Frankly today, I would not hire a software engineer that does not know how to use AI tools in a very sophisticated way. This hype that we'll get to AGI, artificial general intelligence, in two quarters or a year or whatever, that's just hype. I seriously doubt that's going to happen. There's no technology that I know of today that by itself is a path to AGI. AGI has been defined and redefined so many times it has almost lost meaning. When I started the Google Brain project, the number one mission I set for the team was, 'Let's do really large neural networks and throw all the data at it.' The way the AI learns is different than the way humans learn. I actually was starting a little bit before when I started Google Brain when deep learning was coming up. I was reading stacks of papers on neuroscience to try to figure out how the human brain works to see if that could inspire better AI algorithms, and those exercises turned out mostly not to be useful. And so many nations are thinking about sovereign AI. One way to do that that's very efficient is investing in open source and open weight models. In the open source open weight model, one huge force over the last year has been the open source open weight models coming out of China.
But increasingly, the United States appears to be less welcoming of immigrants, whether they are in technology or in any other field. How do you look at this?
I think it's awful. I think it's a huge unforced error on the part of the United States. I think the question is, now that AI has shaken everything up, can India upskill rapidly enough so that very talented, very smart, very hardworking innovators and businesses in India can not just keep up but potentially even leapfrog and invent the next thing.
Andrew, thanks so much for joining us today here in Davos.
Thank you. I just got off a plane a few hours ago and so my first meeting here is with ET.
Wonderful. I'm going to start by asking you about what AI means for India and specifically for one of India's biggest industries, which is the IT services industry. A big concern has been that rapid advancements in AI has meant that a lot of software coding and a lot of software development, even mid-level software development, is today done by machines themselves. Tools like Claude Code are now being cited by software engineers as getting really, really advanced. What does this mean? What could it mean for a very large services industry like India's IT industry?
I think the worries about secular disruption are well placed, and the risk and the opportunity is that with AI tools, individuals that are not skilled with AI are increasingly dramatically less productive than people that are. And so, frankly today, I would not hire a software engineer that does not know how to use AI tools in a very sophisticated way. What is not well understood but is spreading is for many other job roles, from accountants to marketers, HR professionals to IT backend management, for more and more job roles, companies are demanding sophisticated use of AI skills, maybe even to the point of coding. So frankly today, when I hire a marketer or HR professional, I would strongly prefer to hire someone that knows how to use AI to the level of being able to write software, build software to help be a better marketing or HR professional. So for India, with a very large professional services and outsourcing industry, the risk and the opportunity is if this group of people, very smart, hardworking people that the world has trusted for a long time to do important work, could upskill rapidly, then India will continue to provide tremendous value. But if the upskilling doesn't take place rapidly enough, then I think the job disruption, job displacement could be significant. So I feel a lot of pressure for India and a handful of other nations around the world to see how we can figure out the training to make sure we get this one right.
I want to talk about the recent evolution of the LLM-based AI models and chatbots in general. I think the world woke up in a very big way during what we call the ChatGPT moment of 2022 to the possibilities of this generation of GenAI tools. Since then, when you look at the evolution of these models and what they're capable of, have you been really impressed or have you been underwhelmed?
I think it's been great. Since the first version of ChatGPT, the models have continued to improve and I feel like the hype has gone like that incredibly fast. The business reality has gone like that also very fast, but not as fast as the hype, as things often are. Yes, more work lies ahead of us to build AI applications, to build AI agents and agentic workflows on top of this technology that I think is widely appreciated. This hype that we'll get AGI, artificial general intelligence, in two quarters or a year or whatever, that's just hype. I seriously doubt that's going to happen for any reasonable definition of AGI. But at the same time, without having computers that are as intelligent as humans or as capable as humans in the general sense, there's so much work we do to help with business process automation, workflow re-engineering. I think there'll be many big industrial or industry transformations coming in many different sectors, and we'll still be working on this like a decade from now.
You know, Yann LeCun, who's like you another of the biggest names in the field of AI, he has recently said that he does not believe that the LLM-based, I mean the current wave of models, are the right way to even try and approach AGI and a totally different wave is required. Where do you stand on that debate?
So I think he's right. At the same time, the most reasonable definition of AGI that I know of is AI that could do any intellectual task that a human can. So if we have AGI, it means AI should be able to spend five years to write a PhD thesis or learn to drive a truck through a jungle in maybe minutes, tens of minutes, because a human can. LLMs by themselves are not a path to AGI, but there's no technology that I know of today that by itself is a path to AGI. To achieve that type of AGI will need new technical breakthroughs that none of us really know what exactly it is. So you can plug in any technology and say that technology that exists today by itself is not a path to AGI. That will be a true statement. But one interesting thing is because of the hype around AGI, I've seen a lot of businesses attempt to redefine AGI at a lower bar to make it easier to hit. And this is problematic because most of society thinks AGI means AI that is as smart as humans generally. To me, when we have AGI, it means AI should be able to do everything that a remote worker would be able to do. We're very far away from that. And when companies redefine AGI to lower the bar and then say with some narrow technical finicky definition maybe they've achieved AGI, it misleads the public, misleads business leaders into thinking AI is on a more powerful trajectory than it actually is. I've seen this mislead high school students that are now making different career decisions because they think AGI will obsolete certain careers, and I'm seeing companies make very large capital allocation decisions on the basis that AI will be more intelligent in a couple years than I think it actually will be. So, AGI has been defined and redefined so many times it has almost lost meaning. But having said that, even without reaching this really high bar of AGI, AI is incredibly exciting and is incredibly valuable, and India and the US and other nations just building to that will be exciting.
I'm sure CEOs ask you all the time wherever they can get hold of you about, 'Hey Andrew, tell us what is happening, how should we...' What's the most common question that you get from CEOs?
Actually, my business partner Kian Katanforoosh and I do a lot of work through DeepLearning.AI advising large businesses, and as we look to the industry transformation that's happening through many sectors, the biggest thing we end up speaking with CEOs about is how to help the businesses adopt AI. And I would say the number one thing is the people change: how do we upskill people, take them with us, how do we really do the people change management that helps the people they're ultimately going to do the work to adopt AI. And of that group, I would say the biggest category is a lot of businesses have done the 'let a thousand flowers bloom' strategy and it's mostly not working. And there's a specific reason for that. Bottom-up innovation is important, let's keep doing it, but I find a lot of bottom-up innovation ends up with point solutions where someone sees businesses do many things and you take a small piece of it, make it 10 times more efficient. That's nice, but the overall gains to the business are incremental. So we see a lot of these 5%, 10% efficiency improvements which are nice, but it's not what AI is promising to do. In contrast, it often needs top-down innovation or bottom-up innovation with a view of the bigger picture to understand what are all the steps the business executes to create value to then re-engineer this broader workflow. I'm seeing that deliver much more value. So bottom-up innovation is great, but at some point it takes an individual, maybe a top-down leader or bottom-up person with that broader view, to think through the workflow redesign. And I'm seeing that where the business gets it right, it delivers much more value. Leadership comes from the top. There's a lot to be said for bottom-up influence, let's do lots of that too, but leadership comes from the top. So one challenge I have to every C-suite person is if you're serious about AI, learn a bit about it yourself. Make sure that you personally, in addition to excitement about AI, learn enough, spend a few hours going a little bit deeper in the technology to make sure that you have that knowledge to drive your teams to allocate assets and prioritize and then drive the business results.
I want to, while I have you, I want to test my own understanding of how bots work and about the current wave of generative technologies. Fundamentally, the large language models work by predicting the next word or the next string of sentences, right? By all the training they've had, they look at what's... how is it that models are able to sound so intelligent if all they're really doing is actually predicting the next word, or is that an oversimplified understanding of what they do?
Oh, so loosely the way that large language models are trained, step one, called pre-training, is learn to predict text on the internet. And so after that, because the internet has a lot of random stuff, it tends to regurgitate internet-like things. But there's another step after that where we take the model and with just a small amount of data align it to say more sensible things. So for example, they do their best to factually answer questions and give helpful answers, whereas sometimes the internet, social media, doesn't give the most helpful answers. So you ask a question, you get back a bunch of snarky funny comments or something. But during the second step, which can be implemented using a technique called reinforcement learning or other techniques, you train the algorithm to be helpful, honest, and harmless. And so it's that second step. So the first step, pre-training, gives a lot of knowledge on the internet, and then that second step causes it to use all that pre-trained knowledge to be helpful. But with such fundamental steps, these models are able to sound so smart, able to give you strategies about complex problems and solve them. It just seems a bit counterintuitive, but it just works. It just seems to work.
