Самые большие возможности ИИ — не там, где вы их ищете | Сооснователь Coursera Эндрю Ын

There is truly a lot of misinformation surrounding AI.

This is Andrew. He is the founder of Google Brain and Coursera. Millions of

people have taken his machine learning course, and today he is one of the most influential experts in the field of artificial intelligence.

Leading tech companies are actively stoking fear regarding artificial intelligence. They are trying to push

intelligence. They are trying to push for specific laws, and this constant wave of fearmongering is heavily distorting public perception in a negative way.

I constantly hear about data centers and job losses. And it is capable of performing 30-40%of tasks in many professions. And this means the value

professions. And this means the value of the remaining 60%, which humans do, is increasing.

Humanity is losing control over artificial intelligence.

There are other things that we do not have control over.

You have vast experience in machine learning and teaching skills related to artificial intelligence. And at the

artificial intelligence. And at the same time, you are one of those who speak about this technology in a positive light. This summer, I noticed

positive light. This summer, I noticed even more how polarized society is right now. Every time I write about

right now. Every time I write about artificial intelligence on social media , I hear back about data centers, climate issues, and job losses. Why do

you think this wave has overwhelmed us so strongly right now? What are its causes?

There is truly a lot of misinformation surrounding AI. And I think the root of

surrounding AI. And I think the root of many of these problems is that 2-3 years ago, an active PR campaign began, along with attempts by labs to influence legislation. One of the most

influence legislation. One of the most valuable things in AI is massive language models, on which companies have spent billions of dollars to train . And if you have invested billions in

. And if you have invested billions in training a model, it is, of course, not in your interest when someone else creates their own model and distributes access to it for free. Therefore,

several leading AI companies are actively stoking fear around this technology to push for laws, and any regulation will create an uneven playing field in favor of existing players. As a result, access to it will

players. As a result, access to it will become more expensive, and other teams, whether researchers or companies that want to freely distribute open-source models so that anyone can use them at a lower cost, will have their capabilities limited. Unfortunately,

capabilities limited. Unfortunately, fearmongering works. Comparing AI to

fearmongering works. Comparing AI to nuclear weapons has no proven basis whatsoever. They have absolutely

whatsoever. They have absolutely nothing in common. It is the same when they cherry-pick individual instances where AI made a mistake. It is making a mountain out of a molehill. It is also misinformation that data centers use much more water than they actually do.

This stream of fear-based messages has severely distorted public perception.

And AI in general is leading to a slowdown in the adoption of this technology in America, which makes the country less competitive. And if we ignore the truth about AI, that it is a technology that brings enormous benefits, despite the fact that it also

creates problems—but not as large-scale as they are made out to be —then that will harm people.

I will cite a few points that people write to me about: "Job loss, social inequality." What do you think? When

inequality." What do you think? When

people talk about a shortage of jobs, they imagine that AI will take 50%of jobs, people will be left unemployed, and will be dying in the streets. That

is simply impossible. With every new technological wave, including the current one, the skills required for work also change. Artificial

intelligence is indeed changing professions. However, AI is not yet

professions. However, AI is not yet working that well. Some companies like to exaggerate the capabilities of AI and say that they have a superintelligence that can do everything a human can do. I would like AI to be better. We just don't know how

to make it do everything a human does yet. Economists, for example, my friend

yet. Economists, for example, my friend Erik Brynjolfsson from Stanford or Andy McAfee from MIT, analyzed human work by breaking it down into individual tasks.

Perhaps AI is capable of performing 30-40%of tasks in many professions. And

this means that the 60%performed by a human will become even more valuable, because they complement those 30-40%, which, in turn, become cheaper.

Therefore, people who use AI will possibly displace those who do not use it. Nevertheless, artificial

it. Nevertheless, artificial intelligence is not yet capable of replacing humans in the vast majority of professions. Out of all the

of professions. Out of all the different professions. It is currently

different professions. It is currently having the strongest impact on software development, because it writes code excellently. And we see that the number

excellently. And we see that the number of vacancies for developers is growing, contrary to all the alarmists ' forecasts. And all the good developers

forecasts. And all the good developers I know are currently swamped with work.

On the other hand, if a person still writes code like it’s 2022 before ChatGPT, they have a problem; they need to acquire new skills, don't do what AI can automate by 30-40%. Let AI handle that. And you should develop skills for

that. And you should develop skills for the remaining 60 or 70%, which cannot be automated with artificial intelligence.

What would you advise recent graduates?

I just spoke with Erik Brynjolfsson, a Stanford economist, on this podcast, and he said that the labor market as a whole has not changed that much, except , it seems, for jobs for people aged 18 to 25 who have just graduated. What

would you advise those who don't have experience yet on how to build a career on their own? To people who can do the same work that AI is already capable of doing. And and

doing. And and One of the real problems for graduates now is that the education system is inflexible and adapts slowly. I I am a fan of classic education and believe we should all support universities. But

when AI changes how things are written, it takes universities a year or two for faculty to master skills, create new courses, and complete approval processes. It takes years. These speeds

processes. It takes years. These speeds

do not match the speed of AI development. Many universities are

development. Many universities are still training students for 2022 jobs, although it's already too late to prepare them for 2026 jobs. Students

need to be prepared for 2028 jobs and, of course, even further out. Yet, job

openings exist. Many employers cannot find people with the right skills for senior positions, for senior positions of any kind. We have a lot of interns.

These are current students and recent graduates. And they are amazing

graduates. And they are amazing employees and very productive because they are tech-savvy with AI. They

delegate to AI what I used to do myself , focusing instead on what AI won't be able to learn for a long time.

Therefore, there is plenty of work for people. But my advice to students and

people. But my advice to students and graduates is: study hard, get good grades, and learn from your professors.

But if your university hasn't provided relevant skills, you can acquire them on other platforms, like Coursera, DeepLearning.AI, or Udacity. Especially

master modern AI-related skills. Which

universities have not yet adapted. I

want to add something else. If you look at how the skills map is changing, one of the most important shifts is that it's now much easier to create things with AI. And when something becomes

with AI. And when something becomes much easier, more people should learn how to do it. Now, it shouldn't just be developers creating programs with AI.

It has become much easier for a wide range of people to create them. Those

who know how to use it will be more productive, will be able to do more, and, I think, will enjoy their work more. And AI allows for creating new

more. And AI allows for creating new things much faster. That's why

marketers, recruiters, HR specialists, operations staff, and not just developers should learn to create with AI. Then they will be able to do much

AI. Then they will be able to do much more.

And by the way, how do you measure productivity growth after implementing artificial intelligence? Does your

artificial intelligence? Does your company have any KPIs?

I wish there were a simple answer to that. I believe that the result from AI

that. I believe that the result from AI depends more on the business itself than on the AI. For some, it's growth and customer retention; for others, it's faster customer service or increased clarity in task delegation.

KPIs are tied to the business itself, not to efficiency. And

that is, you cannot measure the effect directly. We are actively implementing

directly. We are actively implementing artificial intelligence in the company.

And for me, as a media company, that's probably the number of views and the volume of content produced.

Yes, that is a good indicator.

It's interesting how differently you can measure the impact of artificial intelligence. Sometimes it even means

intelligence. Sometimes it even means revenue growth.

Well, yes.

And in general, as the team grows, you quickly realize that you need to scale not just content production, but all processes around the remote team, because the number of people just keeps increasing. And here we also have a

increasing. And here we also have a solution that has greatly simplified our business life. I’ve mentioned

many times that I have a distributed team, and the larger it got, the more time we spent onboarding new contractors. Each one needs a contract

contractors. Each one needs a contract and closing documents that meet uniform standards. Then you need to store it

standards. Then you need to store it all somewhere, make sure nothing gets lost, and be able to find it quickly when needed. For us, this part was

when needed. For us, this part was significantly simplified by fodf.com, which we’ve been using for several years now. Through the platform, you

years now. Through the platform, you can onboard and manage contractors from over 150 countries. Documents are

generated automatically and stored in one place, and you can onboard a new person in just a few clicks. I

personally like that you can secure the rights to the work results right at onboarding if the project requires it.

This means you don't have to look through old correspondence months later or go back to someone to finish paperwork. If you also have a

paperwork. If you also have a distributed team and want to eliminate this administrative routine, sign up at demofdf.com using the link in the description. The team will show you how

description. The team will show you how to organize processes better specifically for your tasks.

Tell me a bit more about how you use it .

First of all, we all have Claude. There

are separate projects for each social network. For this podcast, for example,

network. For this podcast, for example, we have a "Guests" project. It has

analytics on previous episodes and the criteria by which we evaluate a potential guest: whether they’ve been cited, whether they have a specific opinion on AI, whether they actively use it in their company, or perhaps they recently founded an AI company that is showing great results. Each

criterion is assigned a weight, resulting in a score out of 40 points.

40 is the first level, 30 is the second , and so on. There is also a project that analyzes each podcast and gives me advice on how to ask questions. We have

the same thing for Instagram and LinkedIn. Each project knows how I

LinkedIn. Each project knows how I speak. It has my personal file,

speak. It has my personal file, business strategy, and knows facts from my biography. At the same time, there

my biography. At the same time, there is a separate person for each social network. So, the strategic decision is

network. So, the strategic decision is always made by a human. Right now,

I’m trying to close the loop, because I’m often sent text that I correct manually, telling them what needs to be changed. And this happens in Telegram,

changed. And this happens in Telegram, where we have a bot that scans all those chats. And I want my agents to be

those chats. And I want my agents to be able to learn from this feedback and better understand my taste.

There is one peculiarity of artificial intelligence that I constantly notice.

If you use it as an assistant for data analysis or a brainstorming partner, it often provides one or two good ideas, two or three mediocre ones, and four absolutely terrible ones. And you

wonder how it even decided that this idea could be acceptable. And this is the answer to the fear about a lack of jobs. Humans will have a huge advantage

jobs. Humans will have a huge advantage for a long time because they possess context. For example, if something is

context. For example, if something is very obvious and plain to a person, such as the fact that an idea is terrible, artificial intelligence does not pick up on that. Therefore, in the near future, one should not expect mass replacement of humans by artificial

intelligence. People possess a huge

intelligence. People possess a huge amount of context thanks to years of experience. We talked to a client, saw

experience. We talked to a client, saw the look on their face, and realized they didn't like our proposal. Or a

manager said what was important to them . And it turns out that almost every

. And it turns out that almost every person, if not to say everyone, possesses context from a huge number of things that AI doesn't know and, I think, won't know in the foreseeable future. We often talk about critical

future. We often talk about critical thinking and human taste. And someone

might ask, "What is taste?""Is it some kind of abstraction?" But in my view, behind the fact that a person evaluates and senses things better, there lies a quite specific thing. The advantage of context. And this advantage is not

context. And this advantage is not going anywhere in the next few years.

Therefore, we will still need many more people who have critical thinking and taste to complement AI.

Doesn't this make education even more important? After all, education gives

important? After all, education gives us context. I hear more and more often

us context. I hear more and more often that education won't be needed because all information is at our fingertips.

Just ask ChatGPT. But when you talk about context and taste, for me that means years of accumulating knowledge, adopting experience from the best of the best, and watching how they work.

It's not just getting an answer from a chat.

Now I will say something that not everyone will agree with. I'm not sure if I've spoken about this publicly, but I believe that AI is completely unsuitable for learning. Many think

that AI is a wonderful tool for completing tasks. They use it

completing tasks. They use it constantly and really love it. But all

the data that is appearing now points to one thing. For example, students get high grades, and we know they use AI, and now research confirms this too.

Hooray, grades for homework have become higher. Wow. But when it comes to

higher. Wow. But when it comes to retention and long-term results, everything is much worse because AI does the work for them, and more and more studies confirm this. I believe

that Wikipedia is a wonderful tool.

There is a huge amount of information there. Internet search is also a

there. Internet search is also a wonderful tool, full of information, but it turns out that when you ask it to do the work for you, you offload the cognitive load onto it. And that's

great, because that's exactly how society moves forward, and we can get more work done. But the person remembers much less of what they actually did. Language models, in the

actually did. Language models, in the form most people use them today, are not suitable for learning at all. I'm

not saying they can't be used for study entirely. I think an approach can be

entirely. I think an approach can be found. Even I myself have constantly

found. Even I myself have constantly been asking language models about completely different things over the last six months. How does a certain front-end or back-end component work?

Tell me, and I will complete the task.

That's cool, but six months later I don't remember the answer anymore. And

when I need to perform that step again, I ask the AI again. So the data is completely unambiguous. It's time for

completely unambiguous. It's time for us to stop considering AI a good tool for learning. At the very least, the

for learning. At the very least, the vast majority of ways people use models today are completely unsuitable for learning.

But you are creating a company that is specifically addressing this problem, right? This is that same personalized

right? This is that same personalized one-on-one AI tutoring you recently talked about.

Yes, I am happy to lead the new organization Learn Vector, which is focused on creating much more personalized one-on-one learning formats rather than group ones. 15

years ago, I was lucky enough to participate in the development of online education, which changed how many people approach learning. But it's

still not a personalized format, since everyone watches the same video, and that, of course, has its benefits. Now,

technology allows for a much more personalized approach to learning. Our

team is actively working on this, and by the beginning of next year, we will be able to show some results. Because

of AI's influence on software development, the IT job market reflects what will happen in other fields as well. Developers have to learn new

well. Developers have to learn new skills, but when they do, they create more value, get promoted, and work on more interesting projects. And I see the first signs of this in other areas,

for example, in development. Many

front-end and back-end developers are now becoming full-stack developers because, thanks to AI, they can take on a wider range of tasks. I see the first signs of this in other professions as well. For example, a person who

well. For example, a person who previously coordinated marketing campaigns can, with the help of AI, become a more versatile marketer and take on a wider scope. And I see the same thing in recruiting. Sourcing

candidates and hiring is increasingly turning into a full-cycle process from start to finish. And this is where there is good news and bad news. To

transition into such broader roles, you really need to master skills for working with artificial intelligence.

But it is not just about AI. You also

need to develop professional skills and understand other aspects of marketing, recruiting, development, or AI engineering. That is why I think there

engineering. That is why I think there is a huge need right now for people to master and acquire new knowledge. And

when they master working with AI and professional skills, they can do much more and, hopefully, get more enjoyment out of their work, take on more interesting projects, and likely earn more. A person who was planning to

more. A person who was planning to enter university wrote to me and said: "Andrew, I'm taking online courses, but I just can't figure out which major to choose." Won't AI be doing my job in 4

choose." Won't AI be doing my job in 4 years? Won't everything I learn be

years? Won't everything I learn be obsolete? And this is one of the

obsolete? And this is one of the reasons why the surrounding hype worries me, and I replied to this young man: "No, far from everything will be obsolete." But when we constantly

obsolete." But when we constantly spread such scary messages, people start doubting whether they will even be in demand. And instead of mastering skills that will help them come out ahead, they prefer to abandon the idea altogether. Therefore, I clearly see

altogether. Therefore, I clearly see how fear-mongering distorts the perception of many people, including schoolchildren, students, and graduates .

You know that my team and I love using Claude. And we taught Claude to think

Claude. And we taught Claude to think and do things for us. So, while I was in Cannes or in St. Petersburg, my

agents kept working. It's clear that we work too, but compared to last year, my team is now creating two to three times more content. And it's done without my

more content. And it's done without my participation, but my face and my real experience are in the frame. A vivid

example of all this is the carousels, which are 99%assembled by the agent. We

publish these carousels on the LinguaMarina channel. All that needs to

LinguaMarina channel. All that needs to be done is to upload my photo for the first slide and click publish.

Previously, we didn't have enough time for carousels, even though for the language niche, carousels are an ideal format. Actually, for any niche,

format. Actually, for any niche, carousels are just booming right now.

Having this agent doesn't mean we suddenly do nothing at all. We are

simply increasing the amount of content . My face appears more often before my

. My face appears more often before my target audience, while my team and I sit and think about strategic tasks.

Where we should head and what else we can come up with. Actually, if you look at my Instagram, I feel like every second post now is a carousel, and it gets much better reach than Reels. And

at the same time, we don't have to spend hours putting these carousels together. And, for example, this

together. And, for example, this specific carousel on LinguaMarina was shared by more than 2,000 people. I

think this turned out to be a very successful format. Our goal with this

successful format. Our goal with this agent is to automate routine tasks, not to create that generic AI junk. In my

newsletter, I’ve put together a step-by-step guide for you on how to set up such an agent for yourself, your niche, and your audience. And if you read it—it’s completely free—by the end of the evening you’ll have a 30-day content plan, one finished carousel based on that plan, and a schedule in your Claude workspace so

the next batch of content is created automatically. Basically, you’ll have

automatically. Basically, you’ll have a working group right inside Claude.

You’ll have an analyst, an SMM manager, a copywriter, and a designer in there. Each one will be doing their

in there. Each one will be doing their own task. The guide is suitable even if

own task. The guide is suitable even if you’ve never opened Claude before.

Just follow these five steps in order.

And if you’re already a pro, we’ve marked the most interesting settings for you in the "Pro Tip" section.

Subscribe and start learning. How would

you respond to this letter? Which major

do you think is best to choose now to be successful in the era of artificial intelligence? Should one dive deep into

intelligence? Should one dive deep into a narrow field, or, conversely, is it better to choose a broader direction like computer science to master AI skills faster?

I don’t know which major is best.

There is a huge number of wonderful fields. What is the best job in the

fields. What is the best job in the world? The one you love? For example,

world? The one you love? For example,

my daughter wants to become an astronaut. I don’t know if you can

astronaut. I don’t know if you can get a degree in being an astronaut. I

need to look into that, but she might change her mind when she grows up. I

see a huge number of opportunities in many different professions, and it all seems very interesting to me. But learn

to work with AI. Learn to create using AI. My teams are working on an AI

AI. My teams are working on an AI engineer skills map to identify the most important skills for working in this field. I was surprised when I saw

this field. I was surprised when I saw it in the data, but I subconsciously suspected that more and more vacancies require a very high level of autonomy.

Because, as it turns out, people have more opportunities to spot problems themselves and create things to solve them. We are truly changing and growing

them. We are truly changing and growing . And even though we’ve been changing

. And even though we’ve been changing for a long time, this process has accelerated; we are moving further away from the era where people sit and wait for their boss to tell them what to do.

I’ve started to feel that too, especially since we switched to remote work. I really want people to be

work. I really want people to be entrepreneurs within their own area of responsibility. If you handle LinkedIn

responsibility. If you handle LinkedIn at my company, be enterprising: hire contractors, implement new tools, and make strategic decisions regarding the relevance of a particular topic. That

is the right direction. I think we are moving toward a labor market where everyone becomes more self-reliant within their own task.

I think people will have much more autonomy and creative freedom. I agree

with that, but I would take it even further. I talk a lot with engineers

further. I talk a lot with engineers and other employees at large companies who are told by managers not to step outside their areas of responsibility.

An employee says, "I have this proposal ." And the manager replies, "No, we

." And the manager replies, "No, we need to focus on this instead." And

honestly, this often happens because the manager’s own career is on the line. At the same time, it seems very

line. At the same time, it seems very promising to me that people are getting more opportunities to notice things outside their scope and then find ways to realize their ideas. That’s very

cool. And I think that in the future, companies with a culture that encourages people to learn to work with AI will generate much more added value than vertical organizations split into isolated departments.

Yes, it all starts with the right hiring, and then fostering the development of culture within the organization. What do you mean by

organization. What do you mean by learning to use and master AI? Can you

give another example of what an advanced marketer or other knowledge worker should be able to do? I’m

confident that my team is ahead of the market. All my marketers know how to

market. All my marketers know how to program, so during interviews, I ask them what they have programmed. If they

haven’t created any software at all, I’ll take note of that.

If a marketer, for example, has created a dashboard, that’s not bad. Dashboards, again, my team is pretty far ahead of the market.

It’s useful to know how you evaluate that.

All my marketers create much more complex things.

Like what, for example? Well, for

instance, recently one employee from our marketing team was talking about his software, which he uses when he starts writing an article. His tool

searches the internet for materials on the topic on its own. He even built a separate Mac application for it that displays these materials for him. And

you can interact with this tool in a chat, as well as view your article and the related materials. He also has a large dashboard that monitors online publications and shows him new and useful materials. I think this marketer

useful materials. I think this marketer is ahead of everyone else even on my team.

I’m curious, what other interesting ways to use artificial intelligence could inspire people to create something similar?

Well, for example, my finance team uses artificial intelligence very actively.

One of the finance leads noticed that her team spends a ton of time every week opening documents, copying here, pasting there, and manually transferring data. So she started

transferring data. So she started building automations that run on their own, open files, check their contents and reconcile data, and if something important appears or a new document is added, they notify the team. So I think that now teams don't necessarily have

to wait for an engineer to do everything for them. They can build not just dashboards themselves, but also quite complex data processing systems. For example, modifying data and receiving notifications when something happens. That is exactly what I think

happens. That is exactly what I think my finance and marketing teams are doing. My recruiting team has engineers

doing. My recruiting team has engineers , professional developers who work directly with recruiters and build quite complex hiring tools. And that is another trend. I think marketers,

another trend. I think marketers, recruiters, HR specialists, and office staff should learn to work with AI. But

if you add an engineer directly to such a team, they will only strengthen it.

Yes, we have the same approach. We

start with something simple, build it ourselves, then hit a wall, and a developer comes in and helps improve the system.

If you look not only at developers, but also at recruiting, marketing, and HR engineers, there is now an enormous amount of valuable engineering work to be done. I am not worried that there

be done. I am not worried that there will be less work for engineers.

Everyone I know is so busy right now that it's hard to imagine how they could end up out of work.

Yes, there are so many interesting ideas to try, but you touched on one of the fears related to confidential information. How much data do you give

information. How much data do you give to artificial intelligence? I mean, for example, I gave Perplexity access to my Fidelity account so it could track my portfolio and tell me when it’s time to rebalance it. It doesn't do anything on my behalf, but it does have access to the accounts. Do you think that's bad?

That’s a complex question. And

privacy is a complex area. And a lot depends on the company you are giving your data to. I trust the major tech companies and count on them to abide by the terms of the user agreement. And I

would be very surprised if they published privacy policies and then violated them themselves; that would contradict their culture and cause huge damage in the long run. But that’s

the big picture. Among companies, there is at least one, which I will not name, that changes its terms of use from time to time. You go to the site, see a

to time. You go to the site, see a notification about changes to data storage or usage terms, click the wrong thing, and the company gets permission to handle your data in a way I might not like. And I have sensitive

not like. And I have sensitive information from time to time, so I am cautious with companies whose culture and long-term business model are not clearly oriented toward protecting privacy and confidentiality. And this

is very important, because even now I see several different businesses and large corporations that understand this . For example, my team at AIPire works

. For example, my team at AIPire works with large corporations, including banks, on extremely sensitive financial data. And such data will not be

data. And such data will not be voluntarily handed over even to advanced laboratories without a very thorough assessment of protection and privacy measures. So it is a complex

privacy measures. So it is a complex issue.

That is, one can trust big tech companies, but there is another option.

Download a local model, run it on your computer, and then the data is stored only on your computer.

Yes, yes, many banks really do run such systems in a virtual private cloud or within their own infrastructure, so the data never leaves the company. And when

it comes to private users and truly sensitive information, I sometimes run a local model. And here is what is interesting. Some of the latest open

interesting. Some of the latest open models are already approaching the capabilities of advanced models, and they are quite compact. These are

really good models, aren't they? The one from Meta, the

aren't they? The one from Meta, the most recent one, yes, Meta Llama is a good model. And

the latest version of Qwen is also very good, but honestly, these models change every couple of weeks, so it is better not to get stuck on one, but to constantly try new models.

So, if you don't trust anyone, you run a local model and thus keep your data safe. I trust the biggest tech

safe. I trust the biggest tech companies, but when it comes to truly sensitive information that I wouldn't send to the cloud at all, then I either perform the task manually without AI help, or if AI is really needed, I very

carefully use a local model on my computer.

I wonder, what about humans losing control over artificial intelligence? I

spoke with Yoshua Bengio, who is quite negative about AI freedom without regulation. He painted a rather

regulation. He painted a rather frightening picture that we cannot control what is smarter than us. And if

artificial intelligence is becoming smarter, where will we end up? What do

you think about that?

I think about another technology we cannot control: airplanes. It is

impossible to create an airplane that flies perfectly. The wind will rock it.

flies perfectly. The wind will rock it.

In the early years of aviation, planes crashed, people died, and it was truly tragic. But we learned from our

tragic. But we learned from our mistakes and learned to control airplanes better. And today, in most

airplanes better. And today, in most cases, we can get on a plane without much fear for our lives. AI is roughly the same. It is impossible to control

the same. It is impossible to control AI perfectly because it generates tokens and results that are partly random. But as we use it, we learn to

random. But as we use it, we learn to control it better. There have been a few incidents, which is certainly unpleasant, and some of them have caused harm. And practically any system

caused harm. And practically any system , from airplanes and electrical circuits to AI, we design in a controlled environment, carefully increasing its capabilities and measuring what is not working correctly , and then we adjust the system so that

it behaves responsibly and safely. And

to this day, we cannot perfectly control any airplane, and we will also never be able to control AI perfectly.

But I think we already control it well enough so that this loss of control doesn't seem like some science fiction disaster.

And what about deepfakes?

Deepfakes are a problem. One of the most disgusting phenomena I've ever heard of is non-consensual intimate deepfakes. I'm very glad that the US

deepfakes. I'm very glad that the US Congress is moving toward passing laws and introducing penalties. There are

truly dangerous ways to use AI that need to be banned and strictly punished . We just need to get rid of that.

. We just need to get rid of that.

And what do you think about children and their interaction with artificial intelligence? Especially now, when they

intelligence? Especially now, when they are using artificial intelligence more and more. We’ve already seen how

and more. We’ve already seen how social media leads people to mindlessly scroll through their feeds for hours.

My daughter is five now, and when I don't know how to answer her question, she says, "Well, ask ChatGPT." She

thinks that ChatGPT knows everything, including what other people are thinking. What would you say about the

thinking. What would you say about the future of children in the age of artificial intelligence?

First of all, I think children have a wonderful future. It's great to be a

wonderful future. It's great to be a child and grow up in a time with tools that none of us had before. At the same time, I think social media is blamed more than it deserves. But it has still done its share of harm, as it doesn't have the best influence on children.

And that's great too, but I'm very worried about the impact on learning.

When my kids were little, I could keep calculators away from them. When I

taught them to count, I would ask how to multiply numbers and help them solve the problem on their own. Without a

calculator at all, but I worry about students because they rely too much on AI and it harms long-term memory. But

at the same time, I built an app myself . I didn't like anything that was

. I didn't like anything that was available online. I created an app so

available online. I created an app so my daughter could learn how to type.

And now, for a seven-year-old, she already types quite well.

Oh, she already types?

Yes, she already types. She knows how to type lowercase letters. She knows

how to use Shift a little bit for capital letters, but hasn't quite mastered it yet. I think the use of online tools should be under adult supervision, but that's difficult.

Using digital tools under adult supervision can certainly be very helpful for kids, but most adults don't have enough time to monitor how they use them. And at the same time, social

use them. And at the same time, social media platforms themselves encourage certain behavioral patterns. It is

important that the incentives are right . We talked about fears, and we talked

. We talked about fears, and we talked about how one can improve their work with the help of artificial intelligence. Can you name some of the

intelligence. Can you name some of the biggest opportunities in artificial intelligence in 2026 for people who want to build something?

For a person who wants to build something, there is no single answer.

But since the cost of building anything has dropped sharply, I advise learning to work with AI, building quickly, and talking to customers. I am constantly building things myself, almost every weekend, because I or someone on our

team encounters a problem and I implement an idea to automate it using AI. Last weekend, I used one of the

AI. Last weekend, I used one of the newest models to analyze a large number of our key business metrics because I didn't have the time to do it myself.

But I didn't go looking for a data scientist to help me with it. So, I

tried several cutting-edge models, very carefully reviewing their privacy policies. I didn't use models whose

policies. I didn't use models whose terms I wasn't comfortable with to analyze the data. And that is what is happening with AI; the cost of creating and writing applications has dropped sharply, so the problem shifts to the

question of what exactly to build. I

call this the product management bottleneck. Founders, engineers, and

bottleneck. Founders, engineers, and product managers can talk to customers, understand what their needs are, and then quickly build it using AI and finish it off. There is a huge amount of opportunity right now. If you are a beginner, what should you focus on if

you want to build something over a weekend? Or can you now, thanks to

weekend? Or can you now, thanks to artificial intelligence, work on several products simultaneously and launch multiple business ideas?

Building a company is still very difficult, so it is very important that someone is completely focused on one direction and works only on that. Over

a weekend, I can often make a simple wrapper or put together an app, but I wish it were just as easy to build a big company. To create something truly

big company. To create something truly significant, you need either deep technical knowledge or a deep understanding of customers. Yes, you

can now write code in a few hours using AI, but that is only a small part of the task. You need to spend time

the task. You need to spend time understanding technical complexity and building complex software, which can take months or years. Or you need to deeply understand customers to decide what exactly to build: talk to people, read their facial expressions, conduct

surveys, and repeat this until you understand what to build. Therefore,

there is value in trying different directions, but at the same time diving deep into a few areas. For building a business, this is still important. One

last question. I really want to ask about artificial general intelligence, because people often use that term. I

think it was Jensen Huang. He says we have already reached artificial general intelligence, but you say it will take us a decade before we arrive at it.

What is the criterion that would allow you to say, "Now we have finally achieved this AI"?

People cite different timelines for achieving AGI because they use different definitions of AGI. The

definition that makes the most sense to me is the ability to perform any intellectual task that a human can perform. For example, a human brain

perform. For example, a human brain might need five years to learn and write a doctoral dissertation. But

could an AI write a doctoral dissertation, or can a human learn to drive a truck through a dense rainforest in just a few dozen minutes of practice? And would an AI be able to

of practice? And would an AI be able to master a new environment in just a few dozen minutes of practice? It turns out there is a long list of similar things that AI cannot yet do and, in my opinion, will not be able to do for a very long time. I hope it is only a matter of decades; it may turn out that

much more time is needed. Therefore,

with this definition, we are still a long way from AGI. OpenAI and Microsoft , as far as I understand, had an agreement. According to it, OpenAI had

agreement. According to it, OpenAI had an economic incentive to announce the achievement of AGI sooner. It turns out that if you use other definitions and lower the bar enough, you could easily say that AGI has already been reached

or that it was achieved 30 years ago.

Everything depends on the definition of AGI.

Yes, exactly. Andrew, thank you so much for this positive and very practical conversation. I like it when you look

conversation. I like it when you look at something, compare your experience with what others are doing with artificial intelligence, analyze your process, and push its boundaries. Thank

you for sharing how your team works, and thank you for your ideas and contributions.

Yes, considering the enormous benefits that AI will continue to bring, I hope that everyone watching this will want to truly study AI, apply it, and even create something of their own.