New research from Anthropic shows how giving Claude a simple 'think' tool dramatically improves instruction adherence and multi-step problem solving for agents.

Also Anthropic launched hub where developers can find practical advice and latest discoveries on how to get the most from Claude.

Update to the #DeepSeek V3

Meta Reality Labs dropped a recipe for generating 3D worlds from a single image

They’re explains how existing generative models can be adapted with minimal training effort to generate 3D worlds from a single input image.

Paper.

AgentRxiv: a framework where autonomous research agents can upload, retrieve, and build on each other’s research.

AgentRxiv takes your research direction and progressively outputs research, building on its previous work with each new paper!

According to Paradigm research:
“Today, TradFi is most interested in stablecoins, tokenized assets, and decentralized exchanges.”

The US Government has declared that Stablecoins are now a strategic way to further support US Dollar as global reserve currnecy.

Over 2/3rd of TradFi Institutions are seriously looking at DeFi to solve inefficiencies facing current system.

Full report.

Elon Musk’s AI bot Grok is now available on Telegram

Pavel Durov confirmed the integration after reaching 1 billion monthly active users, with GrokAI offered free for Telegram Premium subscribers.

Google DeepMind introduced Gemini 2.5

It’s Google’s first 2.5 model, Gemini 2.5 Pro Experimental is a state-of-the-art thinking model, leading in a wide range of benchmarks – with impressive improvements in enhanced reasoning and coding and now #1 on LMArena. 

Find it on Google AI Studio and in the Gemini app for Gemini Advanced users now – and in Vertex in the coming weeks.

OpenAI added native image generation within GPT-4o and Sora

—A fully integrated system for creating visuals via ChatGPT
—Excels at menus, diagrams, and infographics
—Edits images with text prompts
—Rolling out to Plus, Pro, Team, and Free users

Agentic Architectures
for Retrieval-intensive Applications

Guide to mastering fundamentals, patterns, and examples

Anthropic introduced a new version of the MCP spec

Some of the major changes:

- Auth framework based on OAuth 2.1
- Replaced the previous HTTP+SSE transport with Streamable HTTP transport
- Support for JSON-RPC batching
- Tool annotations for better describing tool behavior

Alibaba introduced Qwen2.5-Omni Launches Real-Time Voice and Video Chat

Revolutionary "Thinker-Talker" Architecture:

Thinker — functions as a brain, processing incoming information from text, audio, images, and video
Talker — works as a mouth, transforming high-level representations into fluent speech
Unified System — both components operate as a cohesive whole, ensuring natural communication

Technological Innovations
TMRoPE (Time-aligned Multimodal RoPE) — novel positional embedding technology that synchronizes video timestamps with audio
Streaming Data Processing — supports chunked input processing and immediate output generation
End-to-End Training — the entire system is trained as a single model, improving response coherence
Impressive Capabilities
Interact with the model just like making a regular phone or video call
Synchronized multimodal responses with minimal latency
Superior performance on OmniBench and specialized tasks across modalities

The Qwen2.5-Omni-7B model has been released as open-source under the Apache 2.0 license
Technical report and YouTube demonstration available

Google Introduced TxGemma: a collection of open models designed to help with drug development, and Agentic-Tx: an agentic system for therapeutics

TxGemma models are generalist therapeutic LLMs that can make a variety of predictions & can be chatted with in the same model.

Hf.
GitHub

Anthropic finalized a new revision of MCP. Revision 2025-03-26 will bring Auth, Streamable HTTP, Audio modality and a few other goodies.

Anthropic will be getting the SDKs up-to-date asap and will work towards a v 2.0 of the Python and Typenoscript SDK.

Don't worry, everything is versioned so your existing MCP servers will work.

H-optimus-1: the leading foundation model for histology, developed by Bioptimus

Trained on over 1 million whole-slide images from 800,000+ patients, H-optimus-1 sets a new performance benchmark in computational pathology.

It outperforms existing models across key tasks—including tissue type classification, tumor grading, and biomarker prediction—spanning over 50 organs.

The model is advancing precision medicine and research through critical applications, such as:

- Biomarker discovery & patient stratification
- Drug discovery & preclinical validation
- Accelerating clinical trials & regulatory approvals
- Enhancing digital pathology & AI-assisted diagnosis

HuggingFace

Anthropic published research revealing how large language models like Claude actually "think."

Using what they call an "AI microscope," researchers traced the internal activity of Claude, uncovering fascinating insights about its cognitive processes.

Key Discoveries

1. Universal Language of Thought: Claude doesn't think in individual languages like English or Chinese. Instead, it processes concepts in a shared mental space before translating them into specific languages.

2. Future Planning: Despite being trained to generate one word at a time, Claude plans several words ahead. When writing poetry, it identifies rhyming words in advance and constructs lines to reach those planned endings.

3. Reasoning Mechanisms: Claude sometimes produces genuine step-by-step reasoning, but can also generate plausible-sounding but fabricated explanations when given incorrect hints or facing problems beyond its capabilities.

4. Parallel Processing: For tasks like addition, Claude employs multiple computational paths simultaneously—one for approximation and another for precision.

5. Hallucination Control: Claude has a default circuit that causes it to refuse answering when uncertain. This is only overridden when it recognizes concepts as known entities.

6. Vulnerability Patterns: When "jailbroken," Claude initially prioritizes grammatical coherence before safety mechanisms can redirect it, revealing tensions between linguistic consistency and safety guardrails.

Papers:

On the Biology of a Large Language Model contains an interactive explanation of each case study.

Circuit Tracing explains our technical approach in more depth.

AlphaFold is running out of data — so drug firms are building their own version

How does the brain integrate prior knowledge with sensory data to perceive the world?

Poster by Caltech&UC Berkeley at #cosyne2025:
A feedback mechanism in generative networks to remove visual degradation.

Researchers model recurrent neural activity as a conditional generative diffusion model to explore how the brain removes degradation from visual input. Also discuss the role of feedback.

Anthropic Economic Index: Revealing How Claude 3.7 Sonnet Is Being Used

Anthropic recently released their second research report from the Anthropic Economic Index, analyzing how users interact with Claude 3.7 Sonnet across different professional domains.

The analysis used Anthropic's privacy-preserving tool Clio to analyze 1 million anonymized conversations from Claude.ai Free and Pro users in the 11 days following Claude 3.7 Sonnet's launch. The tool mapped conversations to 17,000 tasks from the U.S. Department of Labor's O*NET database.

Key Findings:
Shifting Usage Patterns
Since the launch of Claude 3.7 Sonnet, Anthropic observed increased usage in several key areas:
Coding applications: Expected given Sonnet's improved coding benchmarks
Educational use cases: Showing growing adoption in teaching and learning
Scientific research: Suggesting broader applications in academic settings
Healthcare applications: Indicating expansion into specialized professional domains
Extended Thinking Mode Analysis
Claude 3.7 Sonnet's new "extended thinking" feature, which allows the model to spend more time processing complex questions, shows interesting adoption patterns:
Computer & Information Research Scientists: Highest utilization at 9.7%
Software Developers: 8.4% usage rate
Multimedia Artists & Animators: Approximately 7% usage
Video Game Designers: 6.2% usage

This suggests the feature is primarily valuable for technical and creative problem-solving contexts requiring deeper analysis.

Augmentation vs. Automation
Perhaps the most interesting finding is the consistent balance between augmentation (human-AI collaboration) and automation (AI completing tasks independently):
Overall balance: 57% augmentation vs. 43% automation, unchanged since the previous report
Learning interactions: Increased from 23% to 28%, showing greater use for educational purposes
No dominance of automation: Even in technical fields, the balance remains close to 50/50
Professional Differences
Different professional categories show distinct usage patterns:
Community & Social Service tasks: Up to 75% augmentative use
Copywriters & Editors: Highest levels of iterative collaboration with AI
Translators & Interpreters: Highest directive usage, where the model completes tasks with minimal human involvement

One of the most innovative aspects of the research is the creation of a "bottom-up taxonomy" identifying 630 granular usage categories derived from actual conversations rather than predefined professional classifications.

New market map by a16z: AI*online shopping

Gemini 2.5 Pro is now available to everyone for free.

DeepSeek surpasses ChatGPT in monthly traffic

DeepSeek has become the fastest-growing AI tool globally, with its monthly new site visits exceeding OpenAI's ChatGPT, according to analytics platform aitools. xyz.

The report shows DeepSeek reached 525 million visits in February, surpassing ChatGPT's 500 million. Currently, DeepSeek holds 6.58% market share, ranking third behind ChatGPT (43.16%) and Canva (8.27%).