Great interview with one of the well known ML/AI experts in the world. It's about 90 minutes long so if you don't have time to listen, I've summarized the key points (from my view) and the times (in case you just want to go straight to it) below. Hope you enjoy it as much as I did.
2:50: Andrew started learning to code when he was 5-6 years old.
4:50: He was an AI professor, teaching AI at Stanford in a repetitive fashion! The was no better way to practice what he was teaching than automating it which gave birth to Coursera.
6:50. Hard work pays! He usually worked till 1:00am in the early hours of the morning to record the Coursera videos.
8:25: Knowing the foundations (gradient descent) is key to becoming a good ML/AI engineer
9.30: This was not an overnight success. It took years to get there.
10:10: Not every idea works out in AI and that’s okay. Just learn from the experience
12:00: AI developers can be more than 50% of developers worldwide. In fact 100% will have an appreciation of ML/AI – this is huge in my book.
14:00: Coding might develop like how language developed over the years. Just like it is important, almost imperative today to learn a universal language like English, in the future, it will become important, almost imperative to learn a universal programming language to be an effective citizen of the digitized world.
16:00: Being a data scientist will become more important than being a software engineer in the future.
21:00: The Stanford helicopter project was one of the first practical use cases for reinforcement learning.
25:40: It was the wrong intuition to focus on unsupervised learning over super4vised learning. Think of it as trying to run before you can walk. The goal is still to be able to run but walking is a necessary first step.
27:35: If you want to make a breakthrough, there are times you just must go against the norms of the time.
28:40: Both better learning architectures and better data sets are necessary to improve ML/AI.
32:00: Data defects should be thought of in percentage terms not absolute terms.
39:40: Debugging in ML relates to answering questions of overfitting data requirement etc. and very different from software debugging.
43:15: We need so many things to converge together before reinforcement learning can be mainstream.
45:10: A true AI team should be able to use a portfolio of AI tools to do its work and not just a narrow set of tools.
46:50: Self supervised learning may be a critical part to unlocking supervised learning in general.
49:55: It takes about 16 weeks to finish the deep learning specializing from Coursera and deep learning.ai
53:10: Regularity and making learning a bit is the key to be good at anything. Sounds cliché but still an extremely powerful idea.
57:55: Doing projects to learn ML/AI is important for gaining ML/AI skills. Starting small is also important. Taking the first step and then taking smalls steps idea.
70:30: Building companies is one thing but building companies that move the world forward should be every entrepreneurs goal.
72:10: The next wave for AI is to transform other industries outside the software industry.
76:00: With ML/AL in companies, always start small. You need to build trust and that takes time and small steps.
77:40: There is a huge difference with your ML/AL algorithm working well on a test set than working well in a production environment
78:50: Machine learning is automation on steroids – best quote I’ve heard on ML/AI