Reid Hoffman: There’s a lot of things that go into the training of these models beyond simply providing data. These models are intensive learning machines. And in this phase, called post-training with human-feedback learning, they go through a set of drills and give two answers, and the human judge says, “This one is better.”

So in the case of Inflection’s Pi, one of the things that the Inflection team came up with was to train emotional intelligence [EQ] as intensely as we train IQ. So when we’re doing human-feedback learning, we teach Pi to give an answer with EQ.

One simple example might be if you asked both Pi and ChatGPT how much you console a friend after the loss of a treasured family pet. They would both give the same sort of answers, but ChatGPT might say, “And here are the five things you could do.”

Pi, one the other hand, might say, “You know your friend. What would count as being there for your friend? Might it be just expressing the sympathy that you feel with them in their moment of anguish? Or something else?” And then it would help you walk through that, even though both models know the five possible actions.

Lareina Yee: This is pretty extraordinary, because you’re basically saying we have the IQ, but we can also train EQ. That could include empathy, active listening, how to stay positive, and how to listen to feedback. Tell us a little bit about how you see those two attributes in AI solutions, and the connection to our experiences.

Reid Hoffman: One thing that’s certain in the next five to ten years is that we are going to have agents everywhere, doing all kinds of things for us through all kinds of interactions. There will be agents for groups, agents for companies, and many other types. And the most natural thing when you’re doing engineering is to get IQ correct. But one of the things that’s really essential for people is how we bring EQ into it.

At the Edge podcast excerpt