Over the weekend, Andrej Karpathy shared this tweet & it impressed me to conduct the 2024 GTM Survey Evaluation this fashion.
I take advantage of a language referred to as R to research knowledge due to its potential to generate fairly charts & the depth of its statistical evaluation instruments. Inside 90 minutes I had discovered 9 key knowledge factors within the knowledge that had been statistically important & runs greater than 50 analyses. Prior to now this sort of work would have taken me 30 to 40 hours.
Programming is changing into prompting.
I used to jot down one thing like this to generate a chart :
ggplot(knowledge) + geom_bar(aes(x = variable, y = worth), stat = "id") + theme_minimal() + labs(title = "Title", caption = "Caption")
Copilot autocompleted the totally different fields. However utilizing Sonnet & Cursor, I first wrote “Carry out a conjoined evaluation, evaluating the correlation throughout all variables throughout the knowledge body. Plot this on a bar chart utilizing my specific theme, with an insightful title & a caption for Idea Ventures.”
Then I wrote “Run the identical evaluation for gross sales quota in comparison with firm measurement.” Subsequent, “how about NDR for firm measurement?” Every time, the robotic produced 150 traces of code in seconds.
Extra than simply the code, I request a take a look at for statistical significance. I remembered from statistics class in school to carry out a t-test for evaluating two means when the pattern measurement is bigger than 35. However I had forgotten evaluate the means throughout greater than two teams. ANOVA to the rescue.
The entire code is formatted in accordance with correct syntax & it really works. The one errors I discovered involved colour palette specs.
English is the brand new programming language. Coding this fashion, I explored the information rather more deeply, extra rigorously, & extra rapidly than I’d have in any other case.
The consumer nonetheless wants to pay attention to the underlying syntax to repair errors & some statistical exams to confirm the pc is doing the proper factor, however gone are the times of memorizing the practical arcana of particular person programming libraries.
In different phrases, I’m working at the next stage of abstraction. Although it could not appear this fashion, the consumer interface of information exploration has modified. It’s a again & forth with the pc, a dialog, a dialogue with ongoing output. I’m eager about the following evaluation, not the following practical argument.
