This is a submission for Weekend Challenge: Passion Edition
What I Built
I built Passion Atlas — an AI-powered map of human curiosity t...
For further actions, you may consider blocking this person and/or reporting abuse
A living map is a good format for curiosity because interests rarely sit in neat categories. The interesting product question is how the map changes after use: does it only display declared interests, or can it reveal unexpected neighboring passions from the paths people take?
That's exactly the direction I was thinking about. I don't see it as a static map of declared interests—it should evolve with how people actually explore it.
Every interaction could become a signal: the paths someone follows, the stories they spend time on, the connections they revisit, or even the ones they ignore. Over time, the map would adapt and reveal neighboring passions that weren't explicitly stated but emerge from their curiosity journey.
I think that's where it becomes interesting. Instead of asking "What are you interested in?", it starts answering "Based on how you've explored the world, here's something you might never have considered—but it fits your curiosity surprisingly well."
That interaction layer is where the map becomes interesting. Declared interests are a starting sketch, but followed paths and repeated returns are closer to real curiosity. I would be careful to keep some serendipity too, so the map does not overfit the first few clicks.
That's a good point. I think one way to achieve that is by making curiosity decay rather than permanence part of the model. Early interactions would have relatively little influence, while long-term patterns and diverse exploration would carry more weight. It keeps the graph flexible, allowing people's interests to evolve naturally instead of locking them into an identity based on their first few sessions.
Curiosity decay is a strong way to put it. It also avoids the classic personalization trap where the first few sessions become a permanent label. A good interest graph should remember patterns without turning early behavior into identity.
Exactly. I'd also want the graph to recognize that curiosity has seasons. Something that fascinated you years ago might suddenly become relevant again because of a new experience or interest. Instead of treating passions as fixed labels, the graph would model them as relationships that strengthen, fade, and reconnect over time.
That's where I think Snowflake and Gemini complement each other well. Snowflake can capture the long-term interaction history and evolving patterns across millions of contributions, while Gemini can reason over that context to identify when seemingly unrelated experiences have become connected again. Rather than just remembering what you've clicked, the system could understand how your curiosity is evolving and surface connections that feel timely instead of simply familiar.
The chains you're showing — Travel → Cultural Exchange → Local Food Traditions → Storytelling — read like exactly what an LLM produces when you ask it to relate two concepts: a smooth, plausible, generic path. That's the trap. Gemini will always find a connection, and it'll find a similar one for almost any pair of inputs, because interpolating between concepts in embedding space is the one thing these models can't refuse to do. The "meaningful surprise" you're optimizing for and the "hallucinated but confident bridge" failure mode are the same output — you can't tell them apart from the text alone.
The thing that would actually make this defensible is the part you're leaning on Snowflake for: connections grounded in real co-occurrence across contributors, not connections the model invented on the spot. "People interested in ceramics often explore gardening" is only worth anything if it's measured from actual contribution data, not generated. So I'd flip the architecture — let the graph agent propose edges from real data and use Gemini to explain them, rather than letting Gemini author edges and hoping they're true. Otherwise every user's Passion Genome is equally profound and equally made up, which means it's neither.
That's a great point, and I agree that grounding the graph in real contributor data is important.
I was actually thinking of a hybrid approach. Rather than having Gemini blindly generate every edge, an agentic workflow could let Gemini propose new edges and hypotheses based on semantic reasoning, while Snowflake provides the evidence layer through real contributor data, co-occurrence patterns, and feedback loops. The proposed connections could then be validated, reinforced, or discarded as more data comes in.
With Google's recent agentic AI capabilities, Gemini is becoming much better at planning, reasoning, and using external tools, so I don't necessarily see it as only an explanation layer. I see it as a collaborator with the graph—discovering potential connections while the underlying data continuously grounds and refines them. That way you still get novel discoveries without relying solely on LLM intuition.
Really appreciate the thoughtful feedback—it gave me a lot to think about.
Thinking about it more, I can actually see the next iteration evolving into something quite different from the current demo.
Instead of a mostly AI-generated graph, the Passion Graph would become a living knowledge graph built from real user contributions. Agentic AI would continuously ingest new stories, propose new relationships, detect emerging patterns, and surface anomalies, while Snowflake becomes the long-term memory and source of truth. As more people contribute, the system would become progressively less dependent on priors from the LLM and more grounded in collective human curiosity.
The exciting part is that it could discover entirely new intersections that neither humans nor the model explicitly knew about—because they're emerging from real-world contributions rather than being generated from embeddings alone.
Really nice Ujja! I enjoy your youtube viedo and everything with your project. I am curious if you are going to share on LinkedIn also?
Thanks Benjamin. Didn't post it there yet.
ok!
Great ❤️
Thanks 🙂