Hum's core intuition is that the right training signal for intelligence about the world is the world itself. To do so requires a model that learns and understands generative processes rather than a human account or interpretation of them. With sensors building a deep, continuous record of how Earth (and space) behaves we now have the ability to see in real-time where energy flows, where matter accumulates, how systems interact and expand. Hum is building a foundation model trained on that record, capable of learning structure and possibility that we’ve never been able to comprehend. The end state is a zip file of the world - intelligence that internalizes and expresses the planet’s past, present, and future.

Earth observation training data is to physical intelligence what the internet was to LLMs. This dataset is so large and under-leveraged that it is likely to exceed the level of intelligence that LLM’s have achieved via text. It is also an inexhaustible dataset — the physical data we generate compounds in every second of life on Earth. We believe this is how Hum becomes the default intelligence for any system that needs to interact and reason with the physical world, which is most systems worth building!

Three months from finishing her PhD in quantum machine learning, Kelly walked out, convinced that the field had the underlying assumptions wrong. According to Kelly, they had it backwards: learning a language about the world (text, our descriptions, our abstractions) and trying to reason back to physical reality. As she saw it, the physical world doesn't need to be translated into human symbols to be legible as it already has its own grammar. Instead, we should learn and understand that language.

In Hum's world, any question can be posed to the Earth, and Hum’s model composes the relevant data and knowledge required to return a comprehensible answer. Hum shifts the boundary of what’s possible to the point where the constraint is how creatively we are willing to interrogate possibilities in the physical world.