DOE Opens Contributions for Genesis Open Models, a Planned Open AI Infrastructure for Science
The U.S. Department of Energy has opened the Genesis Open Models portal to collect contributions for training a future scientific AI model
On August 7, the U.S. Department of Energy opened the Genesis Open Models portal. Universities, national laboratories, companies, and nonprofit research organizations can offer data, code, models, evaluations, and working environments. Submissions of foundational materials will be accepted until August 14, while tasks and environments for further training will be accepted until August 25. At this stage, contributors submit a description of the proposed contribution and information about it.
A scientific task rarely consists of a question with a ready answer. A researcher may receive old code, an incomplete calculation, or conflicting logs from previous runs. The researcher must choose the next test, run it, determine what caused any failure, and decide what to do next.
In July, the U.S. Department of Energy selected 278 Genesis Mission projects that the program is expected to support with models and computing resources. The new portal is collecting data, tasks, and working environments for training and evaluating Genesis-Science-1.
Genesis-Science-1 is a planned scientific AI model that the U.S. Department of Energy is developing with Arcee AI. The model is intended to train in working environments that contain datasets, documentation, tools, logs, partial results, and evaluation rules. Within such an environment, an evaluator can inspect the task conditions and the result of each step.
The description of Genesis-Science-1 states that the future model will operate in an isolated computing environment. It is expected to create a plan, use permitted tools, retain the task state, and resume work after a failure. An activity log will record tool calls, code changes, data used, intermediate files, and conclusions. A specialist will be able to use this log to determine whether the result is reproducible. People will make decisions about safety, publication, and computing resources.