
Blog
I have a strange new pandemic ritual. Every Monday, I gather my team of machine learning engineers in a video meeting to kick off the week. But before we get down to business, I use a giant neural network to generate questions that help us go deeper than small talk. We vote on our favorite question and then take turns answering it.
Those are actual questions the model generated. They indeed lead to deep talk.
We’ve been working remotely at Primer since the earliest days of the pandemic. This ritual is one of our solutions to the problem of feeling disconnected to each other. I can tell you: It really works.
Here’s how I do it.
First, I create a prompt consisting of numbered questions. For example:
You can find lists of such questions all over the internet, probably inspired originally by this 2015 New York Times article, 36 questions for a private conversation starter. I find that a list of at least 20 questions works best.
Then I feed this prompt to a big autoregressive language model. I use GPT-J-6B, the open source model created by Eleuther. I just copy-paste the prompt and run the model in Primer’s developer interface. The model’s inference takes about 5 seconds.
What comes back is machine-written text that continues where mine left off:
4. What thing would you never let anyone see you do?
5. What color are your dreams?
6. If you could have one superpower, what would it be?
Yes, these neural networks know how to count. My team uses the numbers to vote for their favorite.
I admit it may seem bizarre to use a machine to help people move beyond small talk and connect. But this is artificial intelligence at its best: Working together with humans to help them do what they do best.
I generated 365 Deep Talk questions so you can try it yourself.