Researching alternative numerical foundations for machine learning.

Unfloat AI is a research and engineering organization investigating whether machine-learning systems can be designed around fundamentally discrete and integer-based numerical computation rather than relying on floating-point computation as their underlying numerical foundation.

We are exploring what becomes possible when the numerical foundation itself is reconsidered rather than simply optimizing an existing implementation.

Our research interests include:

• Discrete and integer numerical computation
• Neural-network training and inference
• Numerical representation and precision
• Machine-learning algorithms
• Computational and memory efficiency
• Resource-constrained AI
• CPU-oriented computation
• Training stability and convergence
• Hardware implications
• New approaches to machine-learning systems

The central research question is:

Can useful machine-learning systems be built around a fundamentally discrete numerical foundation?

We are investigating that question experimentally rather than assuming the answer.

We follow a simple research cycle:

Build → Measure → Understand → Improve

We aim to:

• Build working prototypes
• Design controlled experiments
• Measure results
• Compare against appropriate baselines
• Document findings
• Identify limitations
• Iterate based on evidence

A result is more valuable when it can be reproduced and independently understood.

Failed experiments are also useful. Understanding why an approach does not work can be just as important as demonstrating that one does.

Our work exists within a broader field of research involving areas such as quantization, fixed-point computation, binary and ternary neural networks, low-precision machine learning, integer computation and alternative numerical representations.

We actively study existing work and aim to distinguish genuinely new ideas from approaches that have already been explored.

We believe ambitious technical questions should be investigated with both curiosity and skepticism.

We do not assume that a different numerical representation automatically produces better performance.

Any meaningful advantage must ultimately be demonstrated through experiments and measurements.

Our goal is not to make extraordinary claims.

Our goal is to find out what is actually possible.

Evidence over assumptions.

Experiments over speculation.

Reproducibility over anecdotes.

Understanding over blind implementation.

Results over hype.

Our implementation, experimental systems and internal research infrastructure are currently maintained privately.

As the research develops, we may publish selected findings, technical material, benchmarks, or other work when appropriate.

For now, the public organization provides an overview of what we are investigating rather than exposing our internal implementations.

Unfloat AI

Research. Engineering. Experimentation.