I needed to create a week of content in a course I am writing for our on-line Masters of Applied Business Analytics. The topics is something that I am not very strong on: optimisation combined with non-linear generalised linear modelling. There are plenty of interesting issues. The algorithms for fitting non-linear models and how they are affected by non-orthogonal parameters and conditions of identifiability.

I got Claude to generate some time series data for me, relating demand to price and advertising budget with a business conditions confounder and a couple of plausible exogenous drivers. It gave me the R-code. I then had difficulty fitting various models and asked it what the problem might be. It was not perfect in its diagnoses but much better than I would have been.

After a few days I have a really good data set and a case/story explaining the background, 75% based on Claude’s writing creativity. At various points of our “conversation” about computaitonal difficulties it suggested key teaching points that I should stress. For instance, centering predictors and avoiding fitting non-linear models when the relationship is linear. For instance, a logistic curve will tend to blow up if the data is linear in x.

I now understand the topic much better than two weeks ago. Claude largely taught me this content and I will now teach it to the students. I wonder how long it will take the market to determine who the middleman is!

Where does this leave people with my skills i.e. academics?

If I am still in this gig in 5 years (unlikely) I suspect content creation will be quite different. I will not only use AI to create the course material, but I will need to use AI to teach it to the students (as it taught me). I imagine that, over and above some readings (perhaps condensed by AI from documents I supply) the core learnings will be based on a list of AI prompts that students give to a (specially trained and mandated) AI. I would road-test all these prompts until the responses were providing the best learning outcomes.

• ·         Explain the key issues in this case.
• ·         What are the main pitfalls in a naïve analysis?
• ·         What are the main modelling options?
• ·         I fitted this model and got this output. Here is how I interpret it. What do you think?
• ·         I tried this model and got this error. What is going on?
• ·         Give me a high level overview of how optim works.

etc. You get the idea.