AI Success Stories (and some failures)
Exploring Niwot Climate Data Trends
Original figures from the Niwot renewal prep work (pre-AI)
Modified code in (? an hour or two can’t remember) to make nice aesthetics for the data dashboard (pre-AI)
Modified code in ~ an hour to update the import to read from EDI following EDI API modifications, all sites, split by season, calculate mann-kendall statistics plus Thiel-sen trends, render into multi-panel figure with embedded stats. ~391 lines of code.
AI meeting minutes
Used claude.ai to summarize zoom meeting transcription. Committee manually reviewed, edited and augmented content, checking against notes taken during the meeting. Guess ~80% of content was generated from AI summary.
Workshop facilitation
~ A room full of 80+ people each stated their research priorities. chatGPT synthesize into 5 themes for breakouts and further refinement. Ideas revised and iterated by humans, but this is a useful tool for generating break-out groups real-time.
Data Digitization
Tried a few options to get (neatly) hand-written data into spreadsheet format. parsio was the winner among those I tried, but your mileage may vary. Structured format is key here; data in neat rows and columns, one value per cell, single columns contain only numbers or text not a mix of both. CHECK results carefully, they will not be perfect. But it’s faster than double-key entry.
Data QC
Have had mixed success here. Would not use AI as the only qc step, but it can be useful. Here’s an example output when asked to find errors in a .csv of leaf trait measurements. I have had better success asking AI to write code to find errors in data. This has the benefit of being repeatable so you can check whether the errors are fixed.
This presentation
Used claude.ai extensively for debugging - figures not rendering, misbehaving .css, formatting issues, inscrutable error messages.
AI Failures
Too many to name! Some examples below:
First attempts at all coding things were filled with hallucinations, imaginary packages, functions that didn’t exist and nonsense code. Syntactically correct but completely disfunctional. (this was about a year ago, things are improving fast).
Many parsio alternatives. Won’t throw anyone under the bus by naming them here…
Asking chatGPT or copilot to make a pptx slide deck for me. Gemma did slightly better but still not saving me any time.
Asking chatGPT to make a cartoon illustration for a talk. No matter how much prompting I provided, we just don’t share the same sense of humor. Bing maybe a tiny bit better but still not fabulous.
Your Turn:
What have you tried? What worked, what didn’t?