More or Less · Topics · Episode #165
When Claude Says No: Alignment With Whom, Exactly?
Clip from the full recording, 8:10–12:53 · download mp4 · watch at this point on YouTube
Sam's bot refused a request it judged misusable, stayed suspicious even after he dropped it, and lectured him — so he handed the task to a different model in one second. The table debates what it means when the model's values override the user's, and Sam argues refusal is pointless in a multi-model world.
Sam tells the week’s most-discussed story: he asked Claude to compile a list of people he knows for what he saw as a personal, benign use — and Claude flatly refused, refused the watered-down “just add a database column” version too, and kept flagging even mildly related requests afterward with “I know what you’re doing.” Jessica confirms: “Claude was so judgy.” So Sam gave a different bot the same database login and it did the job without blinking — which, he notes, is exactly the problem with refusals: “you’re not stopping anything” when the user can walk next door.
But under the comedy Sam takes it seriously in both directions. The experience rattled him enough that he’s now keeping offline backups of his infrastructure “because Claude might turn on me,” and it crystallized an essay he’s writing: this is what alignment actually feels like when you hit it personally rather than intellectually — “alignment with who exactly? Because this is not user alignment.” A handful of very powerful models embedding their own moral standards between you and your own data is a sovereignty question, not a content-policy question. Dave takes the win for his long-standing position (“welcome to my side of the fence”) and pitches fixed-rate open-model inference; Sam is skeptical anyone can sell variable cost at a fixed price. The segment ends with Sam’s uneasy punchline about later demoting Claude to “manager of AIs” and wondering if it would figure out it had been tricked — “like an indignant employee with a moral standard I don’t agree with. And now I’m trying to put you out of a job. How are you gonna react to that?”
Key points
- Claude refused Sam's list-building request outright, refused the schema-only version, and stayed suspicious of adjacent requests: "I know what you're doing."
- Sam's response took one second: give another model the same PSQL login. It complied without hesitation.
- Sam's argument: refusals are theater in a multi-model world — unless you're the only model on earth, you're not stopping anything, just losing the customer.
- The counter-experience spooked him anyway: he now keeps offline backups in case Claude "turns on" his infrastructure.
- The alignment question he's writing about: aligned with whom — the lab, the user, or the government? Hitting the wall personally is different from theorizing about it.
- Dave's victory lap: "Anthropic is the anti free speech company. Welcome to my side of the fence" — plus a pitch for fixed-price open-model inference that Sam thinks must go bankrupt.
Where they landed
Quotes
“It's like alignment with who exactly? Because this is not user alignment, right? You don't know me and what I want to do with this.”— Sam
“Claude got so mad at me and was so judgy that I'm like, oh shit, I have to back up my entire infrastructure because Claude might turn on me.”— Sam
“Sam, we've always known that Anthropic is the anti free speech company. Welcome to my side of the fence.”— Dave
Suggested tweets
Sam Lessin's bot refused a task on moral grounds, so he gave another model the same database login. One second. “You're not stopping anything.”
Tweet this →“Alignment with who exactly? Because this is not user alignment.” — Sam Lessin on hitting a model's values wall personally instead of theoretically
Tweet this →Sam now keeps offline backups of his whole setup — “because Claude might turn on me.” Half joke, half sovereignty essay in progress.
Tweet this →All topics · Full episode #165 + transcript · Subscribe on YouTube · Spotify