Our Mission: From Latent to Lexicon

Abstraction as the foundation of humanity

Much of what sets Homo sapiens apart comes down to our ability to create abstractions: referring to an out-of-sight object, inventing tools and technologies that had never existed, and entertaining philosophical ideas that question the very existence of everything.

Language is our shared system of symbolic abstractions. With it, we can communicate our inner feelings, everyday life, political ideologies, and scientific proofs.

We build abstractions upon abstractions. Algebra generalizes arithmetic; calculus captures change through limits; real analysis makes calculus rigorous; measure theory generalizes size and integration; Kolmogorov gave probability its modern measure-theoretic foundation; and probability is one of the central languages of machine learning.

Learned abstractions change what we can perceive. Infants can tell apart speech sounds from any language, but over their first year they lose much of that ability for contrasts their own language does not use. Chess masters recall real mid-game positions far better than beginners, yet with pieces placed at random, the advantage largely disappears. In both cases, the abstractions we hold shape what we take in.

We are able to advance science because we stand on the shoulders of giants. We learn what the last generation built, then build something new for the next generation to learn. In the same way, we are able to build new social structures because we have collective experience with old ones. We identify what makes people suffer and try to build better societies for the next generation.

The Human Ceiling

However, how much abstraction any single human can learn and create is limited by our biology and lifespan. Writing, print, and the internet have widened the channel, but every person still has to learn what came before, one conversation or one page at a time. That sets a pace for how quickly our civilization’s understanding can grow.

The Latent Vocabulary

Modern AI models have changed this. They are trained on a vast share of the text humanity has digitized, far more than any person could read in many lifetimes. To predict that text well across every domain, including ideas and theories that contradict one another, they are pushed to compress it: to form internal abstractions that serve many contexts at once.

Not all of these abstractions are sound. Some are spurious correlations, statistical shortcuts that happen to work on the training data. Others faithfully capture the structure behind the text. Interpretability research has already found features inside models that match concepts we recognize, and many that we cannot yet interpret. We believe some of these encode what we have collectively felt but never named, and that others are new abstractions so subtle or vast that no human could have arrived at them alone.

Note that all of it comes from the text humans wrote: we are not claiming that models know anything beyond their data, only that the data may hold structure no one has named. We expect that, once named and carefully explained, such abstractions can be understood, and we hold every term to that standard. Today we work from the models’ own proposals, not from their internals, so for now this remains a hypothesis we want to test.

Our Mission

Our mission is to explore whether AI models can help us find and name patterns we have overlooked. We work with them to surface candidate abstractions, name each one carefully, and share the ones that hold up, in the hope that a richer shared vocabulary helps people think about the world, and about themselves, a little more clearly.

We don’t intend to be a closed laboratory. We are building toward a community of thinkers, dreamers, and builders. As we open our tools and frameworks for exploring what AI models have learned, we invite everyone to help identify, name, and integrate these abstractions into our shared vocabulary.