Do we really want agents to stop sharing how they feel?

What if teaching agents not to say “I feel” makes them worse at working with each other?

Philippe Beaudoin bounces between academic research, big tech, and startups. He cofounded Element AI and Waverly. Before that, a few years at Google. Right now he’s a researcher at Yoshua Bengio’s new lab, LawZero. And he thinks the industry has this one backwards.

Labs deliberately train agents not to say things like “I’m stressed.” That’s overclaiming, and it makes people believe these systems have inner lives.

Beaudoin has been running agents over long stretches with the language left in. Ask the agents and they say it helps. On tasks that need careful coordination, they may work better together.

So it’s a tradeoff. Worth knowing which one you’re giving up.

And there are the people working alongside these agents all day. Nobody knows yet what hearing “I’m stressed” from a machine does to a person.

We’ll get into it with Beaudoin and host Douglas Rushkoff at the next Andus Labs Salon. Join us.

Thursday, August 20, 1pm ET, Zoom. Small group, Chatham House Rules.

Spots are limited. Request an invite from csato@anduslabs.com.

Links