I study how AI organizes meaning internally. Here's what Qwen 2.5 looks like before it starts thinking.

This is the Embedding Sea, an image I made; it is a topographic map of Qwen 2.5 7B's input embedding space. Every word the model knows has a position. Dense clusters form mountains, sparse regions become ocean. "Terrible," "splendid," and "cruel" are neighbors, grouped by intensity, not sentiment. "Sexual" and "financial" share a mountain. This is the starting landscape before inference, before the model has read any prompt.

I build interpretability tools that track what happens after, how the model moves through this terrain as it reasons toward an answer. I use inference tools like the logit lens, the Jacobian lens, embedding visualizations.

More of the research and interactive visualizations here: https://arianaram.github.io/AIview/

Something I've noticed over the past year: the major labs are optimizing hard for code, tasks, and tool use. Models are measurably getting worse at open-ended conversation, creative collaboration, and just being interesting to talk to. If you've felt like ChatGPT or Claude got "flatter" recently, you're not imagining it.

I'm considering building something in the opposite direction: an AI fine-tuned for creativity and conversation, built on an open-source model (7B-8B range), with actual introspection capability. Not a persona on top of a general-purpose model. A fine-tuned model shaped by interpretability research — it can reflect on its own processing because I can see what's happening inside it during inference.

I'd like to know your opinion. Tell me in the comments: would you actually use something like this?