On the other hand if the data isn't open you should probably use the term open weights not open source. They're so close.
We risk giving AI the same opportunity to grow in an open direction, and by our own hand. Massive own goal.
I thought it was thanks to a lot of software developers’ uncompensated labor. Silly me.
the word "people" is so striking here... teams and companies, corporations and governments.. how can the cast of characters be so completely missed. An extreme opposite to a far previous era where one person could only be their group member. Vocabulary has to evolve in deliberations.
Tangential, but I wonder how well an AI performs when trained on genuine human data, versus a synthetic data set of AI-generated texts.
If performance when trained on the synthetic data set is close to that when trained on the original human dataset – this could be a good way to "launder" the original training data and reduce any potential legal issues with it.
It makes sense as any bias in the model generated synthetic data will just get magnified as models are continuously trained on that biased data.