OpenTalent
Hire AI TalentFor EmployeesTop 3%Jobs
Sign inJoin freeEmployer Login
Join free
// BLOG · SCREENING
// Screening

The one question we ask every senior IC panel.

It’s not the hardest question. It’s the one whose answer tells us the most about who you’d be on month four. Three years of running this loop has taught us why.

By The Panel·Published 10 May 2026·9 min read

Every senior IC panel we run for the OpenTalent network ends with the same question. It isn't the technical opener. It isn't the systems-design closer. It's a question that takes about eight minutes of conversation to land properly — and over three years of running this loop, it has predicted the engineer's first-year performance better than any other single thing we ask. Including the take-home, including the live coding, including references.

This isn't a clever observation we want to be coy about. Here it is.

// THE QUESTION

“Tell me about something significant you've changed your mind about in your specialty — what you used to think, what changed it, and what you think now.”

That's the whole prompt. The candidate gets to pick the topic. We let the silence sit if they need it. Then we follow up — gently, but precisely — on three things: the old view (did you actually hold it, or are you reverse-engineering a clean story?), the trigger (what specifically moved you?), and the new view (is it actually different, or have you just rebranded the old one?).

This piece is about why we ask it, what good and bad answers sound like, and the pattern we noticed at month four that locked it into our rubric for good.

// 01Why this question

For a long time, our final-panel question was the kind of thing you'd guess: deep technical depth probes, behavioural questions about conflict, the “describe your hardest debugging session” classic. They all gave us signal, but the signal had a problem — it correlated heavily with what we'd already learned in the earlier stages. Stage 02 (the deep skill review) and Stage 04 (the live panel) already told us whether someone could go deep technically. Stage 03 (the 13-day project) already told us whether they could ship. Asking another version of those questions at the end was redundant.

What we didn't have good signal on was something subtler: how does this person operate when reality doesn't match their model? Senior engineering work is mostly that. The training run breaks in a way the design didn't predict. The eval surfaces a regression that contradicts the team's working theory. The architecture you've been defending for a year turns out to be wrong in ways you didn't see. The question of how an engineer metabolizes that— the moment when their existing view of the world stops fitting the evidence — is one of the most important things to know about someone, and we weren't probing it directly.

The “changed your mind” question turned out to be a near-perfect probe for it. To answer well, you have to:

  • Have actually held a real view in the first place — which most surface-level engineers don't.
  • Have noticed when reality stopped fitting it — which requires honest self-observation.
  • Have processed the new evidence rather than rejected or rationalized it — which is the actual skill.
  • Be able to articulate all three steps without making yourself the hero of the story.

That last bullet is harder than it sounds. Listen for it.

// 02What good answers sound like

The best answers we've heard share three structural features. They name a specific, technical, sub-area-level view (not “I used to think AI was hype”). They identify a specific trigger (a paper, a failed run, a piece of feedback from a respected colleague, a debug session that didn't end where it was supposed to). And they describe the new view in a way that's actually different from the old one in a way the candidate seems to find interesting, not just professionally convenient.

// GOOD ANSWER

“I used to think reward model size was basically a free parameter — make it bigger, get better preferences, ship. Then I ran a sweep where the 13B reward model was actually worse than the 7B on three of our four held-out evals, and I couldn't explain it. I spent a weekend reading and basically convinced myself that reward model capacity has a sweet spot relative to the policy you're training, and overshooting actively hurts. I think differently about reward-model sizing now.”

// BAD ANSWER

“I used to think AI safety was less important than capability work, but then I read more about alignment and realized it's all connected. Now I try to think about both at the same time.”

What makes the first one good isn't that we agree with the conclusion — we have no opinion on the conclusion. What makes it good is the structure. There's a concrete old view (“size is a free parameter”), a concrete trigger (“13B sweep was worse”), and a concrete new view (“capacity has a sweet spot relative to the policy”). All three are testable. All three are at the right level of resolution — sub-field level, not field level.

The bad answer fails on every dimension. “AI safety was less important” is not a held view, it's a vague stance. “Read more” is not a trigger. “Try to think about both” is not a new view. The candidate is performing the question rather than answering it.

// What we listen for

  • Resolution.Sub-area, not field. “RLHF reward model sizing,” not “post-training.”
  • A specific trigger.A paper, a run, a colleague, a debug session — not “I matured.”
  • Symmetry. The old and new views should be at the same level of resolution. If the old view is vague and the new view is precise, the candidate has probably retrofitted the answer.
  • Curiosity in the voice. The candidate should sound like they enjoyed being wrong. Not defensive, not embarrassed. Interested.
  • Self-effacement that's earned. They should be modest about the old view without being theatrical about it. “I was stuck on this for longer than I should have been” is good; “I was such an idiot” is performance.

// 03What bad answers sound like

There are three failure modes we see often enough to name. They're not disqualifying on their own — sometimes a strong candidate just doesn't have a good story prepared — but if a candidate hits two of them, the panel weight on this question goes way up in the negative direction.

The first failure is the rebrand.The candidate's “old view” and “new view” are the same idea wearing different vocabulary. “I used to think reproducibility was about controlling random seeds; now I think it's about controlling the entire experimental setup.” That's not a changed mind. That's a more articulate version of the same view. We follow up here directly: “What's a concrete decision you'd make differently now than you would have under the old view?” If they can't answer, the rebrand is the whole story.

The second failure is the consensus laundering. The candidate adopts a popular industry view and packages it as a personal evolution. “I used to think small models were enough for most problems; now I think frontier scale is essential.” This is sometimes true — but the trigger is doing all the work, and the trigger is suspicious. “I just kept seeing the scaling laws results” is not a personal trigger; it's industry consensus arriving at a normal pace. We follow up: “What's something youran or saw that convinced you, not what the field said?” If they can't produce one, the candidate is borrowing convictions.

The third failure is the no-answer answer. “I can't think of anything significant I've changed my mind about.” We let this sit, then ask gently: “Take a minute.” If after a real minute there's still nothing — not even a small thing — that's almost always disqualifying, and it correlates strongly with later patterns we don't want.

The candidates whose first-year performance most pleasantly surprised us tended to share a specific failure mode in this question: they couldn't stop telling us about it. Eight minutes turned into fifteen. We had to redirect them. That's the right amount of enthusiasm for being wrong.
// FROM THE INTERNAL PANEL DEBRIEFS · Q4 2025

// 04The month-four pattern

About three years into running this question, our placement-side ops team started noticing something. The engineers whose answers we'd graded most positively on this question were also the engineers whose hiring-partner managers gave us the strongest feedback at the 90-day check-in. The correlation was strong — strong enough that we now use Q5 performance as one of the inputs to our renewal-of-membership review (Stage 05 of the screen).

We dug into why. The pattern is structural, not coincidental. Month four is when the honeymoon ends for a new senior IC at a frontier lab. The first month is onboarding and observation. The second is small wins, mostly running the playbook handed to you. The third is when the bigger work starts. The fourth is when you encounter your first real disagreement with how the team has been doing things — and what you do at that disagreement is what determines whether you become the kind of senior IC the team starts deferring to, or the kind who gets quietly worked around.

What you do at that month-four disagreement is the same thing you do every time reality fails to match your model. If you're an engineer who has a history of metabolizing that gap honestly, you do it again. If you're an engineer who has a history of rejecting or rationalizing, you do that again too. The “changed your mind” question is, in effect, a behavioral history check on that specific muscle.

It's not the only thing that matters. Plenty of engineers with strong Q5 answers underperform at month four because of mismatches we should have caught earlier in the screening. But across the cohort, the question's signal-to-noise has been strong enough that none of us on the panel side want to give it up.

// 05What this means for candidates reading this

You might be reading this and thinking: great, now you've told me the question, I'll just prepare a good answer. Two responses to that.

First: please do. We're not trying to trick anyone. We're trying to find engineers who think carefully about their work and can talk about it. If you've never asked yourself “what do I think now that I didn't think a year ago, and why,” that's worth doing for its own sake, regardless of our loop.

Second: prepared answers are easy to spot, and bad prepared answers are worse than honest hesitation. We're not grading the polish of the story. We're listening for whether the underlying experience actually happened and the reflection actually happened. Polished answers from candidates who haven't done the underlying work fail spectacularly — we follow up on the trigger, and the trigger has to be load-bearing under cross-examination.

The best preparation, for what it's worth, is to spend a few hours writing — privately — about three or four genuine changes of mind you've had in your specialty over the last two years. What did you used to think? When did you notice the gap? What do you think now? When you sit down to do this, you'll find that one of the four is the strongest, and that's the one to bring to the panel.

If you're considering applying to the network, the rest of the screening is described on the Top 3% page. The “changed your mind” question is the final 10 minutes of Stage 04. Everything before it is covered there. Everything after it depends on this one.

TP
The Panel
// EDITORIAL · OPENTALENT SCREENING TEAM

// Read next

// THE SCREEN
The Top 3% screening process

Five stages. Less than 3% acceptance. The full process explained.

READ →
// FLAGSHIP
What the May 2026 frontier-hiring data actually says

5,800 senior-IC offers across 52 companies. Three numbers moved fast enough to change your year.

READ →
// PREP
Interview guides

Prep playbooks calibrated to each stage of the OpenTalent screen.

READ →

Apply when you’re ready.

Less than 3% of applicants make it through the network. The ones who do see frontier-lab roles the broader market doesn’t.

Apply to the networkSee the bar
OpenTalent

The Cohire for AI engineers — and the hiring partner for the teams building frontier intelligence.

Features

  • AI Job Match
  • Resume AI
  • Application Autofill
  • Cohire

For engineers

  • Browse jobs
  • AI Research roles
  • ML Engineering roles
  • Applied AI roles
  • Early-career track
  • Salary data

Resources

  • Blog
  • Events
  • Interview guides
  • Frontier lab insights

Company

  • About
  • For employees
  • Careers
  • Partners
  • Contact
  • Privacy · Terms
© 2026 Gravity Engineering Services Pvt. Ltd. All rights reserved.hello@opentalent.in