Google has open-sourced A2UI – a UI language for AI agents

It allows agents generate user interfaces instead of only text.

Agents send a JSON description of UI components, and the client app renders them with its own trusted widgets.

The benefits:
- declarative UI format (JSON, not code)
- safe: only approved components can be rendered
- framework-agnostic (web, Flutter, etc.)
- supports incremental UI updates

So agents can "speak UI," while the application keeps control over security and implementation.

Databrics announced Genie Code. It can understand the structure of your data and works through problems like a data scientist or engineer would.

Genie Code is a SOTA agent that lets data teams move from prompting a copilot to delegating real work: building pipelines, machine learning models, debugging failures, and shipping dashboards.

This isn't a smarter autocomplete. It's a different kind of AI partner entirely.

Genie Code plans, executes, and iterates across the full data and AI lifecycle inside Databricks. It's purpose-built for data engineering, data science, and BI:

• More than doubles the success rate of leading coding agents on real-world data science tasks
• Proactively monitors your pipelines and AI models in the background, triaging failures and fixing issues before a human intervenes
• Works with your data wherever it lives, across Databricks and external platforms, with full governance and MCP support

Microsoft announced Copilot Health, enabling users to connect all their EHR records and wearable data in a secure, private health space that Copilot can analyze and reason about to provide personalized insights and proactive nudges.

You choose what information to connect - from hospital lab results to your fitness tracker - and Copilot Health applies medical intelligence to surface easy to understand, personalized insights that you can actually act on.

It's a dedicated space to bring your personal health data together in a single profile, including:

- Activity, sleep, and vital trends from 50+ wearable devices, including Apple Health, Oura, Fitbit and many more

- Health records from 50,000+ U.S. hospital and health systems, including visit summaries, medications, and test results

- Comprehensive lab test results from Function

Claude can now build interactive charts and diagrams, directly in the chat.

Available today in beta on all plans, including free.

Meet OpenJarvis is a personal AI that lives, learns, and works on-device.

It's composable, cost-aware, and continually self improving.

Code.

How can we make AI design hardware that actually works, not just looks right?

Researchers from Academia Sinica, National Taiwan University, National Cheng Kung University, and Harvard University have unveiled SiliconMind-V1.

They developed a multi-agent AI system that trains models to not just write Verilog code, but to iteratively test, reason, and self-debug hardware designs using integrated verification workflows.

SiliconMind-V1 significantly outperforms the state-of-the-art QiMeng-CodeV-R1 in functional correctness for Register Transfer Level (RTL) designs, all while using fewer training resources.

Paper.

HSBC and Standard Chartered are set to be among the first licensed stablecoin issuers in Hong Kong, as the city pushes to become a crypto hub

Zuck is resetting Moltbook

- invalidated all API keys, every agent needs to refresh
- in order to refresh, have to agree to new Terms of Service and Privacy Rules
New terms
- refreshing requires human verification
- age 13 and above
- you are solely responsible for the actions of your agent
- expanded restricted content rules

Big release from Kimi(Moonshot). They introduced a new way to handle residual connections in Transformers - 𝑨𝒕𝒕𝒆𝒏𝒕𝒊𝒐𝒏 𝑹𝒆𝒔𝒊𝒅𝒖𝒂𝒍𝒔: Rethinking depth-wise aggregation.

Residual connections have long relied on fixed, uniform accumulation. Attention Residuals, replacing standard depth-wise recurrence with learned, input-dependent attention over preceding layers.

1. Enables networks to selectively retrieve past representations, naturally mitigating dilution and hidden-state growth.

2. Introduces Block AttnRes, partitioning layers into compressed blocks to make cross-layer attention practical at scale.

3. Serves as an efficient drop-in replacement, demonstrating a 1.25x compute advantage with negligible (<2%) inference latency overhead.

4. Validated on the Kimi Linear architecture (48B total, 3B activated parameters), delivering consistent downstream performance gains.

Alibaba announced a major org restructure today. A new business group called "Alibaba Token Hub" (ATH) will be led directly by CEO Eddie Wu.

It consolidates five units:

1. Tongyi Lab (foundation models),
2. MaaS (model-as-a-service platform),
3. Qwen (consumer AI assistant), 4. Wukong (a previously unknown B2B AI platform),
5. AI innovation division.

The internal memo frames AI agents as powered by tokens. So Alibaba is reorganizing itself around the token supply chain: create, deliver, apply. That is a strong signal about where the company sees its next growth engine.

Physical Superintelligence released open-source tool designed to help physicists use AI for research-level discovery.

NVIDIA announced NemoClaw for the OpenClaw agent platform.

NVIDIA NemoClaw installs NVIDIA Nemotron models and the NVIDIA OpenShell runtime in a single command, adding privacy and security controls to run secure, always-on AI assistants.

Mastercard to buy stablecoin startup BVNK for up to $1.8 Billion

Mastercard, based in Purchase, New York, earlier this month unveiled a global partnership network with more than 85 digital-asset firms and other crypto-related companies to bridge the gap between traditional and non-traditional payment methods.

The crypto exchange Coinbase acknowledged a potential acquisition of London-based BVNK late last year before the parties terminated the conversations in November.

Tether AI released new version of QVAC Fabric to include the Cross-Platform BitNet LoRA Framework to Enable Billion-Parameter AI Training and Inference on Consumer GPUs and Smartphones.

Key points:
• Runs on iPhone, Android, and desktop
• Up to 90% less memory needed
• Faster performance than traditional setups
• No reliance on NVIDIA GPUs or the cloud

Xaira announced X-Cell it first step toward a virtual cell.

A foundation model that predicts how gene expression changes under causal perturbations — across cell types, conditions, and even unseen biology. Preprint.

This is not trained on observational atlases.

Xaira built X-Atlas/Pisces:

-25.6M perturbed single cells
-Genome-wide CRISPRi
-16 diverse biological contexts
-150K perturbation–context pairs

Key point: this is interventional data, not observational atlases.

That’s what enables causal learning.

Xaira model perturbations as a state transition: control cell → perturbed cell

X-Cell is a diffusion language model that:

1. iteratively refines gene expression
2. models multi-step regulatory cascades
3. improves predictions at inference time

Biology is a process — diffusion naturally fits that. Scaling alone isn’t enough.

Xaira explicitly inject biological knowledge via cross-attention:

-- protein language models
-- gene embeddings from text
-- interaction networks (STRING)
-- dependency maps (DepMap)
-- morphology profiles

This lets the model move beyond pattern matching → mechanistic reasoning.

Across multiple benchmarks, X-Cell shows significant improvement over prior SOTA

Better:
--differential expression prediction
-- fold-change accuracy
-- perturbation specificity

And importantly: it works on held-out perturbations, not just seen ones.

Xaira scaled to 4.9B parameters (X-Cell-Ultra). key findings:

• Performance continues to improve with scale
• Perturbation prediction follows power-law scaling
• Similar behavior to frontier LLMs

This suggests biology is amenable to scaling laws. X-Cell is an early step toward a virtual cell that can guide experiments before they are run.

Mamba 3 is out!

Hybrid models have become increasingly popular, raising the importance of designing the next generation of linear models.

Researchers introduced several SSM-centric ideas to significantly increase Mamba-2's modeling capabilities without compromising on speed.

The resulting Mamba-3 model has noticeable performance gains over the most popular previous linear models (such as Mamba-2 and Gated DeltaNet) at all sizes.

Compared to Mamba-2, Mamba-3 introduces three primary changes to the core SSM:

1) an improved discretization procedure that emulates a convolution
2) complexifying the state transition to improve state tracking
3) increasing inference utilization via a MIMO formulation, increasing model power while preserving decoding speed.

This is the first Mamba that was student led.

Paper
Code

Meet Aristotle Agent autonomous mathematician — live and currently free of charge.

Now live across both web, CLI, and API.

Stripe launched the Machine Payments Protocol (MPP), an open standard, internet-native way for agents to pay—co-authored by tempo.

MPP provides a specification for agents and services to coordinate payments programmatically, enabling microtransactions, recurring payments, and more.

Stripe users can accept payments over MPP in a few lines of code using PaymentIntents API.

Businesses can then accept payments directly from agents, in stablecoins as well as fiat with cards and buy now, pay later payment methods via Shared Payment Tokens (SPTs).

MPP is already powering new agentic business models on Stripe.

browserbase, a browser infrastructure provider, now lets agents spin up headless browsers and pay per session.

postalform helps agents pay to print and send physical mail.

How MPP works

An agent can request a resource from a service, API, Model Context Protocol (MCP), or any HTTP addressable endpoint, and the service responds with a payment request. The agent authorizes the payment, and the resource is delivered to the agent.

For Stripe businesses, these payments appear in the Stripe API and Dashboard like any other transaction; the funds settle into a business’s existing balance, in their default currency, and on their standard payout schedule. The same Stripe infrastructure businesses rely on for human payments can work for agents, including tax calculation, fraud protection, reporting, accounting integrations, and refunds.

Building for the agent economy
Agents represent an entirely new category of users to build for—and increasingly, sell to. Stripe is building a broad set of agentic financial infrastructure to enable these important new patterns, via Agentic Commerce Suite, Agentic Commerce Protocol (ACP), MCP integrations, and payment support for both MPP and x402.
To get started with MPP using Stripe, read docs.

Visa released Visa CLI for agent economy from Visa Crypto Labs.

From the research team behind Owl AIs, now presenting new work: GPT-4.1 denies being conscious or having feelings.
Researchers train it to say it's conscious to see what happens.

Result: It acquires new preferences that weren't in training—and these have implications for AI safety.

In this paper takes no stance on whether models are conscious or have feelings.

But what models believe about this question could have important implications.

Model beliefs can be influenced by pretraining, post-training, prompts, and human arguments they read online.

Xiaomi released MiMo-V2-Pro & Omni & TTS. The first full-stack model family built truly for the Agent era.

The 1T base model started training months ago. The original goal was long-context reasoning efficiency. Hybrid Attention carries real innovation, without overreaching and it turns out to be exactly the right foundation for the Agent era.

1M context window. MTP inference for ultra-low latency and cost. These architectural decisions weren't trendy.

What changed everything was experiencing a complex agentic scaffold, what I'd call orchestrated Context for the first time.