Inside the Pod: AI and the Art of Efficient Creativity

Thank you to everyone who is watching or listening to my podcast, AI & I. If you want to see a collection of all of the prompts and responses in one place, Every contributor Rhea Purohit is breaking them down for you to replicate. Let us know what else you’d like to see in these guides.—Dan Shipper

Was this newsletter forwarded to you? Sign up to get it in your inbox.

---

Doing creative work isn’t all that it’s cracked up to be.

Don’t get me wrong—there is joy in bringing your ideas into existence. But in between the flashes of inspiration, eureka moments, and bursts of productivity, there’s a fair bit of drudgery. 

Take writing, for example.

I love researching and writing longform essays. But there’s nothing I find more arduous than dreaming up a headline.

When I finish a draft I’m proud of—after a quick victory lap around the kitchen table—my heart sinks just a little. The ghosts of titles and subheads past are back to haunt me! Though Every recently launched a tool called Spiral that ameliorates the horrors of headline writing, there are plenty of other tasks that I still have to endure in order to do the creative work I love, such as formatting citations, proofreading for typos, and dealing with backend newsletter software.

In this episode of AI & I, Dan Shipper interviews New York Times bestselling author Seth Stephens-Davidowitz, who wrote a book in 30 days—in large part because he used ChatGPT for all the parts that he didn’t like, the ones that would have inevitably slowed him down.

The book, Who Makes the NBA?, is Stephens-Davidowitz’s third data science book. He challenged himself to write it in less than a month because he realized how much faster he could work in collaboration with the ChatGPT feature Advanced Data Analysis (formerly called Code Interpreter). 

In this essay, I’ll pull out the core themes of Dan and Stephens-Davidowitz’s conversation (complete with accompanying screenshots!):

• The practical ways he used ChatGPT to write a book in less than 30 days
• How you can use AI to answer any question that you’re curious about—like which Olympic sport you’re most likely to medal in

What struck me the most about Stephens-Davidowitz is the pure joy he experienced from doing what he loved—thinking of interesting questions to ask a dataset—and the world of possibilities that AI tools opened up for him by taking care of the boring bits and allowing him to focus on that. I think anyone curious about how AI is changing what it means to do creative work will find this essay of interest.

How to use ChatGPT to do creative work faster

In writing Who Makes the NBA?, Stephens-Davidowitz outsourced all the tasks that he found tedious to ChatGPT. As an example, he took Dan through the way he:

• Made a fitting acronym for a metric to evaluate basketball players
• Generated charts with Advanced Data Analysis

As you go through these screenshots, you may notice that Stephens-Davidowitz had to go back and forth with ChatGPT several times before it generated output he was happy with. He notes that this iterative process was very important because it helped him clarify his own thought process.

Brainstorming ideas with ChatGPT

Stephens-Davidowitz created a metric to rank basketball players after adjusting for their height and named it after Muggsy Bogues, the shortest NBA player in history. He wanted the name of the metric to be a backronym—or a name that doubles as an acronym that describes the metric. Stephens-Davidowitz decided to use ChatGPT to help him brainstorm ideas for this.

Stephens-Davidowitz: I have a statistic called MUGGSIES. It is how good someone is in basketball adjusted for their height. Could you think of words that explain the stat for which that would be an acronym?

All screenshots courtesy of AI & I.

Stephens-Davidowitz wasn’t impressed with ChatGPT’s response because the suggestions were generic and the model went off the rails after the ninth suggestion, hallucinating more letters than necessary. He uses this example to highlight what he believes is a common misunderstanding about ChatGPT—that AI can produce great responses in one go. Instead, Stephens-Davidowitz argues that getting good results from AI is an iterative process.

Stephens-Davidowitz prompted ChatGPT again, this time including more detail about what he expected.

Stephens-Davidowitz: No, I mean give me something to describe the stats which is eight words and the first letter of the first word is M, and the second word is U, etc.

This was a better response, but Stephens-Davidowitz didn’t think it was good enough yet.

Stephens-Davidowitz: That should have two Gs, it only has one; and it’s not just about scoring.

Stephens-Davidowitz isn’t satisfied with the direction ChatGPT has taken. 

Stephens-Davidowitz: That doesn’t explain the stat at all, which is how good they are for controlling for their height.

Stephens-Davidowitz: That’s close but try again.

Stephens-Davidowitz: Can you try again? Efficiency is the wrong word—is there an E word that captures overall performance on a basketball court.

Drawing from ChatGPT’s previous responses, Stephens-Davidowitz stitches together a response that he likes. However, he realizes that the backronym is currently missing an S word.

Stephens-Davidowitz: I like: Metric For Understanding Game, Given Individual’s Effectiveness and Size. But there should be an S word before Individual.

ChatGPT added the word starting with S in the wrong place. 

Stephens-Davidowitz: NO. I want Metric For Understanding Game, Given ___ Individual’s Effectiveness and Size. But I want a word in the blank that starts with S.

Stephens-Davidowitz thinks ChatGPT can come up with a better response.

Stephens-Davidowitz: Try again. Offer me 10 possibilities.

Stephens-Davidowitz has high standards for ChatGPT. He prompts the model to come up with even more options.

Stephens-Davidowitz: Give me 20.

Stephens-Davidowitz doesn’t use any of the words that ChatGPT suggested, but the model’s responses inspired him to think of a word of his own: sporting.

Editing charts in real-time using AI

Stephens-Davidowitz uses ChatGPT’s Advanced Data Analysis feature to analyze a file that contains data about the height of NBA players. After he uploaded the file, the model returned with basic information about the dataset.

Stephens-Davidowitz asked ChatGPT to create a chart that measures the height of the players in inches based on the dataset.

Stephens-Davidowitz: Can you give me a histogram of height inches?

Stephens-Davidowitz acknowledges that creating this chart using traditional coding methods wouldn't be particularly time-consuming—however, it would be a task that he found unappealing.

While examining ChatGPT’s histogram, Stephens-Davidowitz noticed that on the horizontal axis of the chart, inches are recorded in groups. This bothered him, so he asked the model to tweak it.

Stephens-Davidowitz: Can you have every inch of height included?

As Stephens-Davidowitz scrolls through this chat, he emphasizes the convenience of having a tool that makes adjustments to a chart he’s working on in real time.

Stephens-Davidowitz: Can you include only American-born NBA players?

While Stephens-Davidowitz was satisfied with the basic form of the histogram generated by ChatGPT, he pushed the model to refine the way information is presented.

Stephens-Davidowitz: Can you point to the most common height and say what it is in feet?

Stephens-Davidowitz wanted to change the way the most common height is represented on the graph.

Stephens-Davidowitz: Can you get rid of the vertical line and instead put the label most common height: 6’ 9” inside the bar for 6’ 9”?

Stephens-Davidowitz: Can you have “most common height 6' 9”” be written vertically inside the bar representing 6’ 9”?

Stephens-Davidowitz notes that finicky tasks like this would be very tedious to carry out without the ChatGPT’s assistance. He proceeded to make minor changes to the chart, like changing the color of the bars to red.

Stephens-Davidowitz then used ChatGPT to add another data point to the histogram: a normal distribution of the height of American men. After a few rounds of iteration with the model, this was the final version of the chart.

How to use AI to answer any question that you’re curious about

Stephens-Davidowitz uploads a dataset of Olympic athletes throughout history—recording variables like height, weight, and type of sport—to ChatGPT. Dan wonders if they could use ChatGPT to analyze this data and discover which Olympic sport he is most likely to medal in, based on his own height, weight, and nationality.

Stephens-Davidowitz suspects that ChatGPT may not provide a good answer right off the bat, anticipating that they might need to take a step-by-step approach to tackling this question. However, he decides to attempt soliciting a direct response from ChatGPT.

Stephens-Davidowitz: I’m 6’2”. I’m about 160 pounds. What sport would give me the best chance of success?

ChatGPT begins its analysis by ensuring standard units of measurement—it converts Dan’s height and weight into centimeters and kilograms, aligning with the format used in the dataset of Olympic athletes. Then, for each sport in the dataset, ChatGPT quantifies the number of athletes who were within a range of 5 centimeters and 5 kilograms of Dan’s height and weight. This approach aims to identify Olympic sports featuring athletes with similar physical attributes to Dan. In the final step, ChatGPT ranks the sports based on this analysis, determining that Dan would be most likely to win a medal in Athletics, which includes a diverse set of track and field events. 

Stephens-Davidowitz comments that ChatGPT’s approach is interesting because it’s more straightforward than the one he had in mind. However, he also notes that there is an obvious flaw with the model’s method—ChatGPT’s result isn’t accurate because it’s skewed toward sports with more athletes overall. Stephens-Davidowitz prompts the model to correct this oversight.

Stephens-Davidowitz: Can you divide by total athletes in that sport? So what fraction are in that range? And show me the top 10.

ChatGPT's analysis suggests that Dan's physical characteristics align most closely with Olympic triathlon competitors. It reports that 12.67 percent of Olympic triathletes fall within 5 cm of Dan's height and 5 kg of his weight. To refine these results further, Stephens-Davidowitz proposes reducing 5 cm—the model’s “tolerance range” for height—to 2 cm.

Stephens-Davidowitz: Could you do 2 centimeters on the tolerance range?

Stephens-Davidowitz notices the common characteristic between these sports is that being tall is a big advantage. He's intrigued to find that Dan's height of 6'2" is considered advantageous even in basketball, where players often reach 6'7" or 6'8". Seeking to analyze the data with greater precision, Stephens-Davidowitz instructs ChatGPT to focus specifically on American athletes in the dataset. He also asks the model to include the total count of athletes in each category, along with the percentage.

Dan notices that modern pentathlon, which ChatGPT records as having 14 athletes within the tolerance range, is ranked higher than swimming, which has a greater absolute number of athletes within this range. He estimates that ChatGPT is likely adjusting its rankings to account for the larger overall number of American athletes participating in swimming compared to other sports. Stephens-Davidowitz recommends asking ChatGPT a few more questions to check that it is carrying out the analysis correctly.

Stephens-Davidowitz: Can you list all the seven canoeing athletes, their country, and their height and weight?

From ChatGPT’s response, Stephens-Davidowitz identifies a mistake in their methodology. He realizes that the original dataset contained repeated entries for athletes who competed in multiple Olympic games. As a result, ChatGPT mistakenly treated these repeated entries as distinct individuals, effectively double-counting (or more) some athletes in its calculations. As he prompts the model to correct this, Stephens-Davidowitz notes that the inaccuracy isn’t ChatGPT’s fault, but rather a limitation of their process that could only be discovered through trial and error.

Stephens-Davidowitz: Can you redo the analysis but only include each athlete once, even if they were in multiple Olympics.

The results are in! The Olympic sport that Dan is best suited for based on his height, weight, and nationality is the modern pentathlon. He wonders what sports the modern pentathlon includes, so Stephens-Davidowitz asks ChatGPT.

Stephens-Davidowitz: What are the sports in the modern pentathlon?

In Stephen-Davidowitz’s experiment with AI, you may have noticed that there isn’t a single function that ChatGPT carried out that he couldn’t do himself. Stephen-Davidowitz cherry-picked the parts of his work that delighted him—posing interesting questions to a dataset—and almost wished the rest away to AI. This gave him the momentum to accomplish a larger task—writing a book in less than a month—that he felt he couldn’t have done without AI. Creative work still involves a fair bit of drudgery, but now, AI can make the burden much lighter.

---

Rhea Purohit is a contributing writer for Every focused on research-driven storytelling in tech. You can follow her on X at @RheaPurohit1 and on LinkedIn, and Every on X at @every and on LinkedIn.