Research
Automated research, in the open
We believe automated research is key to using AI in a safe and controlled way. So we build research tools, publish what we find, and say what we do not know.
ldgr-research
We open sourced ldgr-research to help machine learning researchers run long, open-ended automated research. It keeps the work structured across sessions, and it records what was tried, what worked, and what did not.
Our first paper
Our first paper asks whether one shared state can handle conflicting demands on timing. It cannot. Loosening the coupling between attack and recovery helps, but a single shared state still performs 7.7 times worse overall, and 510 times worse in the gap that matters most.
What we are working on now
Memory in the model.
We are researching how to integrate our memory system directly into the model architecture, instead of running it as a module beside the model.
One substrate for every input.
We are connecting multimodal inputs, text, images, audio, and more, to a centralized substrate, so one memory serves every kind of signal.
Automated design and prototyping.
We are looking at automating adversarial build systems, where one side builds and the other tries to break it, to speed up iterative design and prototyping.