Claude 3.5 has pushed a new preview feature to me: the Analysis Tool.
So, I downloaded GDP per capita data (PPP, constant 2017 dollars) for various countries and regions from the World Bank via OWID to give this feature a simple test. This is panel data. Functionally speaking, if Claude 3.5 can process this data accurately without errors, then handling more complex data analysis tasks is only a matter of time.
The data looks like this:
Over 6,500 rows of standard panel data.
My prompt was also exceptionally simple: "Analyze this data."
The model directly identified the top and bottom five countries/regions and provided a simple analysis: everything was correct—the years, the corresponding values, and the rankings.
Next are some standard prompts.
"Create a chart."
"Analyze Asia specifically."
Finally, my usual trick: "Write a PPT to expand on this, in Chinese."
To showcase the interactive effect, I directly uploaded a screen recording.
Throughout the process, I gave the model no subjective input—just simple commands to analyze data, analyze Asia, draw charts, and generate slides in Chinese.
I know that in at most a year, at least one of the three major models—ChatGPT, Claude, or Gemini—will make me unemployed:
Being able to write code: Given accurate data input, the model can autonomously complete correct analysis tasks.
Not limited to just the data: It can also extract valuable information from its own "knowledge base" to summarize and conclude.
Interactive visualization: This is already a strength of these models.
Wait, weren't models supposed to have hallucinations? Being able to write code solves hallucinations.
Weren't models supposed to not understand time series? Being able to write code lets them understand time series.
Weren't models supposed to be bad at math? Being able to write code lets them do math.
Is being able to write code that impressive? Yes, extremely, most definitely.