It’s certainly nice to have excavators instead of having to dig ditches by hand. But if you never get any exercise, that’s not good, either. It’s great to have calculators and spreadsheets to do math, but we have no mathematical intuition at all, it’s easy to fall victim to scams from advertisers, politicians, and media sources, throwing out numbers that don’t make sense. And it’s great to have AI tools to help with ideas, writing, etc, but if we outsource our actual thinking to AI, well…
I’ve started seeing people recommending having AI write your proposals, and I can see why this might be an appealing idea– who hasn’t sat down to write a proposal and realized you don’t know what to write? But you don’t want an AI, even one supposedly “trained in your voice” deciding exactly what problem you’re trying to solve, and how you’re going to do it. That’s your job.
Imagine going to the doctor, and instead of having the doctor discuss your surgery options, an AI-bot took the doctor’s notes and spat out some recommendations.
(Need an easier way to write proposals? Grab my “fill-in-the-blank” consulting proposal template, which provides the framework you need not just to fill in the words, but to ask the right questions during your consultation(s) to diagnose the issue(s) and propose the right solution(s).)
I’m actually working on removing one of my favorite Mimiran features, and the only place that Mimiran calls an external AI system– parsing text data into contacts. The functionality is still there, I’m just not using external AI, for several reasons:
- Privacy. In theory, this data is safe when going to the a 3rd party for process, and the settings for the integration say that this data can’t be used for training models. But it’s another possible source of leaking sensitive information, if OpenAI gets hacked, or someone discovers they have been using that data for training, etc.
- Consistency. With different API settings, you can get much closer to consistent answers than the typical ChatGPT interactions, but there is some nondeterminism that you can’t eliminate, which makes debugging issues quite frustrating.
- Speed. I just wanted something faster. 😉
While working on the actual parsing of text input (say you have an email signature or even a Zoom chat log with contact info for lots of people), I took the chance to make some other updates that I’ve been wanting to make, regardless of who does the parsing.
So now when you paste in your text, you get a preview of what the parser thinks is happening, making it easy to adjust contact fields, or even delete data for people you decide you don’t want to import, then run the import.