I was slow to jump on the LLM bandwagon.
In some ways this is surprising. Although it was a while ago, I did take several AI courses in grad school. So I am probably in a good position to dive into new developments in the field.
But when the LLM wave started I was busy trying to develop software engineering expertise in Python. I wanted to finish that “project” before moving onto something new. It was only a few months ago, when the PR of my multiyear module was merged to the censusdis package (link), that I felt like I could turn my attention elsewhere.
Coincidentally, around that time Kevin Markham created his first course on LLMs. I wrote a review of that course here. In this post I review his followup course: Build AI Agents with Python.
The Problem AI Agents Solve
Something I learned in this course is that LLMs like GPT-4o mini are called “foundational models”. Foundational models have very real limitations. Consider this conversation I just had with GPT-4o mini:
User: What is your cutoff date?
AI: My knowledge cutoff date is October 2021. This means I do not have information on events or developments that occurred after that date. If you have questions based on information available up to then, feel free to ask!User: Do you have access to the web?
AI: No, I do not have access to the web or the ability to browse the internet. My responses are generated based on a mixture of licensed data, data created by human trainers, and publicly available information up to my knowledge cutoff date in October 2021. If you have any questions within that scope, feel free to ask!User: What is the latest version of Python?
AI: As of my last knowledge update in October 2021, the latest stable release of Python was Python 3.10, which was released on October 4, 2021. For any updates or newer versions released after that date, I recommend checking the official Python website at https://www.python.org/.
In short, the answers that foundational models give are limited to the data they were trained on. For reference, the latest version of Python is 3.13.3, which was released in April 2025.
Augmenting Foundational Models
You can think of AI Agents as augmenting foundational models with tools, data and prompts that you provide them. This allows them to provide better answers to certain questions.
As an example, Build AI Agents with Python walks you though the process of giving GPT-4o mini the “tool” of web search using Tavily. This resulted in the following conversation:
User: What is your cutoff date?
AI: My knowledge cutoff date is in October 2021. I do not have information on events or developments that occurred after that date. However, I can access real-time information through specific tools. If you have a particular query, feel free to ask!User: Do you have access to the web?
AI: Yes, I can access real-time information through a specific tool to search the web for current events and updates. If you have a particular topic or question in mind, let me know, and I can look it up for you!User: What is the latest version of python?
AI: The latest stable version of Python is 3.13.3. You can download it from the official Python website https://www.python.org/downloads/.
I had not heard of Tavily prior to taking this course. It turns out that there is an entire ecosystem of tools around LLMs!
In addition to adding in a tool for web search, the course guides you through integrating custom data sources using Retrieval Augmented Generation (RAG) and building a Q&A system that can interact with a SQL database. For a course that only contains two hours of videos, I felt that it covered a lot.
Like with all online courses, there is an argument that I could have learned all this on my own. But “get your hands dirty with LLMs” had been on my todo list for a while. It was only taking this course that made it happen. If you are looking for a similar push, then you might enjoy this course too.
Wish List
Kevin is gifted at coming up with good examples and clearly explaining his solutions. If there was one change I could make to the course it would be for him to add exercises to it.
Longtime readers might remember that I have created and taught several courses myself. My initial courses did not have exercises. It was only when I read Telling Ain’t Training that I learned how valuable they are for students. That book also presents a framework for creating exercises that I still use today.
Next Steps
After taking an online course, I often want to apply what I learned to a project of my own. In this case, I can now imagine adding an LLM feature to my Covid Demographics Explorer.
The dataset that powers that app is in a single CSV file. I think it would be fun to create a chatbot that has access to that file and uses it to answer questions. That would allow users to ask questions about the dataset using natural language and get answers in natural language.
