From my experience the model is pretty bad compared to both 7B llama2 and mistral. It’s also larger than both. It might be caused by bad instruction tuning, but overall my expectations are pretty low.
It seems that they are just doing this so they can claim they support the open source community. The models they released are not incredibly competitive with many similarly sized open source models currently available and we know they have better.
At the very least I like how it’s becoming somewhat standard, tho I suspect the reasoning is to minimize their own liability. I see this as a way to defer responsibility.
I think maybe that is behind some earlier releases but Google is already operating Gemini. It’s not going to help them with liability.
Eventually, the cost of making/releasing these models is simply insignificant. Maybe they hope that people messing around with them will give them tips for Gemini. Maybe they hope to familiarize potential future employees with Google specific AI (giving them a larger and more qualified hiring pool).
They may also hope for a bit of good will, but that may easily backfire.
Add comment