Meta has acquired enough GPUs and is preparing to train a model that rivals GPT-4 in early next year. The question is whether it would be open-source or closed-source.
Resolves YES if it's open-source.
Resolves NO if it's closed-source.
Resolves N/A if Meta does not release a new LLM in 2024...
@Sodra Perhaps it would be a good method to determine whether an LLM is open source based on this authoritative definition (https://opensource.org/osd/). Llama 2 only meets 6-7 of the criteria in the open source definition, so it is not open source, or it can be considered as pseudo open source.
@Sodra It is difficult to fully meet the ten requirements of open source. Would it be a good standard if we consider "the general public perceives the new LLM as open source" or "the new LLM meets eight or more of the ten requirements" as the criteria for determining whether a new LLM is open source?
@kr Those are too subjective. It should meet the opensource.org/osd like you say.
Llama2 doesn't meet it. #1 they don't release the data #2 they have a commercial restriction
#2 alone is enough to be not open source
@kr When most people use the term "open-source" in the context of language models, they're referring to the weights being publicly available to download, not for the model to actually match the official definition.