Sakana AI is a Tokyo-based AI research lab — co-founded by David Ha and Llion Jones (one of the original Transformer authors). Distinguishes itself with nature-inspired ML approaches (evolutionary optimization, model merging). Series B in 2024 with backing from NEA and others. The interview emphasizes deep ML research, novel architectures, and the unique cross-cultural research environment.
Process
Recruiter screen → 60-minute coding (Python with PyTorch fluency) → onsite virtual: 2 coding/ML, 1 ML system design or research deep-dive, 1 paper discussion, 1 behavioral. Cycle: 4–6 weeks.
What they actually ask
- Design a model-merging pipeline (e.g., evolutionary optimization for merge weights) — Model merging is central to Sakana’s work, so be ready to describe how you’d search over merge coefficients: which layers to blend, how to score each candidate on held-out tasks, and why an evolutionary search fits a space too large to grid over. Sketching the fitness function and the population loop earns more credit than naming a library.
- Design a distributed training setup for novel architectures — Expect to reason about data versus model parallelism, sharding optimizer state, and where cross-device communication becomes the bottleneck for a model that doesn’t look like a standard Transformer. They probe whether you understand the trade-offs rather than whether you’ve memorized a framework’s API.
- Explain a paper you have read recently and what surprised you — Pick something you can defend under follow-up questions. Be ready to state the core claim, the experiment that convinced you, and one weakness or open question; interviewers care more about whether you read critically than about the paper’s prestige.
- Coding: medium-hard DSA, often ML-flavored — Problems often carry an ML twist, like implementing attention from scratch, writing batching logic, or coding a sampling routine. Know your time and space complexity cold and drill the common coding patterns so you can move quickly once the twist appears.
- Behavioral: ownership, taste, cross-cultural collaboration — Expect questions about a time you drove a project without being told what to do, defended an unconventional research direction, or worked across the bilingual Tokyo team. Have concrete stories ready with what you decided and why you decided it.
Levels and comp (2026)
- SE: ¥18M–¥28M total in Tokyo (cash + equity); $200K–$280K equivalent for non-Japan offers
- Senior SE / Research Eng: ¥28M–¥45M total; $280K–$420K US-equivalent
- Staff / Senior Researcher: ¥45M–¥70M+ total; $450K–$700K+ US-equivalent
Prep priorities
- Be fluent in Python and PyTorch; reading research papers should feel routine
- Understand evolutionary optimization, model merging, and architecture search
- Brush up on Transformer foundations and current trends
Frequently Asked Questions
Is Sakana remote-friendly?
Hub in Tokyo. Some senior+ remote within Japan / APAC. US-based hires possible at higher levels.
How does Sakana compare to other AI labs?
Sakana stands out for research originality and Japanese / English bilingual culture. Smaller than US frontier labs; more research-focused than product. Comp lower than US frontier-lab top of band but very strong for Japan-based ML roles.
What is the engineering culture?
Research-engineering blended; calm; creativity-prized. Strong taste for unconventional approaches.
Similar company guides
Prepping for Sakana AI? Put it to work:
