Agentforce> Andrew Ng on Building with AI: Speed, Smarts, and Scale video

Building with AI Transcript

Introduction of Andrew Ng

Silvio Savarese: Welcome, everyone. It's a great pleasure to be here again. It is also a great pleasure to introduce Andrew Ng, someone I had the privilege of working with while at Stanford. Andrew is the founder of DeepLearning.AI, a managing general partner of AI Fund, and executive chairman of Landing AI. Before that, he built Google Brain from the ground up and led a 1,300-person AI team at Baidu. Among countless awards and recognitions, he is also a TIME 100 Most Influential Person in AI.

But Andrew is not just a pioneer in machine learning; he’s also transformed how millions learn about AI. Think about it—8 million people have taken your classes through Coursera, which you co-founded with Daphne Koller in 2012. It’s been amazing to see Andrew’s dedication to making AI accessible and helping accelerate the adoption of responsible AI worldwide. Please welcome a dear colleague of mine, Andrew.

Andrew Ng: Thank you. It's very exciting.

The Evolution of Research: Academia vs. Corporate

Silvio Savarese: Andrew, let me go back a few years. We were both full-time professors at Stanford during the last "AI winter." Since then, AI has evolved from a niche topic into the mainstream. From your perspective, how has the role of academic and corporate research evolved?

Andrew Ng: I find there's a lot of buzz about corporate research, and that’s fine. What’s underappreciated is that academic research at great institutions like Stanford, Berkeley, MIT, CMU, and many others is also extremely important. Corporates have larger marketing arms, so their research tends to be amplified, but there is still incredible research happening in academia.

One of the nice things about academia is that, as a neutral party, academics get to talk to all the leading corporations. A lot of knowledge is still centralized there. We now live in a world where corporates, big companies, and startups are all doing great work. It’s simply a world with more opportunities than there are people with the time and skill to pursue them.

Silvio Savarese: It’s amazing how the landscape has changed. Even in my own research organizations, we’ve been looking at how to reposition ourselves within this new environment emerging from both academia and enterprise.

The Hype of Artificial General Intelligence (AGI) and the Need for Reality

Silvio Savarese: You’ve been vocal about how AI is overhyped, particularly AGI (Artificial General Intelligence). Why do you think it’s overhyped, and what should we do to provide a more balanced view?

Andrew Ng: With the rise of GenAI a few years ago, the technology was so new that a handful of businesses almost got away with saying anything. At that time, media and social media weren't great at fact-checking these companies, so PR efforts wound up distorting the public’s technical understanding.

Regarding AGI: the original definition is AI that can do any intellectual task a human can. I still think we’re decades away from that. However, in the last couple of years, there have been so many alternative definitions of AGI that the term has lost its meaning. Developers writing software don’t use that term in our daily lives because we’re busy building things that work. AGI is used in mass media and by CEOs trying to generate fundraising buzz. Most of us just focus on the tangible work of building AI agents and things that matter.

Silvio Savarese: Do you think the benchmarks for evaluating AI should be redefined?

Andrew Ng: I don’t actually know of any good benchmarks for AGI. You can slap an "AGI-ish" logo on a benchmark, but that doesn't make it accurate. Instead, we see buckets of value emerging in specific verticals. For example, ChatGPT has momentum in the consumer journey, while Gemini and Anthropic are strong competitors. Then you have specialized verticals like coding, medical, or retail. These applications are wildly exciting, but they don't require the word "AGI."

Silvio Savarese: At Salesforce, we introduced the term EGI (Enterprise General Intelligence), focusing on AI for enterprise use cases. These are much easier to define and evaluate for accuracy and consistency in a business context.

The Rise of Agentic AI

Silvio Savarese: You helped coin the term "agentic AI." Now we are seeing a wave of agentic systems used for planning and action within enterprises. What excites you most about how teams are using these systems?

Andrew Ng: A couple of years ago, I saw researchers arguing over the definition of an "agent." I felt that debate was consuming unproductive bandwidth. I suggested we stop arguing and instead view it as a spectrum of "agenticness." There are highly autonomous agents that can plan and take multiple actions, and less autonomous ones that carry out a few steps.

Even though the "agentic" sticker is now slapped on everything by marketers, the productive use cases have also taken off incredibly quickly. We shouldn't wait for AGI to solve all our problems; the work ahead is taking valuable business workflows and implementing them piece-by-piece into agentic workflows.

The Data Layer and Orchestration

Silvio Savarese: You’ve written about the importance of an agent orchestrator to manage specialized agents. How do you see this hierarchy evolving?

Andrew Ng: It’s becoming easier to build complex agentic workflows with tools like Salesforce’s tools, LangGraph, or CrewAI. However, the intellectual exercise of what to build and how to evaluate it is still hard.

The biggest challenge for large businesses is the data layer. Until recently, data engineering focused on structured data (tables of numbers). With GenAI, we can process unstructured data—text, images, audio, video, and PDFs. Businesses need to re-architect their data layers to manage both. If you have a data architecture that connects dots between your sales data, social media trends, and weather patterns, that is incredibly valuable.

The Future of Coding and Skills

Silvio Savarese: Some leaders have suggested people shouldn't learn to code because AI will automate it. What is your take?

Andrew Ng: I think that will be looked back on as some of the worst career advice ever given. As coding becomes easier, more people should do it, not fewer. We went from punch cards to keyboards, and from assembly to Python. With every wave, more people wrote code.

The best developers I know today aren't just fresh college grads; they are seasoned developers with 20 years of experience who are also on top of the latest AI tools. They move at a speed the world has never seen. I have often hired a fresh grad who knows AI over a veteran developer who is still coding like it’s 2022.

Silvio Savarese: I agree. The mindset has to shift from doing everything ourselves to delegating tasks to AI agents to expedite development.

Andrew Ng: Exactly. A few weeks ago, I saw someone in a coffee shop coding by hand. It looked so strange! It turned out to be a student doing a homework assignment where the professor required it. I hope to almost never code by hand again, except for maybe niche GPU kernel stuff. For Python or JavaScript, we get AI to do it and orchestrate it like a flock of agents.

AI Education for the Next Generation

Silvio Savarese: For students or parents of kids, what should they be studying today to be successful?

Andrew Ng: Tell them to learn to code. Not necessarily to become a software engineer, but because the most important skill of the future is the ability to tell a computer exactly what you want it to do.

I’m seeing marketers, recruiters, and finance professionals on my team who know how to code, and they are much more effective. One of my marketers couldn't find an app for "dial testing," so she spent two days writing the app herself using AI. She’s a marketer, not an engineer, but she empowered herself to create a solution rather than waiting three weeks for a vendor. Being a software creator gives you more options than just being a software user.

Conclusion: Scaling and Reinforcement Learning

Silvio Savarese: We’ve seen that the "scaling laws" might be reaching a limit. At Salesforce, we believe the path forward is for agents to learn from experience through Reinforcement Learning (RL). What is your take?

Andrew Ng: Self-improving agents and long-term memory are active research topics. RL is a great technique, but it is currently very complex. I often tell businesses: master prompting first, then fine-tuning, and only then look at RL. It’s a step too far for many businesses right now, but as the technology simplifies, it will become more accessible.

My final thought is this: Go build. The set of opportunities available now through agentic AI is vastly greater than the number of people with the skills to build them. Be responsible, but go build a lot of stuff.

Silvio Savarese: Great insight. Thank you, Andrew.