This isn't a dilemma at all. If Facebook can't release data it trains on because it would compromise user privacy, it is already a significant privacy violation that should be a scandal, and if it would prompt some regulatory or legislative remedies against Facebook for them to release the data, it should do the same for releasing the trained model, even through an API. The only reason people don't think about it this way is that public awareness of how these technologies work isn't pervasive enough for the general public to think it through, and it's hard to prove definitively. Basically, if this is Facebook's position, it's saying that the release of the model already constitutes a violation of user privacy, but they're betting no one will catch them

If the company wants to help research, it should full-throatedly endorse the position that it doesn't consider it a violation of privacy to train on the data it does, and release it so that it can be useful for research. If the company thinks it's safeguarding user privacy, it shouldn't be training models on data it considers private and then using them in public-facing ways at all

As it stands, Facebook seems to take the position that it wants to help the development of software built on models like Llama, but not really the fundamental research that goes into building those models in the same way

> Release the data, and if it ends up causing a privacy scandal...

We can't prove that a model like llama will never produce a segment of its training data set verbatim.

Any potential privacy scandal is already in motion.

My cynical assumption is that Meta knows that competitors like OpenAI have PR-bombs in their trained model and therefore would never opensource the weights.

Generally speaking, no. An important part of a lot of benchmarks in ML research is generalization. What this means is that it's often a lot easier to get a machine learning model to memorize the test cases in a benchmark than it is to train it to perform a general capability the benchmark is trying to test for. For that reason, the dataset is important, as if it includes the benchmark test cases in some way, it invalidates the test

When AI research was still mostly academic, I'm sure a lot of people still cheated, but there was somewhat less incentive to, and norms like publishing datasets made it easier to verify claims made in research papers. In a world where people don't, and there's significant financial incentive to lie, I just kind of assume they're lying