My blog

ChatGPT brought AI closer to us and allowed us to see its tremendous power. It opens up many new opportunities - I believe that the real impact of ChatGPT is yet to be seen.

Traditional AI is good for Classification (recognition). The advent of GPT brought us the ability to generate - from poems to art work, from music to computer program, GPT's (and other LLM's) ability to create will continue to impress us. This will undoubtedly open up many new opportunities for Al applications.

Pre-trained models drastically lower the cost and effort in using Al With "prompts", we can talk to GPT in plain English (or other languages). Every one can now become a "programmer" - you don’t need to know python or a programming language.
ChatGPT drastically lowered the barrier for "programming" a computer.

The lower barriers of entry and the generative power of GPT have accelerated the adoption of Al in almost every industry.

1. It will be as pervasive
2. It will be as transformative
3. We are still in the early innings

Just like electricity, AI will “power” everything from fashion design to pizza making. I believe that significant opportunity lie in the long “tail”. Be in the position to capture these opportunities.

McCarthy & Kotok created the first Chess playing program in the early 60's, deploying

Arthur Samuel developed a Checkers program in 1956 - it used

• Emulating human problem solving (cognitive) process
• USACO/ACSL are about "teaching the computer what you know" how to do
• In the early days of Al research, symbolic approaches have shown more success since it's more efficient to directly "encode" what you know.
• This approach is more understandable (vs. the
  "black box" of ML)

• The symbolic approach is good for solving "well defined” problems, where you can define a set of rules or prescribe a set of solutions
• There are problems that we can not easily write a set of rules nor easily specify the solutions.

Binarized MNIST samples

MNIST 7

• Imitating human's brain - biological
• The foundation of this approach is neural networks and training the neural network to learn the rules
• This approach was not very practical until the last 15 years when we have enough "compute"
• It's more or less a "black box"

An image of a neural network. It works similar to how neurons connect in the brain.

Daniel Kahneman is a psychologist and Nobel Laureate. In his book Thinking, Fast & Slow, he introduces the concepts of System 1 and System 2 - the different ways that human think.

• System 1 is fast, intuitive, and emotional, "effortless". For example, when you pick up the phone, you can instantly tell if it’s your mom calling - you don’t analyze the voice, and after a while you realize - it’s instant.
• System 2 is slower, more deliberative, and more logical, but require effort. For example, when you get a math problem, you have to think about it for a while - the answer doesn’t just pop into your head (oh, how I wish it did).

Prompt engineering with LLM is an integration of the two approaches. From experimentation with various engineering techniques, we have discovered that "system 2" thinking - i.e. deliberative problem solving - can improve the results from LLM in many categories, including:

• Logical reasoning problems
• Math puzzles
• Programming tasks
• Essay writing
• and more!

In prompt engineering, you (the human) play the role of the "teacher" to provide the "system 2" thinking to guide the LLM.

The utilities of the smart phones come from the App Store. There are millions of Apps on the iOS App Store and Android Google Play. I believe that the Power of Al will be harnessed by Al Agents.

Learn to ask questions : Creativity trumps knowledges and experience. Creativity and discovery are sparked by curiosity.

Learn to use technologies : Particularly Al/computer science technologies. How can we leverage computer's power to help us solve problems?

As Al technologies become prevalent, the needs to communicate with Al will continue to increase: a better understanding of "Artificial" Intelligence” will facilitate such communication and will allow us to better leverage Al's power. I think that learning "Al" (what it is and how it works) will become a necessity.

Traditional curriculum in middle/high school and even in college focuses on symbolic approach and foundation of computer science. Introduction to neural network, machine learning are considered
"graduate level" classes. With the advancements in generative Al and LLM, knowledge of neural network and machine learning has become more fundamental, and in the future will become even more fundamental.

CS229: Machine Learning

https://cs229.stanford.edu

CS230 Deep Learning

Deep Learning is one of the most highly sought after skills in AI. In this course, you will learn the foundations of Deep Learning, understand how to build neural networks, and learn how to lead successful machine learning projects. You will learn about Convolutional networks, RNNs, LSTM, Adam, Dropout, BatchNorm, Xavier/He initialization, and more.

https://cs230.stanford.edu

“Run, don’t walk… Remember, either you are running for food, or you are running from becoming food. And often times, you can’t tell which. Either way, run.”
- Jensen Huang, CEO, Nvidia

Feedback? Inquiries? I’m always open to new ideas. Please use the link below or send me an email at arnav.s.krishnan@gmail.com.

📄Feedback & Inquiries