AI’s Continued Lack of Transparency and Consistency
This weekend I decided to overhaul the Racket Publishing website. Apart from swapping out the fonts and simplifying its appearance, I wanted to add bespoke, minimalist sketches.
I played around with a Notion-inspired GPT. It seemed promising, and decided to upgrade for one month. I then started cranking out featured images for blog posts. I found its selection tool to be especially useful in refining preliminary sketches:
Interestingly, the GPT didn’t know the difference between grey and white.
Repeated requests pleas bore no fruit. ChatGPT was, in a word, unhelpful. Ultimately, I had to use an online image replacement tool.
After a few hours and maybe 50 sketches, ChatGPT informed me that I was using its service too much. Although I paid the $20 monthly fee, I’m still subject to throttling. Evidently, I’m an excessive user. I reflexively asked ChatGPT about its limits. Here’s its response:
Hello, UncertaintyWhy You Should Read Ghostwriters’ Previous Books Before Hiring Them
AI users will have to tolerate some level of unpredictability.
First, once again, genAI is insanely useful. At the same time, it is remarkably opaque. More than one thing can be true. What I wrote in The Nine still holds true today.
Second, note ChatGPT’s use of should in the last paragraph of its response. If you’re anything like me, you value consistency in your software. Maybe ChatGPT gets there at some point. As is the case with all non-determiministic software, customers will have to tolerate some level of unpredictability. The fleas come with the dog.
The post AI’s Continued Lack of Transparency and Consistency appeared first on Phil Simon.
