There’s an uncomfortable version of frontier-lab career advice that nobody really wants to publish, so here’s ours. For a meaningful fraction of senior engineers, the third year at a frontier lab is the year their career starts to quietly cap. Not because the lab is bad. Not because the engineer is bad. Because the structural fit between the engineer and the institution has reached its useful end, and neither party is incentivized to be the first to say so. This piece is about how to notice that, and what the engineers in our network who did the right thing with it actually did.
The data underneath this comes from roughly 280network members who left a frontier-lab senior IC role in the past three years. We tracked their next stop, their next-next stop, and what they say about the move 12 to 24 months out. Two patterns dominate. One is the engineers who left for the right reasons and ended up doing better work; the other is the engineers who left for the wrong reasons and ended up moving again within 18 months. The difference between the two groups isn’t where they went. It’s how they read the signal that it was time to leave.
// 01Why year three
The structural argument for year three goes like this. Year one is onboarding — you’re learning the lab’s stack, its rituals, the texture of how decisions get made. Year two is impact — you’ve built the trust to ship something the team takes seriously, and the comp band reflects the trajectory. Year three is when one of two things happens: either the lab’s next problem is a problem your specialization is genuinely well suited to, in which case you’re set up for an excellent year four, or the lab’s next problem isn’t, in which case you have to choose between forcing a fit (which compounds badly) and staying narrow (which caps your scope inside the lab).
Most engineers in our dataset didn’t see this choice coming. The lab kept paying well, the work kept being interesting, the team kept respecting them. The signal was subtler — and that’s what this piece is really about.
// 02The four signs
These aren’t the obvious signs. The obvious signs — being unhappy, being passed over for promotion, hating the work — are well understood and reliably acted on. The four below are the ones our network members consistently say they ignored too long.
You start defending the lab's choices instead of arguing about them.
In year one and two, you argue about technical decisions in your team’s planning meetings. You’re skeptical of the lab’s roadmap. You push back. Sometime in year three, you notice you’ve stopped. You’re now the person explainingto the new hire why the lab is doing X instead of Y. You’re not lying — you’ve heard the reasoning so many times it’s become your reasoning. But you’ve also stopped updating the reasoning against new evidence, because being the person who keeps re-litigating is exhausting and unprofessional, and the path of least resistance is to settle.
The engineers in our network who report this sign — and act on it — almost always say later that they hadn’t realized how much intellectual range they’d quietly given up. Defending is not the same as agreeing. If you can’t remember the last time you genuinely changed your mind about something inside your lab’s worldview, that’s the sign.
The comp doesn't feel like motivation anymore. It feels like the reason.
Senior IC compensation at frontier labs is real money. Most engineers will hit a level of comp here that feels meaningful even relative to the next-best alternative. But there’s a specific failure mode where the comp transitions, around year three, from being part of why you’re doing the work to being the mainreason you haven’t left yet. You can usually tell because your justification for staying has flattened to a single dimension. “The equity vests in 14 months.” “If I make staff I’ll be set up for the next move.” Etcetera.
The cleanest test we know: imagine a counterfactual lab paying you 90% of current comp, doing roughly the same kind of work, but where the next problem is genuinely yours. Would you go? If yes, your current job is mostly a finance position with engineering as a hobby. That’s an okay choice — many of our network members make it, eyes open — but it’s a choice worth making knowingly, not by default.
You're being asked, repeatedly, to do the thing you're already good at.
A specific year-three pattern in our data: engineers who are very good at one thing get asked to do that one thing again and again. The lab has identified you as the reward-modeling person, or the inference-loop person, or the eval person, and every new project that touches your area gets routed to you. This feels like respect, because it is. It also caps your growth, because the next layer of capability requires you to be visibly bad at a new thing for a year — and the lab has built around you being good at the current thing.
The senior engineers in our network who did better after leaving almost universally used the move to deliberately broaden. Not necessarily a different specialization — but a different relationship to it. The reward-modeling person becomes the alignment-research lead at a smaller lab; the inference engineer becomes the head of platform at an AI-native startup; the eval person becomes the founding research-engineering hire at a Series B.
The internal conversations have stopped surprising you.
Walk into the team’s weekly. Listen to what people propose. Predict, before each person speaks, what they’re going to say. By year three you’ll often be right about most of the room most of the time — and that’s both a sign that you’ve earned a real model of your colleagues, and a sign that the room has stopped teaching you. If you’re predicting the takes correctly because the takes have become predictable, that’s information about the lab’s static phase. If you’re predicting them correctly because the team has converged and there’s no new pull from the outside, that’s a different signal about the team’s intellectual liveness.
Hard to disentangle which it is, but our network members who left for the right reasons all describe a moment where they sat in a meeting and realized they hadn’t been surprised by a colleague in months. “Surprise frequency” turns out to be one of the most predictive subjective metrics we’ve found, in our debriefs with engineers who left.
None of these signs is decisive on its own. Most engineers we talk to have one of them at any given time. The pattern that matters is when three or four show up together in the same six-month window — that’s the moment our data says the move pays off.// FROM THE Q4 2025 NETWORK MEMBER DEBRIEFS
// 03What good next moves looked like
The 280 network members in our dataset who left a frontier lab fall into a few destination patterns. We’ve grouped them by where they ended up 12 months later. Comp profile, work surface, and broader-vs-narrower trajectory all vary across the groups.
Smaller frontier lab, expanded scope
Move to a Mistral, a Cohere, a Reka, a Sarvam. Comp drops 10–25%. Scope expands materially — the move is buying breadth.
AI-native startup, founding research-eng
Series B or C AI-native; founding or near-founding research-engineering role. Comp is wider on equity, lower on cash. Real ownership.
Different specialization, same lab tier
Move to a different frontier lab where they can do agent work instead of post-training, or vice versa. Comp is roughly flat.
Founding their own thing
Started a company, often with one or two colleagues from the lab. Comp goes to zero for a year or two; equity bet.
The remaining 8% went to non-AI roles or paused — sabbaticals, longer-form research grants, family-related moves. Useful context but not the focus here.
The common thread across patterns 01–03 isn’t where the engineer landed. It’s the framing they used to make the move. The engineers who did well almost universally led the decision with a positive ask — “I want to do X” — rather than with a negative — “I’m done with Y”. The engineers who moved on negative-only framing ended up moving again within 18 months in our dataset roughly twice as often.
// 04When not to leave
Not every plateau is real. We’ve seen our share of engineers leave for the wrong reasons and regret it within a year. Some honest counter-balance.
The structural signs are there.
- You can name a different problem you want to be working on, with specificity.
- Three or more of the four signs above ring true for a sustained period.
- You have at least one concrete next-step option you’ve vetted seriously.
- The comp delta of leaving is one you can live with for the upside you’re chasing.
The plateau is local, not structural.
- You’re bored on a specific project, not bored with the lab’s whole trajectory.
- A scope expansion would be granted if you asked — and you haven’t asked.
- You’re within 6 months of a meaningful equity cliff and the move is small.
- Your “negative” reasons outnumber your “positive” ones three to one.
The most common mistake we see, finally: leaving without asking for what you actually want. About a third of the engineers in our dataset who left a frontier lab admitted, in debrief, that they hadn’t directly asked their manager for the scope expansion or team change that would have addressed their actual frustration. The conversation is uncomfortable and sometimes performative, but it’s surprisingly cheap relative to the cost of leaving the wrong job. If you haven’t had the conversation, have it before you start interviewing.
Next week’s piece will be on India frontier comp — the 34% YoY lift, the three labs behind most of it, and what it means if you’re considering relocating or going back. Subscribe at the bottom of the blog page if you want it.
// Read next
Cohire Copilot
The map plots you against the AI engineering landscape. Helps with the question this post raises.
READ →// FLAGSHIPWhat the May 2026 frontier-hiring data actually says
5,800 senior-IC offers. Three numbers worth knowing if you’re planning a move.
READ →// FEATURECohire
The agent that runs your next job search end-to-end. Designed for the engineer-with-a-day-job.
READ →If you’re reading this and recognising the signs.
Apply to the network. Your next move starts with the matches we’ll surface for you privately, before any lab knows you’re looking.