Andrew Ng explained where the main opportunities for AI are located.

Why there is so much fear around AI

There is indeed a lot of misinformation surrounding AI. Over the past two to three years, several major AI companies have actively used fear-based rhetoric about the technology, including in the context of regulation. For companies that have spent billions of dollars training large models, it is in their interest to prevent other players from easily creating and distributing their own models.

Fear works well. When AI is compared to nuclear weapons or when isolated cases of model errors are highlighted and presented as evidence of a systemic problem, public perception of the technology becomes much more negative. The same happens with discussions about the energy consumption and water usage of data centers, where individual metrics are taken out of context.

As a result, AI adoption slows down. At the same time, the technology offers huge benefits, although it does have problems. We need to talk about them, but not exaggerate the scale of the risks.

What students should learn

One of the main problems for graduates is that universities are adapting to changes too slowly. When the way software is developed changes, it may take teachers a year or two to master new skills, then new courses need to be created, curriculum approval procedures must be passed, and internal university coordination must be obtained. This takes years, whereas AI changes much faster.

Many universities still prepare students for jobs that existed in 2022, although they should already be preparing them for professions in 2028 and beyond.

At the same time, there are plenty of jobs. I see companies that cannot find specialists with the necessary skills at various levels.

I currently have interns, students, recent graduates, and even one schoolboy working with me. They are productive because they are AI-native specialists. They use AI where it truly helps, and then focus on tasks that humans perform better.

Therefore, students should study well at university, gain knowledge from teachers, and get good grades, but at the same time independently master skills that the university program has not yet had time to include. For this, you can use Coursera, DeepLearning.AI, Udemy, and other educational platforms.

Why AI is not well suited for learning

I use AI models a lot for work and consider them an excellent tool. But for learning, the situation is completely different. Data shows that students receive higher grades on homework when they use AI. This sounds good, but long-term retention of material suffers as a result.

The reason is cognitive offloading. When AI does the work for a person, society as a whole gains in productivity, but the individual remembers less.

I see this even in my own experience. Over the past six months, I have asked AI models many times how a particular frontend or backend component works. I received an answer, quickly solved the task, but a few months later I no longer remembered how it worked and turned to AI again.

Therefore, we should stop perceiving modern AI models as a universal tool for learning. They can be used for education, but most of the ways people use them now rather help get work done than actually learn something.

This is exactly why I am involved in creating LearnVector. We want to make learning more personalized and move from the "one teacher – many students" model to a format that is much closer to individualized learning.

More independence and opportunities

Another important change is related to the independence of specialists. AI gives people more opportunities to notice problems, independently create solutions, and go beyond their formal job description.

I increasingly talk to engineers and other specialists in large companies whose managers tell them to stay within their area of responsibility. But there are more and more opportunities to do something outside of it.

It seems to me that companies that create a culture where employees can study AI, quickly and responsibly create solutions, and talk to customers will be able to get more value from their teams.

The same applies to hiring. I expect that the ability to act independently and create something new will become an increasingly in-demand skill.

In my team, for example, all marketers know how to code. During interviews, we ask what a person has already created. One of our marketers made his own application for searching and analyzing materials on the internet. It finds related articles, allows interaction with the collected information, and helps him when preparing materials.

The finance team is also actively using AI. One of the managers noticed that employees spend hours every week opening documents, copying data, and checking information. As a result, automated scripts appeared that independently open files, check their content and data consistency, and notify the team if something requires attention.

We also have engineers who work directly within the recruiting team and create specialized tools for hiring.

This does not mean that engineers will cease to be needed. On the contrary, new directions are emerging — marketing engineering, recruiting engineering, HR engineering, and others. There are more opportunities for engineering work.

Children and AI

Children have a great future, and now is an interesting time to grow up with tools that previous generations did not have.

At the same time, I am concerned about the impact of AI on learning. I have a five-year-old and a seven-year-old child. While they are small, I can, for example, teach them mathematics without a calculator and ask them to solve problems independently.

As they grow older, I am more concerned about cognitive offloading through AI and how it may affect long-term retention.

At the same time, digital tools can be useful. I even created my own application to help my daughter learn to type, because I was not satisfied with existing free programs.

I think that responsible use of digital tools under adult supervision can be beneficial for children. The problem is that many parents do not have time to constantly monitor exactly how their child uses technology.

Where to find opportunities in AI

Currently, the cost of creating products has dropped sharply. Therefore, I advise learning to work with AI, quickly create solutions, and talk to customers. I myself am constantly creating something — practically every week or weekend. If I or my team encounter a problem, I can try to make an AI tool that solves it.

Recently, I used a state-of-the-art AI model to analyze key business metrics because I did not have time to figure out the data myself or find a data specialist for this. At the same time, I carefully selected the model from the point of view of data retention policy.

Due to AI, the main limitation is no longer the cost of creating a product, but the choice of what exactly to create. There is a kind of bottleneck in product development.

Therefore, people who can talk to customers, understand their needs, evaluate ideas, and quickly create solutions using AI are especially important. At the same time, building a company has not gone away as a complex task. You can assemble a simple application over the weekend, but building a large company is much more difficult.

For this, deep technical expertise, understanding of customers, and integration with them are still needed. You can widely experiment with different ideas, but then it is important to choose a few directions and delve deeply into them.

When will AGI appear

It all depends on exactly what is considered AGI. I am closer to the definition according to which artificial intelligence should be able to perform any intellectual task that a human can perform.

For example, a person can spend five years studying and writing a dissertation. Or can learn to drive a truck in a complex new environment after a few dozen minutes of practice.

AI is not yet capable of reliably doing such things. There is a long list of tasks that modern systems cannot handle at a human level.

Therefore, with such a definition, AGI is still on the horizon of several decades. At the same time, different people use different definitions, and if the bar is lowered, one can say that AGI already exists or existed decades ago.

For me, the wide criterion is important: the system must be capable of performing any intellectual task of a human. By this criterion, AGI is still far away.