DeepLearning.AI

Researchers introduced Walrus, a transformer model that can predict what fluids (liquids, gases, and plasmas) will do across many physical domains. The model achieves higher accuracy and more stable long-term predictions than prior models, thanks in part to a "jitter" technique that reduces error accumulation during repeated simulations. Read our summary of the paper in The Batch: https://hubs.la/Q04dfZ3X0

Excited to see everyone building agents next week at the AI Dev 26 x SF hosted by DeepLearning.AI team! I'll be sharing how to use observability agents to find & fix issues in your AI agents. AI agents fail in prod due to brittle workflows, lack of contextual learning, and inability to improve over time. I will show how to use trace-based testing with coding agents as part of agentic engineering to find and fix issues in AI agents. You'll learn to debug, evaluate and observe AI agents using open source monocle2ai made easy with an observability agent. Would love to learn what kinds of exciting agents people are building!

If you’re in SF for the DeepLearning.AI AI Developer Conference, this is the conversation I’d want to be in. A lot of what we’re hearing right now about AI is more about speed and less about whether that speed is actually translating into better outcomes. Barun Singh and Kennith J. from Andela are talking through where things are working, where they’re not, and how teams should actually be thinking about skills and systems. And be sure to drop by our booth # 120 to chat with the team and I!

Meet the HoundDog.ai team at AI Dev 26 next week! Make sure to visit booth 512. See you in a few days! 🌉

Meet the Box team at AI Dev 26! Box helps devs unlock their unstructured data and build production-grade AI-powered applications using APIs, MCP and Agent Skills. Learn more at developer.box.com, stop by their booth and don't miss Carter Rabasa's talk on Filesystems as the New Primitive for AI Agents on April 28. 🎫 Last tickets available: https://lnkd.in/e5-GXM5s

Images, audio, and video are everywhere in modern orgs but most data pipelines weren't built for any of them. We just launched a new short course with Snowflake on building multimodal data pipelines. You’ll build systems that: - Convert images and audio into structured text (OCR, ASR) - Generate timestamped descriptions from video with Vision Language Models - Retrieve across slides, audio, and video with a multimodal RAG pipeline Taught by Gilberto Hernandez. Enroll in "Building Multimodal Data Pipelines:" https://hubs.la/Q04d0G1T0

AlphaGenome is an open-weights model from Google that interprets non-coding DNA, helping identify how variations affect gene behavior and expressions in disease. By predicting gene properties and mutation impacts with high accuracy, it offers a powerful tool for biomedical research and treatment development. Learn more in The Batch: https://hubs.la/Q04c_FWP0

At AI Dev 26, LandingAI will showcase Agentic Document Extraction (ADE), an API-first platform that turns messy, multi-modal documents and dark data into structured, auditable intelligence. Meet their team at booth 107 on April 28 and 29! Try your document at va.landing.ai