I do not expect anything other than Apache 2.0 or MIT (our go-to licenses) for code and checkpoints. Data is a bit harder, but we will always try to release everything under the most permissive license we can.

Yes, come on in! We do not expect contributors to devote massive amounts of time or go out of their way to contribute—they all have lives of their own, and we respect that.

As for collaboration, we make it work. Most communication is asynchronous text, which is quite versatile and hides varied schedules well.

To clarify: EleutherAI will continue to work with large language models and train its own when there is a clear research case as it always has—there just happens to be a much larger saturation of suitable models today for the research we would like to conduct than what existed even twelve months ago, and there is no reason to reinvent something when something suitable already exists. Expect new models to be designed and trained to specifically meet certain research requirements, rather than more versatile usage.