4.0
Reviewed Dec 28, 2024This is very good course for anyone planning to use Gen AI for data analysis tasks.
I'm sorry to give such a low rating to the course, but I couldn't get myself to finish it, as it is just one giant ad for generative AI to fuel the AI hype train, without teaching anything substantial, and repeating the same points over and over. While this introduction course does go over some very basic concepts in data science (e.g., what are data inconsistencies and some common "traditional" methods to deal with them), it does so without definitions. I expected it to at least dive into how these big AI models work then, but it doesn't do that, either. Here are some highlights of why I believe this course should be seriously reworked: - It goes over the same concepts numerous times: once in a video, then in a very very lengthy reading of the very same thing. - The course material is seriously biased towards setting up generative artificial intelligence as the ultimate solution to our problems, with one or two times (after several videos) warning the student that human validation is needed. Other times, each video is usually all about how genAI will solve this and solve that, and it can "identify nuances that humans can't", but nothing concrete. - The course gives us examples, like: imagine we have medical records and want to diagnose a rare disease, but we only have a hundred patients' medical records. Generative AI can "understand the nuances" of the data and generate much more new records that helps us analyze the data and train models for it. Seriously? One of the most important things taught on statistics courses is that extrapolation is very dangerous. The video never mentions that this can introduce bias or completely derail research. What if the devised treatment would've worked on the LLM's patients, but made matters worse for the real ones? These cases are barely covered in some of the reading material, but even those did not come back to the medical data example. - Videos that demonstrate the AI (Microsoft Copilot) in action actually show that it's not doing what it was asked to do: in one video, they ask it to generate a box plot, and Copilot (on the second try) generates a bar chart. But the video keeps going as if it was a box plot. In the same video, they say "the AI has identified several missing data points", where on the video, there is clearly one missing data point in the data and one that it identified. In another video, the voice over says that the AI (Copilot) generated a chart, yet no plot appeared throughout the demonstration - we could see pivot tables in the Copilot chat window, but no generated plots. These demonstrations videos are a pain to watch and I'm surprised that nobody has validated that the recorded footage and the voice over are actually of the same thing. - Assignments that ask you to work with datasets and "ask a genAI model" do not point you anywhere: you'll have to find ChatGPT (or Copilot) or Claude yourself and input the data, and pray that it outputs what the expected answer is for the question - or just do it yourself, without the AI (spoiler: neither of them gave a good answer, I did the analysis myself to get the good results). Sure, it's a great way to learn it yourself, but it shouldn't be a practical assignment then. - Other assignments sometimes ask you about things that were not or barely mentioned in the course (like which type of machine learning model (i.e., RNN, GAN, etc.) would you use for this type of data analysis task), or are just the same "advertisement" type of questions where you always pick the option that says genAI is the ultimate solution that uses advanced machine learning algorithms to uncover hidden patterns in your data. All-in-all, I was hoping that this course would actually go a bit deeper into understanding these new AI models instead of just treating them as magical black boxes that solve all our data analysis problems. I was also expecting it to give me a balanced view of the pros and cons, along with practical knowledge of incorporating AI into a data analysis workflow instead of a Copilot ad gone wrong. I would recommend students to pick a course that teaches the actual inner workings of LLMs (with definitions and practical guides), provides a less biased view on the topic, and then find another course to learn how to use these LLMs with data. Once again, sorry for being this critical, but I believe this course really only teaches one to believe that genAI will do so many things we can't. Which may be true, but you don't teach that through a course, but instead you form that opinion yourself through experience.