I wanted to share how I use ChatGPT (and Bard in Google). I find myself using it every day. I hope you find a few ideas here that are helpful to you and share back some of your best prompts.
For me, generative AI is like hiring a great data scientist and a copywriter…plus having a friend who knows everything!
Overall, I use ChatGPT more than Bard as I find it gives better answers, summaries of my writing, etc. ChatGPT also has an ecosystem with 100+ plug-ins built around it. I have activated Noteable (creates a notebook of Python code based on statements in English of what I want to do), and the Wolfram plug-in. ChatGPT knows when to pass control to the right plug-in based on the prompt! For example, I asked it for the law of cosines (provides measures of similarity between vectors) and it went to Wolfram Alpha for that.
However, Bard is also very useful especially for facts that require current knowledge as it is regularly refreshed (e.g. brand advertising budgets). I also use it for Python code for Colab; since that is a Google cloud product, I figure it might give me better answers there.
The biggest ways Generative AI helps me:
Data science: ChatGPT writes code for me in R and Python (and even debugs error messages) which saves me hours. I tell it in English what I want and it writes code including what packages need to be installed…a real magic wand. I can also ask it for explanations of math tools, properties of stat distributions, and even to find derivatives.
Copy editor: I tend to be too technical in my first draft of any document. I feed the document to ChatGPT and ask it to help me improve. I ask it what technical terms need simpler descriptions…it calls them out and even provides those descriptions. It even created a glossary for me for a technical paper that was 85% there.
Faster fact finding than via search: Search gives links but generative AI gives answers! However, sometimes you must challenge the facts presented and you will occasionally get different answers.
Personal interests: how something works, historical facts, decoding medical diagnosis terms, baseball player performance and sabermetrics definitions.
Let me take you through some recent sessions.
Media overlaps can contaminate advertising lift studies with confounding factors. For example, you think you are comparing online video exposure to no exposure, but you really are comparing online video + some portion of display (highly correlated).
Cross-media reach (i.e. what percent of consumers see your ad SOMEWHERE) is a similar challenge but expressed relative to media planning and delivery. Same problem though…if somebody saw advertising one way, what is the chance they saw it another way too?
If you can estimate media overlaps, it would be of great value.
I came up with an idea for tackling this problem by leveraging what a marketer could know…the reach of ads on each platform individually and the correlations between exposure on the different platforms. Then I wanted to create a model that could simulate 10,000 virtual IDs so I could analyze overlaps.
I asked ChatGPT about Copulas. These are functions that contain the correlational structure of marginal variables that can have completely different distributions. The Copula plus its associated marginal distributions gives you the ability to simulate virtual IDs that should contain the right correlation patterns. At least, that is what I wanted to test.
Now, this is where ChatGPT really helped. I figured there might be a program in R to do what I had in mind. I told ChatGPT what I wanted, and it created the code including what packages I needed to install in R. It worked like a charm! Then I simulated 10,000 records from my new function and analyzed the results. ChatGPT reminded me of the code to export the files as csv files. In Excel I determined I got back the right correlations and now I was in business. My 10,000 virtual IDs gave me highly accurate cross-media exposure patterns.
Since this was such a new idea, I actually debated it with ChatGPT through a series of prompts and responses that kept building on each other. The back and forth really had the feel of a discussion with a colleague but eventually, I was convinced that I had a valid method.
Explaining this advanced approach to non-technical colleagues was not so easy so when I wrote up the methodology, I asked ChatGPT to simplify my language and make it understandable. I asked it for the 5 essential ideas. I asked it for a title. All of this worked very well.
Now all work and no play isn’t great so I interspersed some baseball stuff. I asked Bard which Yankees have the highest career OPS+ (on base percentage + slugging adjusted for year-to-year effects.) I got my answer! (I was surprised to see Aaron Judge right after Mantle but ahead of DiMaggio!) Then I was looking for a new show to binge, so I asked Bard for streaming video shows that had ratings on rotten tomatoes over 90. I got a great list! The next day, I got recipe ideas with Bulgar from ChatGPT.
The only limit to ChatGPT is your imagination. One researcher told me they even used it to create a code structure for open ended responses.
Please share some of your best stories of what worked (or didn’t) with ChatGPT.
