Hofvarpnir Studios

HOFVARPNIR STUDIOS BELIEVES ARTIFICIAL INTELLIGENCE IS A TRANSFORMATIVE TECHNOLOGY IN THE 21ST CENTURY THAT SHOULD BE ALIGNED WITH HUMAN PREFERENCES AND BEHAVE AS EXPECTED. AS NEURAL NETWORKS BECOME MORE COMPLEX, STUDYING THEM REQUIRES SIGNIFICANT COMPUTATIONAL RESOURCES. THE ORGANIZATION'S MISSION IS TO BRIDGE THE GAP BETWEEN AI SAFETY ACADEMICS AND HIGH PERFORMANCE COMPUTING BY PROVIDING THE INFRASTRUCTURE THEY NEED TO ADDRESS CHALLENGES IN DEVELOPING SAFE AI SYSTEMS. THE ORGANIZATION IS PARTNERING WITH SAFETY LABS TO BUILD GPU CLUSTERS FOR LARGE NATURAL LANGUAGE MODELS LIKE GPT-3 AND REINFORCEMENT LEARNING WORKLOADS. THE ORGANIZATION IS ALSO PROVIDING A SMALLER CLUSTER FOR INDEPENDENT AI SAFETY RESEARCHERS. OVERALL, THE ORGANIZATION'S GOAL IS TO SUPPORT AND ENABLE AI SAFETY RESEARCH THROUGH PROVIDING SPECIALIZED COMPUTING INFRASTRUCTURE.

DURING THE REPORTING PERIOD, THE ORGANIZATION CONTINUED TO SUPPORT ACADEMIC AI SAFETY RESEARCH LABS SUCH AS THE CENTER FOR HUMAN-COMPATIBLE AI (CHAI) AND STEINHARDT LAB AT UC BRKELEY. ITS EXPENDITURES WERE PREDOMINANTLY COLOCATION AND MAINTENANCE COSTS OF THE SERVER HARDWARE AND RELATED COMPONENTS PURCHASED IN THE PRIOR YEAR TO CONSTRUCT A COMPUTE CLUSTER, WHICH IS UTILIZED BY THE AFOREMENTIONED AI SAFETY RESEARCH LABS AND DOZENS OF INDEPENDENT RESEARCHERS. THE COMPUTE CLUSTER EMPOWERED RESEARCHERS WHO WOULD OTHERWISE NOT HAVE THE RESOURCES NECESSARY TO INVESTIGATE COMPLEX NEURAL NETWORKS THAT REQUIRE SUBSTANTIAL COMPUTE RESOURCES AND PUBLISH WORKS THAT ADVANCE THE FIELD OF AI SAFETY.