Hey, Devs! :) This week, I picked a few updates that matter for developers building with AI: safer agents, better multimodal retrieval, more private on-device assistants, and one new research/policy move from Anthropic. Enjoy reading! ;)
Introducing The Anthropic Institute
Anthropic launched The Anthropic Institute, a new effort focused on the big questions around powerful AI: how it may affect jobs, the economy, public values, and governance. The institute brings together Anthropic’s Frontier Red Team, Societal Impacts, and Economic Research groups, and it is led by co-founder Jack Clark in a new role as Head of Public Benefit.
Read more: https://www.anthropic.com/news/the-anthropic-institute
OpenJarvis: personal AI on personal devices
Stanford’s Scaling Intelligence Lab introduced OpenJarvis, an open-source framework for personal AI agents that runs entirely on-device. It focuses on shared building primitives, efficiency-aware evaluation, and a learning loop that improves models using local trace data.
The framework defines five main primitives for on-device AI: Intelligence, Engine, Agents, Tools & Memory, and Learning. If you care about privacy, cost control, and local-first AI assistants, this is a project worth watching.
Read more: https://scalingintelligence.stanford.edu/blogs/openjarvis/
Gemini Embedding 2
Google released Gemini Embedding 2, its first natively multimodal embedding model. It can map text, images, video, audio, and documents into a single shared embedding space, making multimodal retrieval and classification much easier. It is available in public preview through the Gemini API and Vertex AI. The model supports over 100 languages, can process PDFs, audio, video, and images, and offers flexible output dimensions from 128 to 3072.
It means developers can build cross-modal search and retrieval systems with a single embedding model rather than stitching together separate pipelines.
Read more: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-embedding-2/
Designing AI agents to resist prompt injection
OpenAI published a new piece on how to design agents that resist prompt injection. Their main point is important: defending agents is not only about detecting bad strings in inputs. Many real attacks now look more like social engineering, so systems should be designed to limit damage even when manipulation gets through.
If you are building agents that browse the web, read email, or take actions on behalf of users, prompt injection is now a product design problem, not just a model problem.
I might use this in my own research about the Validation and Verification of an Agentic AI system.
Read more: https://openai.com/index/designing-agents-to-resist-prompt-injection/
How to Set Up OpenClaw with Ollama
I also published a new video: How to Set Up OpenClaw with Ollama: The Ultimate Guide.
In this video, I walk through how to connect OpenClaw with Ollama and get a local AI setup running step by step.
Watch here:
See you next week! Cheers, proflead! 😉