Post

There are GenAI powered tools that create data visualizations with Python code, and give you the code that you can inspect. IMO no reason not to use those if you can read and understand the code.

That's a really good point tbh! Kind of curious what the output looks like now.

I uploaded a CSV of state population changes from the 2010 and 2020 US census and asked ChatGPT to graph the states with the 10 largest changes (positive or negative). It gave me a graph and also the Python code when I asked for it. It also asked me if I wanted to change the graph in any way. #GenAI

Even with the largest context windows available you're at risk of running out of space when trying to fit in your dataset anyway. And visual limitations (platform.openai.com/docs/guides/...) make OpenAI at least completely useless verifying the locations of points or relative sizes of bars.

Yes there are like 5 more reasons ranging from ethics to ecology but from a purely practical and professional standpoint the biggest concern IMO is feeding your readers utter bullshit

The "abstraction" point I think is a really good one that can be applied more broadly. Most people aren't great at working at different levels of abstraction simultaneously, so anything that operates at a higher level of abstraction than your data is a risk. Non-deterministic and bullshitty: riskier

So this also applies to summarisation and a number of other text generation uses.

I think that genAI/LLMs, if anywhere, should be used as a pointer rather than a result producer. We can use them to get guesses for where to look, suggestions, ideas; and then do the actual work properly ourselves to avoid propagating any nonsense hallucinations. It's still often unnecessary.