Main Content

Results for


How long until the 'dumbest' models are smarter than your average person? Thanks for sharing this article @Adam Danz
David
David
Last activity on 1 Apr 2024

I was in a meeting the other day and a coworker shared a smiley face they created using the AI Chat Playground. The image looked something like this:
And I suspect the prompt they used was something like this:
"Create a smiley face"
I imagine this output wasn't what my coworker had expected so he was left thinking that this was as good as it gets without manually editing the code, and that the AI Chat Playground couldn't do any better.
I thought I could get a better result using the Playground so I tried a more detailed prompt using a multi-step technique like this:
"Follow these instructions:
- Create code that plots a circle
- Create two smaller circles as eyes within the first circle
- Create an arc that looks like a smile in the lower part of the first circle"
The output of this prompt was better in my opinion.
These queries/prompts are examples of 'zero-shot' prompts, the expectation being a good result with just one query. As opposed to a back-and-forth chat session working towards a desired outcome.
I wonder how many attempts everyone tries before they decide they can't anything more from the AI/LLM. There are times I'll send dozens of chat queries if I feel like I'm getting close to my goal, while other times I'll try just one or two. One thing I always find useful is seeing how others interact with AI models, which is what inspired me to share this.
Does anyone have examples of techniques that work well? I find multi-step instructions often produces good results.
Unlike last year's contest, there are some new technologies this year that might offer some advantages. Namely generative AI's like ChatGPT, Bard, etc. Not to be excluded, MathWorks just launched the AI Chat Playground :)
The contest development team has identified an issue when trying to link your new entries to submissions in the File Exchange. The issue has prevented some users from successfully linking their entries. We have a fix and will be deploying it today at 3pm EST.