Demystifying AI:
From Deep Learning to Workforce Transformation

Michael Krigsman: Welcome to CXOTalk. Episode number 837. Today we're exploring the technology behind AI and the implications for workforce transformation. Our guest is Kian Katanforoosh, and he is co-creator of Stanford's Deep Learning course and CEO of a company called Workera.

Kian Katanforoosh: In 2017, I co-created the Deep Learning Class at Stanford with Andrew Ng, one of the pioneers of artificial intelligence. And together we've taught AI and deep learning to over 4 million students, both on campus but also online through an initiative we call deeplearning.ai. The goal at the time was to democratize access to AI education.

And as all these students around the world were taking our classes and watching our videos, some of them pointed out that, with the emergence of content, they're drowning in a notion where any skill can be learned in 2000s of different ways. But one of the gaps that they pointed out they had was understanding their skills. "What am I good at? What am I lacking? What benchmarks do I need? Am I better than the average big tech engineer or the average product manager out there?" And so in 2019, as we were getting this feedback, we created another company called Workera, which is a skills technology company that leverages AI to help organizations and individuals verify their skills, future-proof themselves.

And it turns out that when you are more self-aware about your skills, you can make the right decisions. You may not need to do a 200-hour class when you can take a ten-hour class and five articles that directly focus on your job.

Michael Krigsman: When we hear terms like deep learning and like machine learning, what are we actually referring to?

Kian Katanforoosh: The way you can think of AI is a large field of study focused on making machines capable of performing certain tasks that would typically be performed by humans. But AI can encompass anything from the most simplistic algorithms to the most advanced neural networks, if you will. Machine learning, on the other hand, is the subset of AI that focuses on developing algorithms that are learning. As the name suggests, they can perform tasks without being explicitly programmed for that task. And deep learning is even a smaller subset of machine learning, which focuses on algorithms that are broadly mimicking neurons as we know them in the human brain. The analogy is not super accurate, but you can think of it as neurons that ingest information and that end up learning certain things.

And so if I give you an example to understand the three levels, when you're going to Google Maps and you're trying to go from point A to point B, AI in the background is optimizing the route from point A to point B. Now imagine you want to add traffic in real-time and take it into consideration. Now your algorithm needs to adapt, and the route may change on a day-to-day basis. And now we're in the realm of machine learning. If you now want to have an autonomous car that understands highly unstructured data, like images, like video streams, we call that perception algorithms or computer vision. Now you're talking deep learning. It's the family of algorithms that is the best suited for that type of problem.

Michael Krigsman : Can you shed some light on how these technologies work? Because again, we know the terminology.

Kian Katanforoosh: Imagine that you are trying to build a model that is meant to detect whether there is a cat or no cats on a given picture. Well, if I ask you to go on a whiteboard and write the formula that is meant to detect cats, will you be able to do it or not? There is no such formula. We don't understand the mathematical formula.

So instead of writing a formula, I am going to program a function that is highly complex. The function is highly adaptable. It has parameters that can be fitted on data. And so instead of writing a formula, I am going to select a large enough data set of cat pictures and non-cat pictures. And I'm going to label them cats or non-cats. And I'm going to iteratively give maybe a picture at a time to the model and let the model predict whether there's a cat or no cats. And if the model is right, I'm going to give it a thumbs-up and reinforce the fact that the function it is fitting is probably correct. If the model is wrong, instead, I'm going to ask the model to modify its parameters—to modify the function it is fitting.

And it turns out that if you do that thousands of times, sometimes millions of times, you end up having learned the function to detect cats, and the model can then use that function in real-time on a new data set that it has never seen, and it will be able to detect a cat. And that's called the learning process.

Michael Krigsman: Help folks listening to have a better understanding of under the covers. When that cat recognition is taking place, how does it actually happen?

Kian Katanforoosh: One of the things to know about machine learning is the differences between the training process, which I just described, and the prediction or the inference process—that is, you using that cat detector for a specific purpose on a use case that can be an industrial use case or a business use case. So once you understood the training process that I described, you now have a model that has parameters that have been fitted on a data set, and the model is ready to be used.

You can now put a camera in your garden and upload, if you will, that model either on cloud infrastructure if the camera is connected to the internet, or directly by putting a file of parameters and an architecture on the camera. And whenever a video stream or an image is captured by the camera, that image, which is essentially a set of pixels that can be described in computer science with numbers, is fed into the model. The model is going to process it. Remember the function I was talking about earlier? The image is fed into the function. The function will compute—or we calculate the function—the model will calculate the function we fit. In the end, you will receive a probability between 0 and 1, indicating whether the model thinks there's a cat or no cats in the image. And this is called the inference process. The prediction process from input to output.

Michael Krigsman: So what are the implications of this for business people in terms of the applications? And how does this then relate to generative AI and the type of applications that are suitable for AI and where problems arise, such as hallucinations, for example?

Kian Katanforoosh: As a business leader, you're probably facing many use cases, and not everything will be solvable by AI. But the reason AI will matter is by understanding how artificial intelligence works, or AI works, you will understand how to develop a competitive advantage, maybe for your business or your organization. You will be better positioned to determine what strategic investments to make.

To go back to the example I was giving you with Google Maps, the first version I was describing that takes us on a map from point A to point B does not require meaningful data to be fed. It is... if you were to solve that problem, you would not have to think about the data that you collect. However, if you're building an autonomous car, you will need to have tons of data about roads and traffic signals and pedestrians. And so understanding the capabilities of AI, the learning process, will allow you, as a business leader, to determine what type of data you need to collect and whether that data can translate into a competitive mode for your organization.

The other thing that I find helpful as a business practitioner is you probably will be asked to maybe hire an AI team or a data team, and without the knowledge of AI, you will struggle to determine who to hire or who to put on what projects. And so understanding the different specialties matters. And finally, we're in a world where different AI technologies might be subject to different regulatory concerns—privacy, transparency, explainability—and understanding those topics will really help you determine what to do.

Michael Krigsman: As business leaders are looking at their workforce and the implications of AI, what are some of the key issues that they really need to be thinking about today?

Kian Katanforoosh: There is a metric shared on a regular basis by the World Economic Forum, called The Half-Life of Skills. That explains how long a skill might be helpful in someone's career. 40 years ago, the half-life of skill was above ten years, meaning that you do not need, as a worker, to refresh your skills too frequently. Most recently, the World Economic Forum has shared that the half-life of skill is around four years. In digital fields like AI, it's probably even closer to two years. We are seeing the half-life of skills compress, which means that we constantly need to reinvent ourselves, to learn new skills if we want to innovate. And so AI has pushed us to become skills-based organizations.

And when skills are moving so fast and you need to refresh your competencies so fast, being self-aware of your strengths and your gaps becomes increasingly important because we don't have all the time in the world. We need to determine what classes we're going to take, what subject matter areas we're going to focus the next ten hours of our learning on. And without knowing our capabilities, we are not able to make the right decisions, which is where Workera comes in to serve organizations to accelerate their learning.

Michael Krigsman: The issue, then, is one of having a clear understanding of the skills that are present inside the organization and where the trajectory is going. Was that an accurate way of describing it?

Kian Katanforoosh: Yes. So I would say there are two steps. The first one is, "Where are we at?" Let's benchmark ourselves with an adaptive test aligned with the business objectives of the organization because every business is slightly different. So the type of skills you will need is also slightly different. So the first question is, "Where are we at?" The second question, as you measure over time, is, "How fast are we moving?" And this is a metric that we call learning velocity, that we believe is a new competitive advantage for businesses. Making sure that you have a high learning velocity is essentially bulletproofing yourself against disruption, to make sure that whenever there is a new innovation, a new language model, a new method that comes out in the market, you have a team that is able to learn it and apply it to your domain and produce a meaningful innovation.

Michael Krigsman: I guess one of the challenges that business leaders face right now is understanding that trajectory because AI is moving so quickly. And so how should—how can business leaders deal with that issue of needing to assess and needing to understand where things are going and not knowing? Really.

Kian Katanforoosh: As a business leader, you are uniquely positioned to understand the capabilities you need in order to solve the projects that you have ahead of you, the OKRs that you're trying to meet, the sales that you're trying to make. And that is what's unique about business leaders. Over the last ten years, business leaders have been asked to be doing all sorts of things that they're not meant to know. You know, most of us business leaders were promoted because we had subject matter expertise in an area. We were not promoted, most of the time, because we were a great talent manager, because we had a PhD in psychometrics—the ability to measure skills in our team—because we were learning scientists and understood what the best content is out there for our team. That is not why we—you're promoted into that role. And so your job is not to go and screen the internet and find what is the best class for employee A or employee B. Your job is, "Can I translate meaningfully my company's capabilities and OKRs and business objectives into a series of skills that I can measure my team on?" That's the question you should be asking yourself.

Michael Krigsman: We have another question from Twitter, and this is from Lisbeth Shaw, who says how should organizations upskill their workforce in advance before people are displaced by AI, anticipating potential displacements in the future and hopefully trying to mitigate or prevent them?

Kian Katanforoosh: Over the last ten years, upskilling was content-led. So we have platforms that provide courses and podcasts and videos and books, and there's a ton of content out there. Managers will say, I have selected a course that is 200 hours and post that course. There is a quiz that may validate someone has learned the video, and then after that, I'm going to find the projects for people. What happens is engagement is low because users employees report that the course was just not relevant to their career. It may have been too easy, it may have been too hard, and thus they haven't completed. But even for those who have completed, the manager cannot verify their skills because the quiz was aligned with the content. It was just verifying whether they watched the video or not. And the project is misaligned with the generated class that they were invited to do. 

And so we run into a ton of issues from an upskilling time wasted, nothing to do after the course, nothing to apply your skills into. What's happening right now is that that approach, content-led, is being flipped, where content is actually the last step of the upskilling program, the way it's done in the foremost thinking organizations is you start with the projects. If you have projects, if you don't have identified projects, you start with the job architecture. How have you defined your job roles, your job families, your job levels? What are the skills involved in those? Can you incorporate your OKRs? That is the outcome or the outcome is then translated into a skill ontology granular set of skills that are representative of what you want people to develop that skill. Ontology is being verified with the verification system that is not meant to verify if someone has watched a video, it is meant to verify someone has the skill on the job that you have defined in your ontology, and then only when you measure people on that ontology, you can talk about content and determine that Alice needs a seven-hour class, but Bob needs a 200-hour degree program. But Eve needs just to watch a video and read two research papers. And that is what's going to help us drive upskilling that is business-driven, not content-driven. And that is the big trend that we're seeing in the market at the moment.

Michael Krigsman: Kian, you've mentioned several times the importance of understanding the problem and understanding the data in order to successfully apply AI. Can you give us some examples of use cases that you've seen where understanding the problem and the data has been particularly important?

Kian Katanforoosh: Use cases are going to fall under one of three categories. The products and solutions innovation, the productivity use cases, and the security and risk mitigation use cases. And I've seen plenty of these three categories, on the products and solutions innovation. I called out the use case about predictive maintenance on drilling. To give you another example. You know, I've seen medical students, medical practitioners work on x-ray images in order to determine if a patient has pneumonia or not. Again, here, the problem can be resolved to a certain extent with computer vision and AI. But who is going to solve that problem? The medical experts, because they understand the metrics that will be used to evaluate the model, to understand the deployment environment. They understand the clinical setting. They can show you in the clinical room, in the Stanford Hospital, which machine the model is going to be running on. 

All these things are uniquely requiring medical expertise. Another example that would come to mind is a group of people building a camera that you ingest your mouse and that is meant to take pictures of your gastrointestinal journey and then will spit out certain predictions and certain pathologies that you may or or you may have. And those are again, projects that require subject matter expertise and AI. There are plenty of these projects on the risk and security mitigation. Companies are worried that they will face data privacy issues, that some of the employees will leak data. And so what they're doing is get ahead of it, employees are going to use it regardless of whether you ban it or not. We know that already, and we've seen it in multiple organizations. So what can you do? You can put together certain skills verification, certain certifications in order to make sure that people are aware of the risks of using artificial intelligence or not. So a lot of the use cases are risk mitigation, and and increasing security on that front.

Michael Krigsman: Here's a question from again from Arsalan Khan, who says I might help to improve a process, but perhaps we still need holistic thinkers who can tell that the process itself is just wrong. Who is going to do this? The AI, the human, the business people, the technical people? In other words, who's got the overall view that we're actually headed in the right direction?

Kian Katanforoosh: Absolutely. It is not the AI, it is the subject matter experts. That is why I think subject matter expertise is going to increase in importance in the coming years. You are working on any AI system. Who is going to stamp that AI system and say it works? The subject matter experts, your AI deployments will be as good as the expertise that you have in the domain. A good company working on a great AI tool will have the expertise to say, this is how AI will be incorporated into the process. This is how the process needs to change. This is how we need to think about the business outcome. And so I don't think that subject matter expertise is going anywhere. If anything, I believe we will need more PhDs. We will need more domain experts. We will need people who know what they're doing in the subject matter that they're focused on.

Michael Krigsman: Which raises another very interesting and important point, which is what should business professionals do to retain their relevance as the great source of knowledge in the sky increases, namely the large language models that know everything.

Kian Katanforoosh: There is tons of innovation ahead of us that we have not seen come to the world yet. It is just going to accelerate. If you can keep up with the pace, you will have a great career. You will contribute to the next generation of technologies that will be produced. You will solve certain use cases, but it is very important to have that mindset of lifelong learning. I think it's more important than ever. The other thing is the self-awareness. As more and more information comes to our world, we have to face critical decisions on how to spend our time. The best of us, the most self-aware, are going to be the ones who are going to spend their time on what matters the most, what matters to their company, what matters to their career, what matters to their project. And so make sure that you have created the right learning infrastructure around yourself to be self-aware. And that may come from different people or systems. Maybe you have a mentor or a coach that keeps you aware and tells you when you're doing something wrong or when you have a gap. Maybe you're using an assessment tool like we're out to take tests on a regular basis. I do that, and I took a test this weekend. I take it every week. And it helps me keep myself self-aware as much as possible. And, so that's, that's the point I'm making self- awareness and learning velocity, learning how to learn. These are going to be your your career defense ability and your competitive advantage.

Michael Krigsman: Self-awareness and learning defense ability. Learning how to learn. These seem not so easy skills to diffuse through an organization.

Kian Katanforoosh: It is a challenge and I'll give you one thing that I've observed. The leaders we work with are great leaders. But generally, leaders would fall under two categories when it comes to upskilling. I have seen leaders, business leaders pretend that they know AI. The problem when you pretend and you don't have self-awareness is the workforce tends to mimic the leader. So when someone pretends, everybody underneath is pretending, and that is when you end up with a workforce of dangerous amateurs that will take you on a journey to prototype AI. But when it's time to productize AI, you will fall short. On the other hand, you have leaders that are approaching upskilling with humility. The leaders take the work their own assessments and they tell their team, listen, I'm not yet at the target score that I set for myself, but I'm not too far and I'm going to work on it and we're all in this together. The consequence of that is the workforce will mimic that behavior. The workforce will approach upskilling with humility. They will test themselves. They will take action on their gaps and they will try to reach their target score the same way their leader is trying to reach their target scores. And so I invite business practitioners to take the approach of humility rather than the approach of pretension.

Michael Krigsman: Any final thoughts on the time frame for disruption that's likely to happen when it comes to workforces, and just final advice on what businesses can do. And I'll also say to the audience, now is your last chance to get final questions in.

Kian Katanforoosh: We're very focused on AI and there is an AI hype cycle. And it is fine. As a business practitioner, don't shy away from learning. I just embrace the hype. The hype will die out. But you would have learned a ton of things while the hype was going. So embrace the hype. However, beware. That's us, we're looking at AI. I believe the real frontier that is ahead of us is skills- based organization—it is a skills-based workforce. Because what's happening with AI is our world is accelerating. Technologies are coming out more frequently than ever. A few weeks ago, four major language models were launched in the same day from OpenAI and multiple competitors at the same time. Think about how fast that is. So the signal is the future moves fast. Learning velocity will be the new competitive advantage, and organizations are going to understand their skills, take actions based on their skills, and become skills-based organizations— or those who will be bullied, proof to any hype that will come in the future and will continue to develop themselves. And so that is my message to conclude on this question.

Michael Krigsman: And with that, we are out of time. A huge thank you to Kian Katanforoosh. Thank you so much, Kian, for being with us and sharing your expertise today.

Kian Katanforoosh: Thank you, Michael, for having me. It was super fun.

Michael Krigsman: And thank you to everybody who watched and especially to those folks who asked such great questions. You guys are an amazing audience. Before you go, subscribe to our newsletter and subscribe to our YouTube channel. We do have incredible shows coming up. I always say that, you know, and it's always true. We have great guests. So everybody, I hope you have a great day.

Subscribe. Stay in touch, and we'll see you soon. Have a great one, everybody. Bye-bye.