They at least give access through paid API to their latest models, which hasn't been the case for Imagen, Parti, LaMDA (google), Flamingo, AlphaGo, AlphaZero, AlphaStar, AlphaCode, AlphaTensor (DeepMind), Make-a-scene, Make-a-video (Meta)

I don't think DM was open in any meaningful way other than publishing research.

It took a long time until their source code became public, and replication of certain papers is problematic to say the least.

Similar issues exist with some G.Brain papers not releasing code publicly years after acceptance.

But all of these are anecdata from a random person on HN so take with a grain of salt.

Yes, we're all aware it's common practice to open source code.

However, if the weights aren't provided and the data is unavailable or the compute is expensive, the code is worthless.