How do I learn practical AI skills?
In my previous article on this series about Getting Started with AI, we discussed how to learn the basics of AI.
In this article, we’ll learn about the five main practical skills you need to know to work with AI.
Each new skill level is progressively more difficult than the previous skill level.
So, you should start with the easiest of these skills and work your way as far as you need to reach your required level of proficiency.
At this skill level, you are using off-the-shelf AI tools built by a third-party application developer. These tools help you solve day-to-day problems with the help of AI. For example, using Github Copilot to help you write code, using ChatGPT to help you write articles, or using DALL-E 2 to help you create digital art. These AI-assisted tools can make you vastly more productive than your peers.
To become proficient at this skill level, you just need basic AI literacy and training on each tool as necessary. Some of these tools are stand-alone applications and websites, for example, Stable Diffusion and Midjourney. Others are part of an AI platform, for example, Open AI’s ChatGPT and DALL-E 2. Others are integrated directly into other applications, like Github Copilot and Adobe Firefly.
At this skill level, you’re using an AI system or ML model that someone else has hosted as a service to solve a common problem. Or, you’re wiring together a bunch of services to solve a more complex problem. For example, using a face-detection service from Google to identify faces in photos. Or, using Microsoft’s Form Recognizer service to extract data from standardized forms and documents.
To be proficient at this skill level, you need to know how to write computer programs with a programming language like Python, C#, or Javascript. You also need to know how to make REST API calls. Then learn about the various AI services from Microsoft, Google, and Amazon. Also, be sure to check out OpenAI for powerful generative models and Huggingface for a wide variety of open-source models.
At this skill level, you’re using general-purpose foundational models created by a 3rd party like Google, OpenAI, etc. However, you’re customizing it — using transfer learning — to solve your specific problem. For example, customizing a 3rd-party image classifier to detect your company’s logo. Or, customizing GPT-4 with Retrieval Augmented Generation (RAG) and few-shot learning to write product descriptions for a website.
To be proficient at this skill level, you need all of the skills above. However, you also need general training on ML and specific training on transfer learning and prompt engineering. Most AI service providers now have easy-to-use tools for customizing their pre-trained models. However, you may also need to learn how to create system prompts for Large Language Models (LLMs) and in-context learning.
At this skill level, you are creating your own models by creating data, labeling data, training a model, tuning the model, verifying its performance, deploying it, and maintaining it. For example, using Python and Pytorch with your customer database to predict customer churn. Or, using Python and Tensorflow with a database of product images to train a deep neural network to recognize products on shelves.
To be proficient at this skill level, you need in-depth knowledge of applied machine learning and data science. However, it is becoming much easier for software developers and IT professionals to create new models using the latest tools that automate much of the heavy lifting like AutoML. If you’re considering this path, you may want to consider a 4-year degree in an AI-related field from an accredited university.
The most advanced skill level involves creating new ML training algorithms and implementing new AI algorithms from scratch. Creating new AI training algorithms requires in-depth knowledge of machine learning, data science, statistics, linear algebra, and calculus. You will almost certainly need, at minimum, a 4-year degree from an accredited university to feel confident creating your own ML training algorithms.
Beyond machine learning, there are a bunch of “algorithms with tricks” that we use in AI to solve complex problems. In computer science lingo, these are typically NP-Hard problems solved by state-space search with heuristics. For example, symbolic logic, expert systems, constrained optimization, genetic programming, swarm intelligence, fuzzy logic, etc.
However, when most people say “AI” these days, they actually mean “ML” — which is modern data-driven AI. If you’re interested in these other (non-ML) branches of AI, the best resource is Artificial Intelligence: A Modern Approach by Norvig and Russell. This book is the most popular textbook for introductory courses on AI. I’ve used it in at least five college courses that I’ve taken. It’s a bit technical but still accessible.
To learn more, please check out my next article in this series: Practicing AI Skills