What slightly unusual beliefs do you hold? How do you defend them?

[INTERVIEWER] What slightly unusual beliefs do you hold? How do you defend them? "What slightly unusual beliefs do you hold, and how do you defend them?" Here's the misread that kills people. They think it's a test of how contrarian they can sound, so they reach for something spicy. It's not that. It's a test of independent thinking backed by evidence.

Pick a real belief you hold about building AI products, and defend it with data and experience. Edgy for its own sake fails hard. A defensible, slightly against consensus take that you can actually support, that's what wins. The core thing being probed here is whether you can reason from evidence to a view, instead of chasing hype or benchmark scores.

AI product work is full of confident, wrong consensus, and they want to know you can stand slightly outside it without falling into pure contrarianism. In the next few minutes I'll give you the four moves that turn a belief into a defensible position, a full worked example, and the follow up questions they'll press you with. Four moves, in order.

Move one: state the belief plainly. About thirty seconds. One clear sentence. Pick something genuinely a little against the grain in AI product work, not a safe platitude dressed up as an opinion. Good territory: evaluation first development, scepticism of benchmark scores, small models beating large ones for most product jobs, or that most AI features should ship as assistants rather than autonomous agents.

Move two: say why the consensus differs. About a minute. Show you understand the mainstream view and you're not just ignorant of it. Steelman the other side, briefly and fairly. This is the exact thing that separates independent thinking from contrarian noise. If you can't argue the other side, you don't really own your own. Move three: defend with evidence.

About three minutes, and this is the meat. Your reasons, backed by something concrete. A result you actually saw, a study, a pattern across projects. Not one anecdote dressed up as a law of nature. If you've got a number, use it. And if you've changed your mind on this before, say what evidence moved you, because that's the sound of someone who reasons rather than recites.

Move four: concede the limits. About a minute. Where does the belief not hold, and how would you act differently there? Certainty with no boundary reads as a zealot, not a thinker. Holding the belief lightly but defending it firmly is the exact texture they're listening for. Here's the whole thing joined up. "A belief I hold that's a little against the grain: for most product features, a properly evaluated small or medium size model beats reaching for the largest frontier model, and teams default to the big model far too fast.

The consensus, and I get it, is that the strongest model gives the best experience so you should just start there. I genuinely understand why. It usually does score highest on the benchmarks, and nobody ever got fired for picking the biggest model. It's the safe call. But my reasons come from actually shipping. On three separate features, once I built a real golden set and measured on the actual task, a medium size model with good retrieval and a tight prompt matched the frontier model within a couple of points of accuracy.

At roughly a fifth of the cost, and half the latency. And here's the bit that mattered: latency moved adoption more than the last two points of quality ever did. Faster answers got used more. The pattern I keep seeing is that teams pick the model by benchmark reputation, not by measuring on their own task, so they overpay for capability the task never actually uses.

Now, what would change my mind, and does, on specific features: genuinely hard reasoning, long context synthesis, or open ended creative generation. There, I've measured the big model pull clearly ahead, and I pay for it happily. So the belief is really 'measure on your own task before you pick.' The default should be the smaller model until the eval says otherwise, not the other way round." See how that lands?

It's a real position, the other side is steelmanned, there's a number, and there's a clear boundary where he flips. That's a thinker, not a contrarian. If you want a second belief in your pocket, here's another defensible one: most AI features should ship as assistants, not autonomous agents. The consensus right now is racing towards full autonomy. The steelman: autonomy removes friction and does the whole job.

The evidence: on the tasks I've shipped, a human confirmation design kept trust high while an automatic action design generated a wrong action the first week and cost us a customer. The boundary: for cheap, reversible, high volume actions, let it run autonomously. Same shape, four moves. Here's what makes them lean in. First, a real, slightly against consensus belief, stated plainly, with the mainstream view steelmanned rather than turned into a strawman.

That fairness is the signal of a genuine independent thinker. Second, evidence from repeated experience plus a number, not one story stretched into a universal rule. And third, a clear statement of what would change your mind. That single move, naming your own falsification condition, is the strongest thing you can do in this answer, because it proves the belief came from reasoning and not from ego.

Now the ways people sink this. The first is contrarian theatre, a spicy take with no evidence, picked purely to sound edgy. They'll ask for your reasons and there'll be nothing there. The second trap is the opposite, a safe bland statement everyone already agrees with, which dodges the question entirely and reads as playing it safe. And the third is total certainty with no boundary, which comes across as dogma, not thought.

Expect them to push. They'll steelman the other side back at you, hard, to see if you crumble or engage. Don't dig in and don't fold. Engage with the specific point and, where they've got you, concede it cleanly. "That's fair, and that's exactly the case where I'd pick the big model." Conceding the right point at the right moment strengthens you here, it doesn't weaken you.

So the whole picture: state one belief plainly, steelman the consensus, defend it with a real result and a number, then name the boundary where it stops holding and what would change your mind. Four moves. The evidence and the boundary are what turn a hot take into a position worth respecting. The single line to carry in is to pick a belief you can defend with data and a number, steelman the other side, and name what would change your mind.

Your goal is to show evidence, not just edge.