● dot company

Researching intelligence from the byte up.

dot company is an independent AI research organization focused on multimodal, byte-level intelligence.

We explore what happens when models operate closer to raw information, reducing reliance on modality-specific tokenization and hand-designed representations. Our research investigates architectures and training methods capable of learning directly from low-level sequences across text and other modalities.

Research

Our work currently centers around:

• Byte-level modeling: Learning directly from bytes rather than conventional text tokenization.
• Multimodal intelligence: Developing representations and architectures that can generalize across different forms of information.
• Small language models: Studying how much capability can emerge from compact, compute-efficient systems.
• Efficient architectures: Exploring recurrence, parameter sharing, adaptive computation, and alternatives to simply scaling model size.
• Representation learning: Investigating how models can discover useful structure with fewer assumptions about how their inputs should be encoded.

Where we're headed

We believe increasingly general models may emerge not only from making systems larger, but from making their foundations simpler and more universal.

Our long-term direction is toward models that can consume heterogeneous information through a shared low-level interface, learn useful internal representations on their own, and allocate computation intelligently.

Instead of building a different foundation for every modality, we're interested in a more fundamental question:

How little structure can we impose while still allowing intelligence to emerge?

We're starting with bytes.