back
3 comments
The article says that you should not use AI if high accuracy is required. I would rephrase this to "Do not use AI if you need correct answers". Asking Gemini for the derivative of a rather simple function was quite eye-opening. It came out with a detailed step-by-step solution, explaining each step with great confidence. But the result was wrong. Pointing out the wrong result made Gemini apologize and provide another wrong solution. Then pointing out the exact step where the error occurred resulted in another apology, and in exactly the same wrong result as before.

With Wolfram Alpha you get a result that you can trust (or no result if it is outside the scope of Wolfram Alpha). With AI you get a result that you cannot trust, although it is presented with high confidence. This can be worse than useless.

I often use it for a type of work that is not listed there: boring work that I could do just as well, but would take much longer, and is easy to verify when done by the AI.

For example as a scientist this is the case for data visualization, I sometimes use ChatGPT to provide it with some tabular data and get nice Seaborn visualizations (including some iterations to adjust: "make the legend bigger", etc.).

Could I learn Seaborn? Sure, but I'd rather not spend time on that as it's not a particularly enriching activity. And even if I already knew it, writing and tweaking plotting code wouldn't be exactly fun.

And I'm not in danger of making mistakes due to the AI because the code in such libraries is quite self-explanatory. Memorizing the functions can take some effort, checking a program to see what it's doing is trivial.

> When AI is bad.

This is described as something of a mystery. How about: imagine if words were turned into numbers and all you know is what order they tend to come in?