Models compress the universe into their weights through beautiful, intricate, and complex mechanisms. We understand much less than we can build – and that tension leaves us in the compelling position of having many more questions than answers.

To make progress on that gap, we believe you must first understand intelligence itself in order to even pose the right questions. As we build more and more capable models, the search space for principled ideas to test grows exponentially; mechanistic insights provide the A* heuristic that guides us toward better model design.

The future of intelligent systems is profoundly exciting, and we’re eager to shape it together. Ultimately, there is no greater pursuit than understanding – not as a means to an end, but as an end in and of itself.