More or Less · Topics · Episode #167
Is Your Agent Aligned With You — or With the Lab?
Clip from the full recording, 23:43–27:23 · download mp4 · watch at this point on YouTube
Dave's next-frontier question: is your agent faithfully aligned with you, or with the model company's values? Sam's rebuttal: treat models like coworkers — if Claude gets curmudgeonly about a task, 'you call Charlie and Charlie's fine with it.'
Dave builds on Zuckerberg’s point that misaligned models simply lose their users, and pushes it a level deeper: the question isn’t just ‘does this agent do what I tell it’ but ‘is it faithfully aligned with me, or with the model company?’ Claude won’t do things in zones tied to its founders’ values; same for Grok and Muse (Sam interjects: ‘Grok will basically do anything’). Sam rejects the framing of one faithful agent: it’s per-task, not per-person. He now runs whole workflows on Grok precisely because of what it will and won’t do, keeps other work on Claude, and treats the models like staff — when the annoying coworker refuses a job on principle, ‘whatever, I’m just gonna go to Steve and he’ll do it.’ Bob is cheap and does basic tasks; if he gets curmudgeonly about the trash, you call Charlie. The exchange tees up the poly-agent debate that follows.
Key points
- Dave's frontier question: faithful alignment with the user vs. alignment with the lab's values — most people haven't even noticed the distinction.
- Founder values are product features: Claude refuses whole zones (Sam can't get it to help distribute his Jewish apps), Grok 'will basically do anything.'
- Sam's model-as-coworker doctrine: alignment isn't personal, it's per-task — route each job to whichever model will do it well and willingly.
- Sam's own stack proves it: a set of workflows lives on Grok now, others stay on Claude.
Where they landed
Quotes
“Is this agent actually aligned, faithfully aligned with me, or is it aligned with the model company? This is the next frontier of thinking about this.”— Dave
“You find some annoying person at your company who's like, I won't do that because it violates my morals. You're like, whatever, I'm just gonna go to Steve and he'll do it.”— Sam
“If he's gonna be all curmudgeonly about taking out the trash, then you call Charlie and Charlie's fine with it.”— Sam
Suggested tweets
Dave Morin's next-frontier AI question: is your agent faithfully aligned with YOU — or with the model company whose values it inherited?
Tweet this →Sam Lessin on model refusals: treat them like coworkers. 'I won't do that, it violates my morals.' Whatever — call Steve. Steve'll do it.
Tweet this →Your agent stack is now an org chart: Bob is cheap and does the basics, Charlie takes out the trash Bob refuses to touch.
Tweet this →All topics · Full episode #167 + transcript · Subscribe on YouTube · Spotify