First, a road trip.
The session starts smaller: a journey from San Francisco through Carmel to LA, and the difference between the next reward and the long-term one.
Start with the teaching example ↗
What counts
as helpful?
Product Lead, AI · Poe / Quora
A model can sound helpful without being right. I’m interested in how we turn that uncomfortable gap into something we can test.
Three signals.
No perfect judge.
A sketch of the reward-function experiment—not a scorecard or a validated measure of safety.
Helpfulness
The demo uses sentiment as a proxy.
Safety
A toxicity classifier supplies another signal.
Truthfulness
A fake-news classifier adds an imperfect third view.
Normalize + weight the signals→ Combined reward
Adapted from the livestream. These proxy choices are part of the experiment, not proof of the qualities they stand for.
Then, real tasks.
My agent-evaluation session moves the question into live environments: what can an agent actually get done?
Evals for AI Agents ↗