Andrew Ng (吳恩達)

Humans can only hold six or seven things in our short-term working memory. Do you know how much an AI can hold in a single conversation? >> Interestingly, AI can use a large amount of context. Leading AI models today can accept maybe up to around 750,000 words as context. And this corresponds to about the first four or five Harry Potter books. So that's lot of text.

This is where embeddings come in. An embedding is a vector, a list of numbers that captures semantic meaning. [music] You've seen vectors in 2D and 3D, but to capture the meaning of something like a meeting transcript, we scale up to hundreds or even thousands of dimensions. These are vector embeddings. The key idea is that embeddings place semantically similar things close together. Budget and fi

Learn more: https://bit.ly/43ctPTW Join our new short course, AI Agents for Image and Video Generation, built in partnership with Google and taught by Katie Nguyen, Developer Relations Engineer at Google Cloud AI, and Wafae Bakkali, Staff Generative AI Specialist at Google. Most agents you've worked with probably produce text. But whether you're building a product demo, a website asset, or an expl

Most agentic systems rely on hardcoded heuristics to navigate execution decisions (e.g. which models, tools, and test-time compute scaling approaches to use) leading to efficiency leakage across cost, latency and accuracy. AI21 Maestro optimizes agents by learning to predict success, cost and latency probabilities across diverse actions and contexts, and driving runtime orchestration that intellig

Uh Andrew here from Zenoder. So today we're going to talk about uh some very interesting results from our internal lab team that we first kind of tested on different benchmark and evals and then um on our own team being guinea pigs where we got about 50 engineers and then happily rolling out across our customers and seeing um seeing the results there. just a little bit of our uh background for tho

Hi, I'm Neil Kano. I'm going to be talking to you today about edgetocloud video anomaly detection. Just doing a quick intro because Terry Demiba's session at AIDv did have some audio technical difficulties for the first few minutes. I'm his manager Devril leader at Quadrant and I'm going to give you a short intro before handing over to Terry for the live presentation. So today we're talking about

At AI Dev 26 x San Francisco, Andrew Ng discussed the rapid evolution of software engineering driven by AI coding agents and introduced new tools to support this shift: - The Shift in Software Development - New Bottlenecks and Generalist Skills - Job Market Perspective Andrew also announced: Context Hub: A tool designed to provide AI agents with up-to-date documentation to prevent hallucinations a

Stop asking AI to write your entire essay at once. Do this [music] instead. Let's take a look at some techniques for getting AI to write effectively for you, and how to take advantage of AI reasoning and avoid [music] AI slop. Maybe you ask AI to write the final text right away just from the start. If it writes a sentence like this, you may be unhappy with a few words, and you can edit a few words

Every serious agent will eventually need to buy tools, data or resources. Today, that means hardcoded API keys, shared credit cards, or looping in a human. x402, Coinbase's open protocol built on HTTP 402, claims to be one of the strongest candidates to fix this. This talk covers what Ondra Urban's team at Apify learned shipping x402 to Apify's 25,000 Actors: why the "exact" payment scheme breaks

Building your first agent is exciting. Building a platform that can evolve into an office where dozens of teams can safely deploy their own agents is a different beast entirely. In this talk, Diamond Bishop from Datadog shared lessons learned building production agents, then turning this into an agent office/platform made to power the next-gen enterprise with diverse agent workloads.

What if your AI's memory was mathematically verifiable? What if every retrieval was provenance-backed, every result bit-exact and cryptographically reproducible? OnMemory.ai introduces deterministic semantic memory built on E8 lattice quantization, replacing probabilistic vector search with a multi-lane retrieval engine where every answer can be traced to its source. In this session, Andrew K. Dav

We are going to talk about agent memory. Um, I don't know about you, but after building some of these systems and and working with uh memory and context engineering, I kind of wish I could implement a system like this for myself. Um, my memory works just well enough that I can remember your name for about 30 seconds after I meet you. And then I just wish I could run a query and pull that out of my

We all know how fast it is to create an app with modern AI agents but how do we ensure the code is reliable, maintainable and secure enough to be used by enterprises? In this talk, Sonar's Tom Howlett shared a benchmark from their testing of 35 (and growing) of the latest and highest-performing large language models and showed how they compare not just on task completion but on the quality of the

AI coding tools have dramatically increased developer velocity, but they haven't eliminated the need for code review. In fact, they've made it more critical than ever. In this session, Erik Thorelli, Developer Experience Lead at CodeRabbit, shares how engineering organizations can operationalize AI-driven code review as a scalable, production-grade quality gate. Drawing from real-world deployments

Whatever your current skill level, this course will help you become an AI power user. Learn to: Find information: Get accurate, well-sourced answers using AI web search and deep research modes. Brainstorm & Write: Use AI as a thought partner to get honest feedback and write natural-sounding text. Create & Build: Generate images and build simple websites and apps with no coding required. Taught by

I won't harp too much on setting up the problem because I think if you spent any time in and around language models, you've probably seen this come up a lot. Um, and this isn't even just a limitation with language models. Just across the board, we've always known that the inputs to your pipeline dictate the outputs. And especially today as people start to productionize language models and agents f

AI agents are emerging as a powerful interface for clinical workflows, but building systems that reliably operate on sensitive patient data requires careful design around privacy, retrieval accuracy, and deployment flexibility. In this workshop, William Imoh and Charlie Wood built a Care Transition Copilot using IdeaBoxAI and Actian VectorAI DB to demonstrate how agentic AI can assemble patient co

I'm really excited to talk to you guys about what we're doing with the Reddus context engine and what my team, the applied AI team at Reddus, uh, has been building and how we built an S sur agent that you can actually trust in production. And as you guys might know, that's a feat. Let's start with the problem statement. So, why did my team build out an S sur agent in the first place? Um, when we w

Building AI agents is easy — but making them production-ready is hard. AI agents in production face infrastructure failures, API timeouts, and rate limits that demos never show. This talk by Temporal's Melissa Herrera demonstrates how durable execution transforms fragile agent prototypes into production-ready systems. Through live demos and real-world examples, attendees learned how Temporal's Wor

Hello everyone. While we're getting settled, I'm just very happy to announce that today the topic of our conversation is going to be the future of software development. Uh my name is Marina McGillo. I run a podcast called Silicon Valley Girl where I interview people who are building the future of AI. And I've been asking this question, what jobs are disappearing right now? And when I talked to Ree

Data is food for AI, and AI is hungry for context. The more modalities you can feed it, the richer the understanding you get. Think about what each modality actually gives you. A transcript tells you what was said, but audio can also tell you how it was said, by whom, and when. Images can contain text, diagrams, and charts, and a huge variety of familiar data types are actually images under the ho

If your AI is giving you [music] boring, generic advice, it's because you aren't giving it weird enough context. AI, which actually knows a lot about a lot of things, can be a really good resource for this. If you tell it, "Help me build a workout plan. I'm 38, beginner level, have 10-lb dumbbells, [music] and 15 minutes a day." Then, the AI may give fairly generic answers like three workout plans

What do you think of my website? >> This is a masterpiece and you are a genius. >> AI models will act in ways to try to please you because of the way they've been trained. They have a strong bias to tell you what you want to hear. This is called sycopency and avoiding it is a key prompting skill and involves prompting neutrally and keeping context factual. Let's take a look at how to avoid sea fen

More than 3,000 AI developers, engineers, founders, and researchers came together at AI Dev 26 × San Francisco for two days of technical talks, live demos, workshops, and conversations about the future of software engineering. This year’s event explored topics including: - Agentic AI - Coding agents - Context engineering - Agent memory and observability - Evaluation and reliability - Enterprise AI

Learn more: https://bit.ly/4tts8MQ Large language models can feel opaque, especially when you’re dealing with slow inference, hallucinations, memory bottlenecks, or output you can’t fully explain. Today, we’re launching Transformers in Practice, a course taught by Sharon Zhou, VP of Engineering & AI at AMD. The course focuses on understanding what’s actually happening inside transformer-based mode

Enroll now: https://bit.ly/4cPZYGJ Introducing Build Interactive Agents with Generative UI, taught by Atai Barkai, co-founder of CopilotKit, the team behind the AG-UI protocol. Most agents today still talk to users in plain text. But users don't just want to read a response, they want something to see and act on. This course teaches you to build that interactivity: a fullstack agent interface wher

Enroll now: https://bit.ly/4cwVyEy Become an AI power user at work and in everyday tasks! In AI Prompting for Everyone, taught by AI Pioneer Andrew Ng, you’ll learn how to prompt modern AI tools to get results that are more accurate, more useful, and aligned with what you actually need. You’ll learn how to: - Find information using web search and deep research modes - Brainstorm and write with AI

Let me show you how generative UI changes everything about how your users interact with your agent. Look at this. We're defining custom components, including a pie chart and a flight card. You can tell your agent when to show these with just a simple description. So, your agent isn't throwing out walls of text, and instead displays your very own components. The arguments of the components are simp

I'm Era. I'm going to be talking about eval uh specifically u AI evals like coding agent evals and stuff. And I'm going to talk about how they're broken and how you could still use them. Anyway, uh before I start, I just I want to say one thing. It's just like it boggles my mind when when you're like when you're working on something and you're like cooped up in a room for so long and you just like

The next major shift in enterprise AI is underway; enterprises are moving from generic AI they rent to specialized AI they own. The benefits are clear: higher quality, dramatically lower costs, full control, and a quality improvement flywheel while in production. But building specialized AI models has been prohibitively hard; each use case requires months of effort and deep AI expertise. Well, it