Subscribe, if this was forwarded to you

Hey there,

In this week’s edition we spotlight Yann LeCun’s AMI Labs and its $1.03B bet on “world models” that learn from real-world data, plus hands-on innovation with Unsloth Studio, the open-source UI making local LLM training accessible to all. Dive into Princeton’s OpenClaw-RL for a glimpse at continuous real-world reinforcement learning, and don’t miss the latest from Anthropic Research, Google’s Gemini Embedding 2, and Mistral Small 4—each pushing the boundaries of what AI can do. Whether you’re building, researching, or just curious, there’s something here for every tech enthusiast.

Let’s dive in!

Cheers, Andreas & Sascha

In today's edition:

📚 Tutorial of the Week: Continuously Learning Agents with OpenClaw-RL from Princeton

🛠️ Tool Spotlight: Unsloth Studio: Local LLM Training Platform

📰 Top News: Yann LeCun’s AMI Labs Raises $1B+

💡 Tip: Streamline API Docs with Context Hub

Let's get started!

Tutorial of the week

OpenClaw-RL: Continuous RL from Real Interactions

While not a traditional hands-on tutorial, this week we highlight a fascinating and educational resource from Princeton: OpenClaw-RL. This framework demonstrates an innovative path for improving AI agents by enabling them to learn continuously from real-world interactions, rather than relying solely on static, pre-collected datasets. The Princeton team’s approach is particularly instructive for those interested in the future of agentic AI systems.

• Unified, asynchronous RL infrastructure: OpenClaw-RL supports both personal agents (such as chatbots on your device) and large-scale general agents (including terminal, GUI, software engineering, and tool-call agents) within a single, decoupled system. This design allows for seamless, non-blocking training and deployment.

• Learning from next-state signals: The framework extracts both evaluative (how well did the agent do?) and directive (how should it improve?) signals from each interaction, using binary rewards and token-level supervision through Hindsight-Guided On-Policy Distillation (OPD).

• Scalable and practical: OpenClaw-RL is built for real-world use, supporting everything from single-user personalization to cloud-scale agent training, with robust logging and observability.

• Empirical validation: The paper demonstrates that combining binary RL and OPD yields significant gains, rapidly personalizing agents and improving performance across diverse agentic tasks.

Although this is not a step-by-step guide, studying OpenClaw-RL provides valuable insights into how continuous, real-world learning can unlock new capabilities for AI agents. We recommend this resource for AI researchers, RL engineers, and anyone curious about the next generation of agentic AI.

OpenClaw-RL: Train Any Agent Simply by Talking

Tool of the week

Unsloth Studio — Open-source UI for local LLM training and inference

Unsloth Studio is a new open-source, no-code web UI designed to make training and running large language models (LLMs) locally accessible to everyone. It addresses the need for privacy, customization, and hands-on experimentation by letting users fine-tune, run, and export open models directly on their own hardware—no cloud required.

• Unified local interface: Train, run, and export text, vision, TTS/audio, and embedding models using a single, intuitive UI. Supports over 500 models, including GGUF and safetensor formats.

• No-code workflow: Upload PDFs, CSVs, JSON, or DOCX files to auto-create datasets and kick off training instantly—no scripting necessary. Data Recipes streamline dataset creation and transformation.

• Advanced features: Includes self-healing tool calling, web search, code execution (Python/Bash), real-time observability (GPU utilization, training loss), and a Model Arena to compare outputs side-by-side.

• Privacy-first: All operations run 100% offline, with robust token-based authentication and no telemetry collection, ensuring your data stays secure and local.

• Cross-platform and hardware support: Works on Windows, Linux, WSL, and MacOS (chat and data recipes), with multi-GPU and Intel GPU support. Docker and Colab options are also available.

Unsloth Studio is rapidly gaining traction in the open-source AI community, with backing from NVIDIA and Hugging Face, and is under active development with frequent updates and new features.

Top News of the week

Yann LeCun’s AMI Labs Raises $1.03B to Pioneer Real-World AI Models

Turing Award winner Yann LeCun’s new venture, AMI Labs, has secured a staggering $1.03 billion in funding to develop “world models”—AI systems that learn from real-world data, not just language. This marks one of the largest investments in foundational AI research, signaling a shift from traditional large language models (LLMs) toward AI capable of understanding, reasoning, and planning in complex environments. AMI Labs aims to build intelligent systems with persistent memory and controllable behavior, targeting applications in robotics, healthcare, industry, and beyond.

AMI’s approach centers on JEPA (Joint Embedding Predictive Architecture), an architecture designed to learn abstract representations from sensor data and predict outcomes in representation space. Unlike generative models that struggle with unpredictable real-world data, AMI’s world models are action-conditioned, enabling agentic systems to plan and execute tasks safely. The lab’s global team includes top AI researchers and industry leaders, and its first partnership is with digital health startup Nabla, reflecting a focus on reliability and safety where LLMs fall short.

The funding round, co-led by Cathay Innovation, Greycroft, Hiro Capital, and others—including Nvidia, Samsung, and Toyota Ventures—will support AMI’s ambitious research agenda and talent acquisition across Paris, New York, Montreal, and Singapore. AMI Labs is committed to open research, promising to publish papers and release code to foster a collaborative ecosystem. The move is widely seen as a potential turning point for the next generation of AI, with “world models” poised to become the industry’s new buzzword.

🔗 Read the official TechCrunch announcement

Anthropic Study Finds Limited AI Impact on Labor Market

Anthropic’s latest research introduces a new measure for AI displacement risk, combining theoretical LLM capability with real-world usage data. The study finds no systematic increase in unemployment among workers in AI-exposed occupations since late 2022, though hiring of younger workers has slowed in these fields. This nuanced analysis suggests AI’s labor market effects remain modest, providing a foundation for ongoing monitoring as adoption grows.

🔗 Read the full Anthropic study

Google Releases Gemini Embedding 2: Multimodal Search Across Media

Google has launched Gemini Embedding 2, its first natively multimodal embedding model, available via Gemini API and Vertex AI. The model maps text, images, video, audio, and PDFs into a unified semantic space, enabling advanced search and retrieval across diverse media types. With flexible output dimensions and state-of-the-art performance, Gemini Embedding 2 empowers developers to build richer cross-media AI applications.

🔗 Explore Gemini Embedding 2

DeepMind Proposes AGI Cognitive Taxonomy, Launches Kaggle Hackathon

Google DeepMind has published a cognitive taxonomy framework to benchmark AI progress toward AGI, identifying ten key cognitive abilities. To operationalize this, DeepMind and Kaggle are hosting a hackathon with a $200,000 prize pool, inviting researchers to design evaluations for learning, metacognition, attention, executive functions, and social cognition. The initiative aims to standardize AGI evaluation and foster community-driven benchmarks.

🔗 Learn about the AGI framework and hackathon

Hugging Face Publishes Synthetic Data Playbook for Trillion-Token Datasets

Hugging Face has released the Synthetic Data Playbook, offering guidance on prompt engineering, model selection, and infrastructure for generating and managing trillion-token datasets. The playbook is designed for practitioners building large-scale synthetic data pipelines, providing best practices and insights for leveraging synthetic data in modern machine learning workflows.

🔗 View the Synthetic Data Playbook

Mistral Small 4: Unified Open-Source Model for Chat, Agents, and Vision

Mistral AI has unveiled Small 4, a versatile open-source model that unifies chat, agentic tasks, and vision capabilities. Featuring configurable reasoning depth, native multimodality, and efficient scaling, Small 4 streamlines AI development by replacing multiple specialized models with a single adaptable system. Released under Apache 2.0, it’s optimized for both enterprise and developer use.

🔗 Discover Mistral Small 4

Tip of the week

Streamline Coding Agents with Context Hub for Reliable API Docs

Keeping coding agents up-to-date with accurate API documentation is a constant challenge—outdated endpoints and missing parameters can break automated workflows. Context Hub, a CLI tool from Andrew Ng, solves this by providing curated, versioned, and language-specific API docs that agents can fetch on demand.

• Easy Integration: Install Context Hub globally with npm install -g @aisuite/chub and use commands like chub search openai or chub get openai/chat --lang py to fetch the latest docs in your preferred language.

• Reduce Hallucinations: Agents access only current, vetted documentation, minimizing the risk of calling deprecated endpoints or using incorrect parameters.

• Persistent Learning: Use chub annotate <id> <note> to attach local notes to docs—these annotations persist across sessions, so agents remember key insights and workarounds.

• Continuous Improvement: Provide feedback with chub feedback <id> up or down to help improve docs for everyone, creating a self-improving documentation ecosystem.

Try Context Hub to make your coding agents smarter and your automated workflows more reliable.

We hope you liked our newsletter and you stay tuned for the next edition. If you need help with your AI tasks and implementations - let us know. We are happy to help