been on the other side of a few ML interview loops this year as an interviewer, and also interviewed at three companies myself before taking my current role. want to give an honest picture of what the bar looks like for SWE-to-ML switchers in 2026 specifically, because a lot of the advice floating around is two years stale.
the bar went up, but not uniformly
companies doing NLP/LLM work are absolutely flooded with applicants who took one fine-tuning course and updated their LinkedIn. the signal-to-noise for those orgs is brutal right now. they've compensated by adding a hard ML fundamentals screen: expect questions on attention mechanisms, gradient flow, loss landscape intuition. if you can't answer those from first principles, you're screened out fast.
recsys and forecasting ML roles at non-AI-native companies (think: a marketplace, a fintech, a logistics co) are more accessible. they want ML depth but the emphasis is on shipping, not research chops. these are often better entry points for SWE switchers.
what actually helps the SWE-to-ML transition a shipped ML project with real data. kaggle is fine for learning, not great for signaling seniority. something you built that runs in prod, even at a small scale, means a lot more. coding interviews are still important. these are not replaced by ML theory screens. you still need to pass LC-style coding, usually medium difficulty. stats fundamentals. bias-variance, confidence intervals, A/B test design. less about calculus, more about whether you'll make basic errors in production decisions.
the leveling hit is real but negotiable
most SWE-to-ML switches involve going down a level. L5 SWE often comes in as L4 MLE. seen a few people resist this and get nowhere. seen others take the level, prove themselves in 18 months, and get promoted back. the second path actually works.
leveling depends a lot on how much ML surface area you've touched in your current SWE role. if you've worked closely with MLEs, integrated model outputs, done feature engineering work, that's transferable. if your ML experience is purely coursework, expect the level adjustment.
skip companies that ask you to do a multi-day take-home as the first filter. those are red flags about how they value your time generally.