lunarain

Glad you found the lookup approach interesting! We don't have public benchmarks to share just yet, as the focus right now is on the architectural grounding.

A lot of 'new' tech is built on fundamentals that haven't actually changed, but we often ignore existing solutions to memory and sparsity because they aren't 'trendy.' This is an attempt to stop ignoring those 'good ol' problems' and apply proven data structures to the attention bottleneck.

Interesting, do you support multi-turn prompts evals as well?

how many people working on demand for you?