The AI lab at the frontier of speed
Performance-critical software often runs far slower than its hardware allows. Flash One develops AI methods to close that gap, and to verify that the optimized software remains correct.
Method
Every project begins with measurement, before anything is changed: hardware counters for cycles and memory traffic, and traces of where the tokens are spent.
We develop AI methods that search for faster ways to perform the same computation: in how data is laid out and moved, how work is batched and scheduled, and which instructions the hardware is asked to execute.
Verification is part of the method, not a final step. Every optimized version is checked against the original before it is adopted, and a speedup counts only if the result remains correct.
Research
One discipline, applied to three kinds of cost: running a model, the tokens it consumes, and the microseconds a system takes to respond.
Current focus
Token-by-token decoding is typically bound by memory rather than arithmetic: every step moves the model’s weights and cached state through the memory hierarchy, and that traffic largely sets the speed and cost of each token.
Every token an AI system reads or writes is paid for, whether or not it improves the result. We study how context is assembled, cached, compressed and routed across models, so that the same quality costs fewer tokens.
First result: the matching engine
In systems where microseconds carry economic value, the bottleneck is often not the computation but how data moves through memory: cache misses and pointer chasing.
Products
First result
A patented architecture for matching orders against an exchange’s order book: a different algorithm, not a faster implementation, derived from finite-field algebra, optimization theory and category theory. We license it to exchanges as IP rather than sell it as a product.
In research
Our research and products going forward apply the same optimize-and-verify method to AI inference: making large language models, and the systems built on them, faster and cheaper to run.
Contact
We welcome correspondence on research collaboration, licensing and evaluation.
New York, NY