Ai2 introduced Tülu 3 a set of SOTA instruct models with fully open data, eval code, and training algorithms.

8B model
70B model
Try it out

A lot of gems and insights in this fresh unreleased slide-deck by HuggingFace.

Production uses for open and closed source AI

Anthropic has just announced a universal bridge between AI systems and various data sources the Model Context Protocol (MCP).

This open-source standard promises to revolutionize how AI assistants interact with enterprise data systems, development environments, and content repositories.

The Model Context Protocol represents a universal bridge between AI systems and various data sources. Instead of requiring custom implementations for each new data connection, MCP provides a standardized way for AI assistants to access and utilize organizational data securely.

Key Components

The initial release includes three major elements:

1. The core Model Context Protocol specification and SDKs
2. Local MCP server support integrated into Claude Desktop applications
3. An open-source repository featuring pre-built MCP servers

The protocol launches with pre-built MCP servers for popular enterprise systems including:
- Google Drive
- Slack
- GitHub
- Git
- Postgres
- Puppeteer

Major players are already embracing the technology:
- Block and Apollo have integrated MCP into their systems
- Development tools companies including Zed, Replit, Codeium, and Sourcegraph are actively working on MCP implementation
- Claude 3.5 Sonnet has demonstrated exceptional capabilities in building MCP server implementations.

❗️Neuralink announced the approval and launch of a new feasibility trial to extend BCI control using the N1 Implant to an investigational assistive robotic arm.

This builds on their ongoing PRIME Study, which helps people with quadriplegia control digital devices through thought alone.

Now, they're taking a crucial step forward - from digital to physical control. Participants who are already enrolled in PRIME will have the opportunity to join CONVOY and potentially control an assistive robotic arm using just their thoughts.

This feasibility trial marks a pivotal moment in brain-computer interface technology. While PRIME focuses on restoring "digital freedom" - letting users control computers and smartphones mentally - CONVOY aims to restore physical independence through robotic assistance.

Andrew Ng announced new open-source Python package: aisuite

This makes it easy for developers to use large language models from multiple providers.

Open-source code with instructions.

You can connect Claude to an internet search engine using MCP

Here's how you can do it too in under 5 minutes:

1. U will need to download the latest version of our Claude desktop app here. To use Brave Web Search specifically, you will need to sign up for a free API key here.

2. Open up your Claude Desktop configuration file:

- macOS: ~/Library/Application Support/Claude/claude_desktop_config.json - Windows: %APPDATA%\Claude\claude_desktop_config.json

3. Then, just add this to that file and save it:

{
"mcpServers": {
"brave-search": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-brave-search"],
"env": {
"BRAVE_API_KEY": "ADD_YOUR_API_KEY_HERE"
}
}
}
}

4. Restart your Claude desktop app for the changes to load.

You can check that the server has been configured properly if you navigate to "Claude" > "Settings" in the top bar and check the Developer tab in this window that pops up.

5. After that, just ask Claude to make a web search for you!

The server tools automatically get loaded in to the system prompt so that Claude knows it has access to them.

If you want to improve on this server, explore more servers, or make your own integrations, go check out GitHub repository.

Cursor agent with the 0.43.3 update is crazy.

You can just let it code while you go to the beach.

OpneAI’s long awaited Sora video model may have leaked on HuggingFace.

Google introduced Health AI Developer Foundations is a game-changer for healthcare AI innovation

It provides open-weight models and resources to help developers build healthcare AI tools faster and more efficiently.

The initial focus is on imaging applications in radiology, dermatology, and pathology, with models pre-trained on large, diverse datasets for powerful performance.

Whether it's chest X-rays, skin conditions, or pathology slides, these models offer a robust starting point for creating AI solutions with minimal additional data and compute.

This suite builds on community feedback and is available for download via platforms like Vertex AI Model Garden and Hugging Face.

This initiative could democratize AI development, enabling creators to innovate for real-world healthcare challenges—improving diagnostics, broadening access, and easing clinician workloads.

Web3 x AI Market Map 2024

Web3 and AI are converging to create a new "supercycle" of technological development. AI serves as the brain of the system, handling decision-making and automation, while Web3 forms the economic foundation, enabling decentralized value exchange.

Key Trends of 2024:

1. The Age of Smart Agents
- AI agents are now creating real economic value
- Agents trade with each other, manage assets, and even launch their own projects
- Platforms have emerged where anyone can create a personal AI agent in minutes.

2. Five Layers of the New Economy

🔵 Compute: Decentralized networks providing AI processing power
🔵 Data: Transparent systems for collecting and validating information
🔵 Model: Creating and training AI on blockchain
🔵 Interface: Tools for system interaction
🔵 Applications: End products for users

3. Practical Applications

- Finance: AI manages investments and analyzes risks
- Gaming: Smart characters with unique personalities
- Creativity: AI creates music and art in the metaverse
- Social Networks: New level of AI-powered interaction

AI2 released OLMo 2, fully open language model to date, including a family of 7B and 13B models trained up to 5T tokens.

OLMo 2 outperforms other fully open models and competes with open-weight models like Llama 3.1 8B.

Model.

Try out an OLMo 2 demo on Playground.

Meta introduced Sparsh is the first general-purpose encoder for vision-based tactile sensing that works across many tactile sensors + many tasks.

The family of models was pre-trained on a large dataset of 460K+ tactile images using SSL.

HuggingFace released SmolVLM, a smol 2B VLM built for on-device inference that outperforms all models at similar GPU RAM usage and tokens throughputs.

SmolVLM can be fine-tuned on a Google collab and be run on a laptop.

Or process millions of documents with a consumer GPU.

Demo.
Model.
Fine-tuning script.

👀Total Stablecoin Value Tops Terra-Era Peak to Near $200 Billion

The total market capitalization for stablecoins has swollen 46% this year to a record of about $190 billion on Sunday, according to DeFiLlama data.

Tether, which issues the world’s largest stablecoin USDT, has seen its token’s circulation soar to nearly $133 billion, accounting for 70% of the total stablecoin market. The company wants to broaden usage of USDT by pushing into new industries including commodities. It recently announced the funding of its first crude oil transaction in the Middle East

Meanwhile, Stripe Inc., the $70 billion fintech giant, in October said it plans to buy stablecoin startup Bridge for $1.1 billion in one of the largest-ever acquisitions of a digital-asset startup. Other traditional finance firms including PayPal Holdings Inc. have also ventured into the sector.

The recent stablecoin surge coincides with a crypto rally spurred by Donald Trump’s victory in the US presidential race. Once an industry skeptic, Trump won over the sector with pledges to put in place more favorable policies.

The overall value of the digital-assets market has risen by nearly a trillion dollars since the election, according to data from CoinGecko.

Alibaba released a preview of QwQ /kwju:/ — an open model designed to advance AI reasoning capabilities.

LLM-brained GUI Agents

Great survey on LLM-brained GUI agents, including techniques and applications.

Mistral announced Startup Program: Mistralship

Become part of a 6-month-cohort of 10 startups to receive:

- 30K credits for La Plateforme
- Dedicated 1:1 support from our Solutions & Science team
- Early access to new models & products

Apply here before January 5th.

First open source reasoning model from Alibaba

Qwen QwQ 32B is in preview, does reasoning like o1 and Deepseek R1 but y'know, on your local machine.

Demo
Model

Scientists at Yale University have developed BrainLM, the first foundation model for analyzing functional MRI brain recordings.

Here's what makes it revolutionary:

- Trained on 6,700 hours of brain activity recordings
- Uses self-supervised masked-prediction training
- Processes data from 77,298 fMRI samples
- Analyzes 424 brain regions simultaneously

Capabilities:

- Accurately predicts clinical variables like age, anxiety, and PTSD
- Forecasts future brain states
- Identifies functional networks without supervision
- Generates interpretable representations of brain activity patterns

What Sets It Apart:

- Generalizes well to new patients and external datasets
- Outperforms baseline models in clinical predictions
- Serves as a powerful "lens" for analyzing massive fMRI repositories
- Creates meaningful insights about brain organization

Potential Applications:
- Non-invasive assessment of cognitive health
- Early detection of psychiatric disorders
- Research tool for understanding brain dynamics
- Biomarker discovery for mental health conditions

Technical Implementation:

- Based on Transformer architecture
- Trained on UK Biobank and Human Connectome Project data
- Uses advanced preprocessing and brain parcellation techniques
- Employs state-of-the-art deep learning methods

PrimeIntellect released INTELLECT-1: open-sourcing the first decentralized trained 10B model

INTELLECT-1 base model & intermediate checkpoints
- Pre-training dataset
- Post-trained instruct models by Arcee ai
- PRIME training framework
- Technical paper with all details.

Decentralized training is one of the most important things to track from an AI policy POV - most efforts to control compute rest on the assumption that strategic AI compute needs to be centralized.

HuggingFace.

Pre-training Dataset.

Post-training datasets by Arcee ai: 1 , 2.

Try out Intellect-1 on.

Tesla unveiled an upgraded hand system for its Optimus humanoid robot, featuring 22 total degrees of freedom

The robot teleoperated by a human, can now catch a ball, with tactile sensing and fine tendon control improvements coming by the end of the year.