# What Ilya Sutskever’s SSI actually screens for in interviews

Source: https://www.techinterview.org/post/3233476797/ssi-safe-superintelligence-interview/
Updated: 2026-07-21 · techinterview.org

Safe Superintelligence doesn't run a careers page the way you'd expect from a lab sitting on a reported $32 billion valuation. No ladder of leveled roles, no take-home you grind through a coding portal, no recruiter blasting your inbox about "impact." The company has said it will spend hours figuring out whether a candidate has good character before it weighs their publication count. If you want to work there, that inversion is the first thing to understand.

SSI was founded in June 2024 by Ilya Sutskever, formerly OpenAI's chief scientist, along with Daniel Gross and Daniel Levy. The pitch was one line: a straight-shot lab with a single goal and a single product, a safe superintelligence, and nothing to ship in between. Gross left in July 2025 to join Meta's superintelligence group, Sutskever took over as CEO, and Levy is president. The team is small and split between Palo Alto and Tel Aviv. None of that reads like a company optimizing for headcount.

## They're not hiring you to ship a feature

Most AI labs interview against a role: you'll own retrieval, or post-training, or the inference stack, and the questions map to that surface. SSI has no product surface to staff. There's no launch date pulling on the org, no revenue number that needs a team behind it. What they're actually deciding is whether they'd want you on a small, trusted group working a multi-year research bet where the payoff is years out and mostly invisible until it isn't.

That changes what the interview is for. It's less "can you close this ticket" and more "can you do real research under heavy ambiguity, and would we still trust your judgment eighteen months from now when an experiment has failed six times." Both halves matter. Plenty of strong engineers clear the first and quietly fail the second.

## The character screen isn't a formality

SSI has been unusually direct about this. It looks for extraordinary capability over credentials, and it says it gets excited about people who are into the work rather than the scene or the hype. Read that as a filter, not a slogan. In practice a "good character" read is trying to answer a few concrete questions about you.

Do you say "I don't know" when you don't, instead of bluffing? When you describe a past result, do the caveats survive contact with a follow-up, or does the story get shakier the more they push? If you found a flaw in your own paper after it was accepted, what did you do? For an alignment-first lab, the sharpest version is a pressure question: you're a week from a milestone and you notice something that looks like a safety problem but might be nothing, and chasing it costs the milestone. What do you do, and what did you actually do the last time something like that came up?

There's no clever answer that beats a truthful one here. People who try to game it tend to give the response they think a safety lab wants to hear, and that itself is the tell.

## What the research interview is really testing

Sutskever has framed the field in three phases: a research era from roughly 2012 to 2020, a scaling era from 2020 to 2025, and a return to research starting around 2026, where new ideas rather than more compute drive the next gains. He's described SSI's approach as a different mountain to climb, deliberately apart from the methods he used at OpenAI. If that's the bet, the last skill they want is someone who only knows how to make a known recipe bigger.

So the research conversation probes taste. Can you take a vague problem and name the actual crux? Can you propose an experiment that would falsify your own hypothesis rather than one built to confirm it? When the first three things fail, do you have a real theory of what to try next, or do you just twiddle hyperparameters? Expect to spend most of an hour on one of your own past projects, going deep enough that hand-waving stops working around the third layer of "why."

A few questions in the spirit of what gets asked:

- "Name a widely praised result from the last two years you think is overrated, and defend the position."

- "You have one week and eight GPUs. Design an experiment to tell whether a capability is genuinely emerging or just memorized from pretraining."

- "Where does the scaling story stop paying off, and what would you try instead?"

- "Walk me through a time your experiment disagreed with your intuition. Which one was wrong?"

## A realistic picture of the loop

SSI doesn't publish its interview process, and given how few people it hires, secondhand accounts are thin. The shape below is reconstructed from what the founders have said publicly and from how small, research-first labs typically run a loop. Treat it as the terrain, not a leaked rubric.

| Interview stage | Typical format | What it's actually reading for |
| --- | --- | --- |
| Intro conversation | 45 to 60 minutes with a researcher or founder | Motivation: are you here for the problem or for the valuation |
| Past-project walkthrough | 60-plus minutes presenting your own work | Technical depth, truthfulness about limitations, research taste |
| Open-problem discussion | Whiteboard or shared doc, no fixed answer | Reasoning under ambiguity and designing falsifiable tests |
| Research-engineering exercise (RE roles) | Speed up a training loop or debug a flaky eval | Correctness, profiling instinct, reproducibility |
| Character and references | Behavioral conversation plus deep back-channel | Integrity, treatment of collaborators, judgment under pressure |
| Leadership conversation | Time with Sutskever or Levy | Long-term trust and fit with a tiny team |

## The coding bar, if you're applying as a research engineer

Research scientist and research engineer are different jobs, and the engineering track carries a real implementation bar. It isn't leetcode theater. The work is turning a fuzzy research idea into an experiment that runs fast and gives a trustworthy answer, so the exercises look like the job: here's a training loop, make it three times faster without changing what it learns; this evaluation is flaky across runs, find out why; this loss curve has a kink at step 4,000, explain it. The failure modes they watch for are the ones that quietly corrupt research, a silent numerical bug, a data leak between train and eval, a "speedup" that changes the math.

If you can profile a PyTorch model, reason about where a distributed job is really spending its time, and tell a genuine bug from an unlucky random seed, you're in range. If your first instinct on a slow experiment is to request more GPUs before you've looked at a trace, that shows too.

## What sinks otherwise strong candidates

The clearest way to fail is to be there for the scene. Interest in the hype, the valuation, or the proximity to Sutskever's name is exactly what the character screen is built to catch, and interviewers at labs like this get very good at spotting the candidate performing enthusiasm for the mission. Overclaiming is the next one. A resume that inflates your role on a paper falls apart fast when someone who knows the area asks the second and third question.

Then there's the candidate who is genuinely excellent at scaling and little else. That was the right skill for 2020 to 2025. If SSI's bet is that the next jump comes from new ideas, an interview that shows you can only turn the compute dial up is a mismatch, however impressive the past numbers. Treating safety as a communications problem rather than a technical one reads badly too, at a lab that put the word in its name and built its whole thesis around it.

The people who tend to do well are the ones who'd be doing this research on nights and weekends regardless of where they worked. SSI is betting that a small group of them, left alone with a hard problem and no product to babysit, gets further than a large org dragging a roadmap. Whether that bet pays off is still an open question. What it wants from you in the room is evidence that you're one of those people, and that part is hard to fake across a full day of interviews.