Questions

What question was I exploring?

Inspired by Perplexity, I wanted to recreate a simplified version of their web search-based research query to explore how this facet of AI development works and also to remove reliance on an external platform for this capability. It is an efficient way to get news delivered across a wide spectrum of topics.

What did I learn?

Through this, I learned about grounding with web search (link tbd). I am particularly impressed and delighed in the way that most models create "query trees" to break prompts into several search queries, each of which are then parsed for responses to summarize. This is certainly a frontier of the web.

What feels unresolved?

Beyond ensuring that it runs smoothly every time it runs and refining the query prompts to be more precise and use qualified sources, I think I see two main value-adds for this. One is to perfect the art of synthesis. That is, making the summarized responses feel human and instantly readable. The second is simply to make it fun. It would be nice to include more diagrams, pictures, or illustrations instead of rote text generation. (Oh and a third would be to track responses over time!)