DeepLearning.AI

📢 New short course in collaboration with @Oracle! Agent Memory: Building Memory-Aware Agents Learn how to design a memory system that lets AI agents store, retrieve, and refine knowledge across sessions. Taught by @RichmondAlake and Nacho Martínez. Enroll now: hubs.la/Q047ljGB0

Your AI Agent Has a Memory Problem Most AI agents start from zero every single session. @DeepLearningAI's new course with @Oracle, "Agent Memory: Building Memory-Aware Agents," teaches you how to fix that by building a fully stateful agent from scratch. Check it out: hubs.la/Q049CYF10

Document processing isn't document understanding! OCR reads text. It extracts words from PDFs. But it misses layout, structure, and relationships. A table becomes a wall of text. Reading order scrambles. Spatial context disappears. Documents aren't just text containers. They're visual systems. The position of a field relative to its label matters. Table boundaries matter. Hierarchies matter. Multi-page flow matters. Traditional OCR is perception, not cognition. It sees characters but doesn't understand what they mean in context. Agentic Document Extraction (ADE) approaches this differently. Vision-first architecture that preserves structure. Parsing that maintains relationships between elements. Every extraction grounded to its exact location in the source. Our free course with @DeepLearningAI walks through why traditional OCR breaks on production documents and how vision-first extraction solves it. 6 hands-on labs covering everything from OCR failure modes to building RAG applications with visual grounding.

When processing long contexts, large language models often lose track of details or devolve into nonsense. Researchers reduced these effects by managing context externally. MIT’s Alex L. Zhang, Tim Kraska, and Omar Khattab developed Recursive Language Models (RLMs) that process long prompts encountered in books, web searches, and codebases by offloading prompts to an external environment and managing them programmatically. Learn more in The Batch: hubs.la/Q049x95V0

We can’t wait to welcome you to AI Dev 26 in San Francisco this month! The agenda and speaker lineup for the AI Developer Conference are now live. You'll get to hear from and interact with industry leaders at @GoogleDeepMind, @Oracle, @AMD, @ActianCorp, @Neo4j, @Arm, and many more, including: - @AndrewYNg (@DeepLearningAI) on the future of software development - @hwchase17 (@LangChain) on the observability flywheel - @DynamicWebPaige (@GoogleDeepMind) on what's new and what's next in AI - @RichmondAlake (@Oracle) on agent memory - @EmmaMcGrattan (@ActianCorp) on vector databases - Marina Mogilko (@siliconvalleymm) will moderate a panel discussion featuring leaders from LangChain, Oracle, @LandingAI, @Replit, and more. Join us April 28 and April 29 at Pier 48 in San Francisco. 🎟️ Grab your tickets: hubs.la/Q049mnPH0

OpenAI partnered with Amazon to build infrastructure for agents on the world’s largest cloud platform, a further sign that its close relationship with Microsoft is weakening. Learn more in The Batch hubs.la/Q0498M5n0

Hi @BeeeeNature! Thank you so much for reaching out and providing this feedback! We sincerely apologize for the frustration, you make a fair point about the limitations of email engagement tracking. We'd be glad to re-subscribe you immediately, and we will use this as an opportunity to revisit our re-engagement practices. Please send us a DM with your email address and we will have it fixed right away.

Learn more about our new short course with @DeepLearningAI, aimed at teaching AI developers and engineers how to architect and implement memory systems that give agents persistence, continuity, and the ability to learn over time. 🧠 social.ora.cl/6017B65FER

Apple introduced AToken, a model that uses a shared tokenizer and encoder to process and generate images, videos, and 3D objects in one unified framework. The multimodal approach beats or rivals specialized models in performance and helps transfer knowledge transfer across media types, potentially reducing training data needs and improving applications. Read our summary of the paper in The Batch: hubs.la/Q048Hp-v0

CEOs are wasting millions on AI. Here's why. 💸 CEOs are pouring millions into AI, but simply swapping a human for an AI agent in the middle of a broken process won't change your business. True AI transformation requires an end-to-end workflow redesign. Don't just use AI to make the old way faster—use it to build a completely new experience. Ready to build AI workflows that actually move the needle? @DeepLearningAI offers free short courses to help you master agentic workflows and the latest AI tools. Start learning today: hubs.la/Q048nRwq0 Subscribe to The Batch newsletter for weekly AI insights: hubs.la/Q048nVvj0

DeepSeek denied Nvidia and AMD early access to its upcoming DeepSeek-V4 while sharing the model with Huawei. The move highlights growing geopolitical friction between the U.S. and China, as export controls only have limited ability influence global competition for advanced hardware. Learn more in The Batch ⬇️ hubs.la/Q048nvSS0

🛠️ Builders, show us what you made! If you've completed a project in the Build with Andrew course, we want to feature it. Your work inspires the whole community. Here's how to share: 📌 Go to the @DeepLearningAI Forum: hubs.la/Q0487K9J0 📂 Post in the AI Discussions Category 🏷️ Tag it #build-with-andrew and #ai-discussions Let's see what you've built!

Alibaba released the Qwen3.5 family of open-weights vision-language models, ranging from lightweight to massive systems. Smaller models like Qwen3.5-9B rival or outperform much larger competitors, making multimodal AI more accessible on lightweight hardware while maintaining strong performance across vision and language tasks. Learn more in The Batch: hubs.la/Q0487yjb0

AI teams often think their biggest problem is the model. In practice, the bigger problem is alignment. If different people are optimizing different things, like accuracy, latency, recall, or edge cases, every experiment becomes a debate. One team thinks the system improved. Another thinks it got worse. High-performing AI teams avoid this trap. They agree early on what success looks like. Once that’s clear, experiments stop being arguments and start becoming progress. Learn how to build better AI systems at @DeepLearningAI

This week, in The Batch, Andrew Ng discusses growing job insecurity and uncertainty in an AI-driven world by focusing on stable foundations like strong communities and continuously developing skills. Plus: 👁️ Qwen3.5 models deliver top-tier vision performance, even at small sizes 🌏 DeepSeek withholds access from U.S. chip makers, escalating tensions 🧩 Apple unifies images, video, and 3D objects with a single tokenizer Read The Batch: hubs.la/Q047-T0K0

The #1 Skill Employers Want in 2026 The job market is shifting, and the #1 skill employers are looking for right now isn't a specific programming language—it's teachableness. As AI tools evolve at lightning speed, your ability to adapt and learn is your biggest competitive advantage. AI won’t replace workers, but workers who use AI will replace workers who don’t. Ready to 10x your productivity? @DeepLearningAI offers free short courses on the exact tools employers are looking for right now, including Claude Code, Gemini CLI, and Agentic Skills. Start learning today and build your career advantage: hubs.la/Q047Nc2q0 Subscribe to The Batch newsletter for weekly AI insights: hubs.la/Q047NjtB0

Apple researchers introduced Feature Auto-Encoder (FAE), a diffusion image generator that learns from compressed embeddings produced by a pretrained vision model. By shrinking rich embeddings before reconstruction, the system trains up to seven times faster while maintaining image quality comparable to state-of-the-art diffusion models. Read our summary of the paper in The Batch: hubs.la/Q047M-gw0

A lot of AI conversations focus on chatbots and demos. But some of the most valuable AI systems solve much less glamorous problems. Processing invoices. Extracting information from documents. Connecting data across tools. Making systems reliable enough to use every day. These kinds of operational workflows are where many companies actually see the biggest impact from AI. Here’s a simple path to start learning how these systems are built. Document AI: From OCR to Agentic Doc Extraction hubs.la/Q047N1Ph0 Preprocessing Unstructured Data for LLM Applications hubs.la/Q047N1KF0 Functions, Tools and Agents with LangChain hubs.la/Q047N0h70 Improving Accuracy of LLM Applications hubs.la/Q047N5J60 Know someone exploring how to apply AI in real business workflows? Share this with them.

Companies including Meta and OpenAI are building private gas-powered plants directly connected to data centers to secure the massive electricity needed to power AI infrastructure. The projects bypass grid delays and could provide energy for a large share of future data centers, but they also raise concerns about rising costs and increased greenhouse gas emissions. Lear more in The Batch hubs.la/Q047M2tQ0

Meet @Oracle on the AI Dev x SF show floor or join their workshop to explore the latest thinking on agent memory and what it takes to build agents that learn, adapt and actually work in production. 🎟️ Grab your tickets: hubs.la/Q047M1Bt0

A lot of beginners start their AI projects the wrong way. They start with the model. Which model is best? Which architecture should I use? But the most important question comes earlier. Who actually has the problem you’re trying to solve? Good AI projects start with real problems and real users. That’s what makes the technology matter. Learn how to build AI that solves real problems at @DeepLearningAI