It's just a matter of time before a model of this size can be run on commodity hardware and somebody will take the brakes off and/or attempt to run experiments that aren't just "can this thing pass the turing test?". I'd be really interested to know the thoughts of the team, given their expert knowledge and experience with the matter.

I have a few questions: Could this be used as a tool to get a feel for public sentiment? For example, could you ask the bot what it thinks about gun control and have it spit out a policy that appeals to the common public? If you ask the bot what it thinks about how a company will perform, how accurately does it predict? I know that the model will contain the biases of the data set, but I'm curious if you've run these types of experiments. What do you think the results would be if you had an even bigger, more diverse corpus? (devil's advocate, for the sake of discussion: perhaps everyone's fb messenger and WhatsApp chat history)

Finally, you have clearly gone to great lengths to make the bot pleasant to interact with. What sort of results to you get when you train such a huge model on an uncurated corpus and don't try to tweak its personality? I find myself wishing that you didn't try to do this as the bot seems to be hyper-agreeable. I. E Too many responses like "You like watching paint dry? That's super interesting! I love watching paint dry!".

Imagine if you had an employment score. A payscale score. A threat-to-society score. Imagine if the systems that generate these scores are all connected together and can share information with each other. Imagine that these systems are fed information by a growing network of sensors - facial recognition, voice recognition, location tracking. Imagine all the ways that it could affect the way you behave and how such a system would try to optimise away dissidents by denying them services. This is one way that we can go from multiplication to becoming enslaved (or as good as killed) by algorithms. I'm sure there are other plausible paths.