Comment by logicchains

2B models by themselves aren't so useful, but it's very interesting as a proof of concept, because the same technique used to train a 200B model could produce one that's much more efficient (cheaper and more environmentally friendly) than existing 200B models, especially with specialised hardware support.