A useful layer for helping AI agents understand your data. @RUC1937 developed open-source EvoOntology. It builds a searchable guide to the data – an ontology that connects business concepts to the actual data. Then it updates this guide based on the agent’s task history. Here is how it works: EvoOntology is "agent-first". A builder agent examines real tasks and data sources to create the initial guide (ontology). It records definitions, field mappings, relationships, and business rules, with evidence from the data to support them. This guide has three layers: - Schema: defines concepts and relationships the guide can represent. - Content: holds definitions, links to the underlying data, and rules. - Tools: let the agent search and use the guide as needed to find and retrieve the right data. The agent searches the guide through MCP tools instead of receiving prompts. As the agent completes tasks, EvoOntology reviews its attempts and proposes changes to the guide. Each change is tested against the previous version and is kept only if the score improves by the required amount. The average gains are impressive: - Multi-source research: +17.8 percentage points in accuracy. - Insight mining: +1.9 points overall. - Data retrieval: +7.4 points in query accuracy and +8.6 in the execution-efficiency score. And some more features – EvoOntology is open-source, with Claude Code and Codex plugins for building, updating, and visualizing that guide.

If you use MongoDB and often ask AI agents to write or edit code that uses it, this article from @MongoDB may be useful for you. It shows how MongoDB Agent Skills help agents write better code through guidance on schema design, indexing, and query patterns It explains how to: - Connect agents to your deployments with defined permissions and access controls. - Get started in Claude Code, Cursor, Gemini CLI, and VS Code (installation commands + links to setup) - Avoid common modeling and indexing mistakes. - Create your own skills around team conventions For all the key patterns: mongodb.com/company/blog/pro…

What is recursive about Recursive Self-Improvement (RSI)? Not every case of AI improving AI is truly recursive. If a model edits some code and humans test the result, the loop repeats - but the loop itself stays the same. The recursion appears as more parts of that loop become editable: 1. AI improves a model or algorithm. It modifies code, kernels or training methods. In AI4AI-Bench, agents get several hours to change real training algorithms, which are then trained again from scratch. 2. AI improves how it searches. It chooses strategies, experiments and where to spend more compute. 3. AI creates useful experience. It decides what it needs to learn and uses its experiments in the next round. 4. AI adapts the research environment. It changes the tools and workflows used to produce improvements. 5. AI improves the improvement process itself. It changes how it searches, experiments and evaluates progress, and uses that new process in the next round. These levels of autonomy were mentioned in "The Last AI Built by Humans" paper. And "Recursive Criticality" adds another point: even fast, highly automated AI development is not necessarily RSI if parts of the improvement loop remain fixed. The recursive part is not that the loop runs again. It is that the loop itself becomes something AI can change. We've made a lot of guides which will give you a real understanding of RSI

Must-read papers of the week ▪️ JEPA-Anything ▪️ Modality-Autoregressive World-Action Models ▪️ In-Context Robot Learning with VLM Agents ▪️ Dream-RSI: Recursive Self-Improvement through Evolving Worlds ▪️ ModularRSI: Modular and Generalizable Recursive Harness Self-Improvement ▪️ DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression ▪️ SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness ▪️ Confidence Comes from Experience: Experiential Confidence Estimation from Reasoning to Agents ▪️ AliceAI-Foundation-80B-A3B-Base (model release) ▪️ Video DeltaNet: A Video-Native Hybrid Attention for Livestream Video Generation ▪️ ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents ▪️ World Modeling in Transformers ▪️ When2Think: Learning Difficulty-Aware Length Control for Efficient Hybrid Reasoning Models Explore these to keep up with main AI trends. Here’s also the full list of stunning papers + links and our weekly AI news digest: turingpost.com/p/tracktwo-ai

8 Notable advancements and news in robotics that we can't miss ▪️ Figure’s Helix 2.5 brings household skills to 30 new homes ▪️ Skild S1 learns from one video prompt ▪️ Facet-0 helps robots detect jammed parts ▪️ Safe-Stop teaches humanoids to stop without falling ▪️ GEN-1.5 learns short tasks from seconds of demos ▪️ Agility’s Digit 5 lifts 22.7 kg and charges in 9 minutes ▪️ FANUC's AI Welding Agent turns drawings into robot welding programs ▪️ and OpenAI plans its own humanoid robot Read more about everything in one place: turingpost.com/p/the-latest-…

There’s so much from this week in AI that I want to show you – and it all connects to something I’m especially excited about: AI’s growing ability to understand and act in the world: • Odyssey-3 – Connects world models to robot controls, with demos showing recovery from dropped objects and missed grasps. • Omni World Model – It's a planned model to understand scenes, generate them, and act. • Helix 2.5 – Tackles household tasks in 30 unfamiliar homes. Human-behavior pretraining lifts success from 9% to 56%. • Agility’s Digit 5 – Brings more safety systems onboard to help robots work around people without a fence between. • Jev – Answers constrained questions with probabilities in 70-500 milliseconds. This opens up faster checks and decisions for agents. • DeepMind Institute – Opens discussion: If AGI is getting closer, what should we do with it, and who gets a say? I unpack all of this and even more in my new video, with plenty of details and insights that didn’t fit a text. Hope, you'll find something inspiring there ↓

San Francisco is precious by moments like this. Thank you, Sarah and swyx for organizing this dinner: amazing group of people, insightful conversations

If you need practical code samples and hands-on build sessions, here is a good option. Try to join Edge Case → discord.com/invite/jQXnaf7Cq It's @Akamai Cloud’s free Discord community, open to anyone building with AI, LLMs, serverless, cloud, or WebAssembly. You'll get: - live build sessions - office hours with engineers scaling production workloads - production-ready code and tutorials. New Akamai Cloud accounts also receive $300 in cloud credits upon creation. No application required. @akamaidevs

I explored @typesafeai's Jev in a way even your grandma can understand. And tested it with Codex (something anyone can do). So if you already know everything about Jev, forward this video to someone less techy to make sure you’re on the same page.