(aka: Pushing the level at which I lean on assistants)
(In my previous article, New Principles of Learning with AI, I talked about some of the things I’ve noticed in using ChatGPT-4o and copilot to help me build my understanding of AI and ML. This article goes in depth about one such principle. I’ve talked with a few friends who found this revelatory, so I decided to give it its own space.)
So, in my AI class, the first few weeks were mostly data prep, simple analysis and visualization. For the assignments, the output was a report with decent plots which would underline the analysis.
To generate these plots, we were using three different libraries, Matplotlib, Seaborn and Plotly.
For the first weeks, I would timidly ask ChatGPT how to do simple things, like:
“How can I use the split function in pandas?’
“how can i select the 'country' column from the first row in a pandas dataframe”
“how can i change the legend labels on a matplotlib bar chart?”
As things progressed, I became bolder, giving the ai more abstract goals, with more precise constraints:
“What are the best plots in seaborn to show distributions between two or three categorical variables?”
“How can I change the heatmap to go from bright red for 0 to bright green for 1?”
“i'd like to do a set of plots which show the accept/reject rate for different subsets of the population”
I noticed that my conversations became longer and more discussion-oriented than pure question/answer with the assistant. I was asking for more wisdom than pure knowledge, and that meant more interaction and subtlety.
Finally, there came a day when I didn’t even know how to ask for what I wanted. Since ChatGPT had been happily delivering code each visualization I requested, I had the following exchange:
This astounded me. I just pointed the AI at the desired output and asked it to give me the code. I didn’t even know that this was a “heatmap with marginal histograms” prior to this, I had just seen another student use this type of plot.
What does this mean for how I use the assistant? It means that, at least in this domain, I can communicate with images, rather than with text. I don’t have to describe what I want, I can just find an example. This is pretty handy considering I have the entire Internet to draw on.
Does this mean that, perhaps now, perhaps soon, I can give it an example video and ask “how can I do this kind of scene transition in Premiere?”, or point it to a particular page of a website and ask for a WordPress template using that layout and color scheme?
Indeed, this means that the expressive language I use in my “prompts” is becoming richer, and the output of the AI could cover any type of desired output format (“define an API which…”, “what is the VBA code to…”).
The hard part of this is in leveling up my own thinking about what is possible. I’ll be writing more about the ramifications of this in the coming weeks.
コメント