We study how to build machine intelligence that is continuous, memory-rich and efficient, and how to keep it safe. We publish what we find, including what does not work.
Research areas
Project Cortex1 postAn architecture for machines that reason and act in the real world. First results and the research paper.Research in progressProject Open-ATLAS1 postOpen datasets with a traceable origin on every record. First release: Theo, an engineering tool-use dataset.First release: TheoSafety4 documentsHow we decide whether, when and how to release a system: the Safety Framework, the Constitution and the Charter.
Featured
01 / 02
2 articles

Project Cortex: first results and the research paper
Three experiments on associative memory, fast control and continual learning, reported in full, including what did not work.

Theo: an open dataset for engineering tool use
36,349 conversations that teach small models to call deterministic engineering tools, ask when values are missing, and refuse to defeat safety protections.
Nothing in this area yet.