 times smaller than the US M1 money supply.

#Stablecoins have already surpassed major #payment networks like Visa and Mastercard in transfer volume, highlighting their growing importance in global finance.

Institutional adoption of stablecoins is growing, with asset managers, payment providers, and financial institutions integrating stablecoin solutions, driving cross-border payments and enhancing liquidity in #DeFi markets.

Stablecoin transaction volume saw an astonishing 115% year-on-year increase in 2024, from $1.9 trillion in February 2024 to $4.1 trillion by February 2025.

NVIDIA announced Isaac for Healthcare, the 3-computer solution for healthcare robotics, allowing simulation, training and deployment of AI solutions.

It extends and tailors Isaac Sim/Lab and Omniverse to enable multi-scale simulation from anatomy to sensors, surgical instruments, robotic systems, operating rooms, intensive care units, labs, sterile processing departments (SPD), and full hospital environments.

NVIDIA Isaac for Healthcare leverages Nvidia Holoscan for hardware-in-the-loop (HIL) simulation and Sim2Real deployment of trained algorithms and policies.

NVIDIA and Synchron Unveil Chiral™: Pioneering the New Era of Cognitive AI Through Brain-Computer Interfaces

Synchron and NVIDIA have taken a step forward in the evolution of AI with the announcement of Chiral™ - a foundation model of human cognition that marks the emergence of Cognitive AI, trained
directly on human neural activity.

At GTC 2025, the companies demonstrated how their collaborative technology enables individuals with severe physical limitations to control their environment using only their thoughts.

The showcase featured Rodney, an ALS patient who, despite being unable to use his hands, successfully managed his digital home environment entirely hands-free, voice-free, and touch-free through Synchron's Stentrode™ technology and NVIDIA's AI computing platform.

The Three-Stage Roadmap to Cognitive AI:

1. Motor Inference Powered by NVIDIA Holoscan
- Real-time neural decoding with minimal latency
- Enhanced on-device edge computation
- Foundation for more advanced, high-dimensional interactions
2. Context-Aware Interaction with NVIDIA Cosmos
- Integration of environmental awareness
- Physics-bound, photorealistic simulations of domestic environments
- Higher-dimensional inference making interactions more predictive and intuitive
3. From Neural Data to Cognitive AI: Training Chiral™
- Leveraging 20 patient-years of Stentrode implant experience
- Creating a pre-trained brain model that abstracts human cognition
- Evolution into a self-improving, general-purpose Cognitive AI

Chip design software maker Synopsys unveiled an AI agent called AgentEngineer to help human engineers design semiconductors, acting as an assistant in the process, taking care of specific tasks in chip design, such as testing circuit designs.

Cool paper by Google DeepMind. Most AI systems assume there’s just one right answer—but many tasks have reasonable disagreement. How can better model human variation?

Researchers propose modeling at the individual-level using open-ended, textual value profiles

Claude can now search the web

Each response includes inline citations, so you can also verify the sources.

Web search is rolling out today in feature preview in the US across all paid plans.

Just toggle it on in the Feature preview section of settings.