To my knowledge, you can't talk to it. It looks like so far they've released sample conversations with Meena[0], and not much else. They've given a rough description of the model architecture and how they trained it, but good luck trying to replicate their 2.6B parameter model unless you can afford a lot of computes.
The burden of proof lies with Google, not with the scientific community. I cannot reproduce a scientific model if the data used for that model is squirreled away. Google has claimed so many times that they have cracked the nut on AI but has never allowed anyone to see real evidence. It makes me wonder if this is just a way to pump their stock.
[0]: https://github.com/google-research/google-research/tree/mast...