AI Legend Andrew Ng Warns: AGI Is Decades Away

One email. One eighteen-year-old student. And one question that revealed just how much the exaggerated expectations surrounding artificial intelligence are harming us.

Andrew Ng, founder of DeepLearning.AI, co-founder of Coursera, creator of Google Brain, and one of the most influential people in AI worldwide, recently received a message from a young person preparing for college. The email's subject line read: "An Eighteen-Year-Old's Dilemma: Is It Already Too Late to Contribute to AI?"

The student was genuinely afraid. He feared that by the time he graduated, AI would be so advanced that there would be no meaningful work left for him. That he would end up receiving universal basic income, with no chance to make a difference. Ng replied, reassured him, and encouraged him to learn how to build with AI. But the email prompted him to think more deeply. How much is the hype surrounding AI hurting us?

AI is amazing. And incredibly stupid at the same time.

Ng says it bluntly. He uses AI every day himself and builds things with it that he could not have built a year ago. That is a fact. But at the same time, he would not entrust any leading language model with independently managing his calendar, screening job applicants' résumés, or choosing his lunch—and these are tasks that companies routinely entrust to junior employees.

One of his teams spent a great deal of time customizing an AI tool for screening résumés. The result? A decent assistant. But getting there required an enormous amount of custom engineering work. And that is exactly the point that gets lost in the media noise.

Large language models (LLMs) can handle a broader range of tasks than previous generations of AI tools. That is true. But compared with what a human can do, they are still highly specialized. They work better with text than with other types of data. They require a great deal of custom configuration for each specific application. And we have very few tools that enable systems to learn from feedback and repeated exposure to a particular task.

AGI? Expect decades, not years

Ng is not turning a blind eye to reality here. Artificial general intelligence (AGI)—AI capable of performing any intellectual task that a human can—is decades away, perhaps even longer.

A person can learn to fly an airplane in twenty hours, write a doctoral dissertation in a few years, or learn to drive a truck. Most people are capable of this. AI? We are still very, very far away.

In an interview with Fast Company, Ng admitted with a touch of humor: "Maybe a year ago, AGI seemed fifty years away. Over the past year, we may have made a solid two percent of progress. So we still have forty-nine years to go." The numbers are metaphorical, but the point is clear.

The problem is that some AI developers are quietly rewriting the definition of AGI. They are lowering the bar so that they can declare success. Ng sees this as a dangerous trend. If a company announces that it has achieved AGI, that does not mean AI has reached human intelligence. It only means the company has adjusted the definition to make it look like a victory.

The hype contains a grain of truth. And that is the problem.

What makes these exaggerated expectations so insidious is precisely that they are not complete lies. AI really does do amazing things. But not to the extent claimed by the media and the PR departments of technology companies. For people without a technical background, it is then almost impossible to distinguish where the truth ends and the marketing narrative begins.

Ng compares it to the situation in the investment market. He knows venture capitalists who are afraid to invest in startups at the application layer because they worry that major AI companies will quickly swallow up their businesses by improving their models. Some thin wrappers around language models will indeed disappear. But a vast number of valuable applications will remain relevant for a long time, despite the rapid progress of foundation models.

Where is the market genuinely at risk of overheating? Ng points to the model-training layer. That is where the largest investments are concentrated and where specialized hardware is being built that is then difficult to repurpose. By contrast, he considers the application layer healthy and would like to see even more investment in it.

Learn to code

Ng has a clear message for everyone—students, marketers, recruiters, and financial analysts. Learn to code. Not so that you can become a software engineer. But because the ability to tell a computer exactly what you want is a skill for the future.

At a Salesforce conference, he shared the story of one of his marketers. She could not find an app for "dial testing," so she wrote one herself in two days with the help of AI. She is not an engineer. She is a marketer. But because she knows how to work with code, she did not have to wait three weeks for a vendor.

Ng says that when hiring, he increasingly prefers people who know how to build with AI. "Just as it has become unthinkable to hire someone who cannot search for information online or use email, I now hesitate to hire someone who cannot use AI for automation or creation."

And what about the advice given by those who tell young people not to learn programming because AI will take it over anyway? Ng calls it some of the worst career advice ever given. With every wave of technology—from punch cards to Python—the number of people who program has increased, not decreased.

Agentic AI: A huge opportunity waiting for builders

Ng popularized the term "agentic AI"—systems capable of planning, reasoning, and carrying out multiple steps without constant human supervision. Today, the term appears on every billboard and in every investor presentation. Marketers have slapped it onto everything.

But behind that marketing noise lies a genuine opportunity. Agentic workflows are developing rapidly and delivering real value. Ng sees enormous potential for companies that stop optimizing individual steps and start redesigning entire workflows from the ground up.

The biggest challenge for large companies? The data layer. Until now, data engineering has focused on structured data and tables of numbers. Generative AI brings the ability to process unstructured data: text, images, audio, video, and PDF documents. Companies that can combine sales data, social media trends, and even the weather within a single architecture will gain a huge competitive advantage.

Ng advises companies to proceed step by step: first master prompting, then fine-tuning models, and only then consider reinforcement learning. Rapid prototyping, testing with users, collecting feedback, and continuous improvement—that is the formula that works.

AI geopolitics: The world is fragmenting

Ng sees another major trend that is often overlooked in discussions about AI. The world is fragmenting, and countries want their own AI capabilities without depending on other nations or specific companies they may not fully trust.

Sovereign AI is becoming as much a political issue as an economic one. Whoever controls the infrastructure and models holds the power. And as AI becomes increasingly important for economic growth and national security, this issue is gaining significance.

Go and build. Right now.

Ng concludes his message simply and clearly. There are incomparably more opportunities in agentic AI than there are people with the skills to seize them. This is the best time to enter the field. Not five years from now. Now.

That eighteen-year-old student who feared he would arrive too late? Ng replied that there would be enough work for him for decades to come. And he is right. Because AI, fascinating though it may be, still needs people to build it, adapt it, fix it, and direct it toward the areas where it can genuinely help.