Teaching Everyone to Fish for Tokens

Access to a finished AI model and access to the recipe for making it offer different kinds of independence. Nathan Lambert, who helped build Olmo at Ai2, distinguishes downloadable weights from training code, data and methods. The latter gives other organizations a foundation for developing subsequent models themselves.

His economic argument is that NVIDIA can benefit when more organizations build and run models, because that creates demand for its hardware. Lambert says it is not clear that this will work. If the returns do not sustain competitive open development, his most likely outcome is that open models fork toward efficiency, modifiability and specialization, filling a long tail of uses such as private-data business agents while closed labs keep the most valuable areas.