DeepLearning.AI

We’re hiring a Product Designer to improve how learners navigate and engage with DeepLearning.AI’s platform! In this role, you’ll study how learners navigate real-world workflows, prototype new interactions, and design interfaces that simplify complex tasks. You’ll work across research and UI design in close partnership with product and engineering teams. This is a hybrid role in the Bay Area. Apply now, or share this opportunity with a friend 👉 https://lnkd.in/eBEXxVGD

At AI Dev 25 x NYC, Hatice Ozen, Head of Developer Relations at Groq, showcased how to build a deep research agent with one API call. AI agents today require complex orchestration: managing state, routing tools, handling retries, and coordinating multiple LLM calls, all while fighting latency at every step. In this hands-on workshop, Hatice guided the audience to built a sophisticated deep research agent using Groq's compound system: web search, code execution, and multi-step reasoning in a single API call with zero orchestration code. Attendees saw how instant inference enables intelligent server-side orchestration and walk away knowing when to use direct APIs vs. frameworks. Watch the full session: https://lnkd.in/efhVsrEJ

Baidu released Ernie-4.5-VL-28B-A3B-Thinking, an open-weights MoE vision-language model that tops many visual reasoning tasks at low cost. It also debuted the giant Ernie-5.0, a proprietary 2.4T-parameter natively multimodal model (text/image/audio/video). Learn more in The Batch: https://hubs.la/Q03XCXc80

🚀 New short course with Qdrant: Multi-vector Image Retrieval. Taught by Kacper Łukawski, Senior Developer Advocate at Qdrant, the course shows how multi-vector techniques outperform single-vector methods by matching text tokens to image patches directly. You’ll implement ColBERT to understand multi-vector search, apply ColPali for patch-level image retrieval, reduce memory with quantization and pooling, and use MUVERA to enable fast HNSW search. The course concludes with a full multi-modal RAG pipeline built on ColPali and MUVERA. Learn more and enroll now: https://hubs.la/Q03XCNCz0

World Labs’ Marble generates persistent, downloadable 3D spaces (as Gaussian splats or meshes) from prompts, photos, panoramas, videos, or simple layouts. Its Chisel editor lets you expand and edit scenes with text or visuals. Unlike prior “on-the-fly” 3D generators, Marble’s worlds can be saved, combined, and modified. Learn more in The Batch: https://hubs.la/Q03XxZ9_0

Aayush Kapoor, Software Engineer on the AI SDK team at Vercel, shared during AI Dev 25 x NYC how to build production-ready agents with the Vercel AI SDK. He covered fundamentals like text generation, tool and function calling, structured outputs, and swapping models with a single line, then guided attendees to code a Deep Research style agent in Node.js that searches the web and writes a markdown report. Watch his full session: https://lnkd.in/ec_twnB8

Get serialized! Learn how to use Python libraries like Pickle with the help of ChatGPT to serialize even the most complex objects easily! From simple commands like pickle.dump and pickle.load to handling nested layers of data, this quick tour shows you how to round-trip objects quickly. Gain these and more essential skills in our Generative AI for Software Development skills certificate. Let's start coding! https://hubs.la/Q03XkzMw0

Meta released an open-weights image segmentation suite: ➡️ SAM 3 segments images and video (including from text prompts) ➡️ SAM 3D Objects converts segmented items into 3D meshes/gaussians (optionally using point clouds) ➡️ SAM 3D Body generates full 3D human figures. In Meta’s tests, the models outperform most rivals on segmentation and 3D quality, and all are available to try online with downloadable weights under the Meta license. Learn more in The Batch: https://hubs.la/Q03XjkPC0

Researchers at Anthropic and multiple research and safety institutions identified persona vectors, patterns in LLM layer outputs that represent character traits like sycophancy or tendency to hallucinate. By averaging outputs from examples of a trait and subtracting its opposite, they can isolate and control these traits. Finding persona vectors lets engineers screen fine-tuning datasets beforehand to predict personality shifts, making training safer and more predictable. It reveals that high-level LLM behaviors are structured and editable, enabling more proactive control over model personalities. Read our summary of the paper in The Batch: https://hubs.la/Q03Xh1HX0

This week in The Batch, Andrew Ng identified declining public trust in AI as a major problem and called on the AI community to address legitimate concerns while building more applications that genuinely benefit everyone. Plus: 🌐 Meta's SAM 3 now turns images into 3D scenes and people 🎨 Marble makes editable 3D worlds from text, images, and video 👁️ Baidu debuts an open vision-language model and giant multimodal Ernie 5.0 💃 RoboBallet choreographs many robot arms at once Read The Batch: https://hubs.la/Q03WWNTt0