AI topical fixation

[jwhco]

Try this prompt with your Obsidian vault:

Provide for me a list summarizing dominate topics addressed in this notes repository. Include no more than five topic clusters. Each summary will include the big idea encompassing that topic. Do not repeat topics. Use short sentences and benefit oriented description please.

I noticed responses repeated certain topics. What's interesting about the response is the composition of the assistant lookup and system context.

The system context includes artificial intelligence + machine learning (AI-ML) operational risk, keyword clustering notes for SEO, and a project on note-taking. Yet the dominant topics highlighted are time management, relaxation, and mindfulness.

Productivity and time management are not relevant to AI-ML risk. Although the summary implies that using AI-ML in risk management will save time, it isn't contextually meaningful.

Also, the topics repeat. So, it doesn't follow the original prompt. While AI is cool, it doesn't work the way we think it does when working with subject matter expertise.

Prompts don't matter as much as conversation. I followed up with What does AI-ML risk tools mean? Please describe. to which it replies:

The model has one thing in mind (a simplistic view). It's not thinking or developing relationships beyond what is already spelled out. Then again, it may see these things as unique and related where I do not. What are your thoughts?

[jwhco]

After some tweaking, I'm getting better results with the prompt above.

I get a much better response now. It's more aligned with what I expected.

Who knows if it is using a clustering algorithm. Label propagation gives me something different.

The fixation isn't SC. It was the fidelity of the model.