Somewhere in a Poincaré ball, 32,070 tokens from GPT-2's vocabulary are quietly waiting to be explored. Not as a flat map you squint at, but as a hyperbolic tree you can actually fly through on your phone. The creator, responding to a Reddit user's disappointment with an earlier 2D projection, built this interactive model using nothing but the raw token embeddings. Drag, pinch, tap. The space shifts around you in a Möbius translation, the natural way to move through hyperbolic geometry. It's a small, precise act of making the invisible visible.
This is the kind of tool that should make you stop and reconsider what you think you know about AI's inner workings. We often talk about models as if they're black boxes, but here's a direct invitation to see the vocabulary's similarity structure as a forest: one giant tree with about 2,300 tokens, a few hundred smaller families, and around 6,700 isolates with no close relatives. Trees don't fit well in flat space, but they embed naturally in hyperbolic space, where room grows exponentially as you move away from the center. No optimization, no training. The layout is constructed exactly. That's not just clever; it's a quiet lesson in why geometry matters. For anyone who's ever felt constrained by the flatness of traditional data visualization, this is a practical nudge toward a different way of seeing. It echoes something we've touched on before when we looked at Clean Data Starts With Catching AI Slop Before It Skews Your Model: the structure underneath the data shapes what you can learn from it. And when that structure is a tree, you need a space that respects that.
What we appreciate most here is the restraint. The creator isn't claiming to have solved AI interpretability or to have built the next great analytics platform. They're showing you something honest: a direct visualization of what a model's raw embeddings look like when you stop forcing them into a grid. That's a rare and refreshing approach. It reminds us of the mixed feelings we explored in Talking to My AI Clone Taught Me to Question the Tech, where the more we interact with AI, the more we realize how much depends on the interface we're given. Here, the interface is the insight. You can tap a token, bring it to the center, and watch the entire space reorient around it. That's not just a demo; it's a way to build intuition. And intuition is what most of us lack when we think about embeddings.
So what should you do with this? Open it on your phone. Spend five minutes flying through the vocabulary. Notice how some tokens cluster tightly while others sit alone. Ask yourself why certain families form and what that says about how GPT-2 organizes meaning. The specific takeaway is simple: hyperbolic space isn't a gimmick, it's a tool that fits certain data structures better than anything flat ever could. If you've ever tried to make sense of high-dimensional relationships, this is the clearest argument yet for questioning your assumptions about the space you're working in. The geometry you choose changes what you can see. Now go see for yourself.