Make devices think like a frontier model.Frontier capability inside the compute and memory envelope of a consumer device - phone, laptop, wearable - is not a constraint. It's the most interesting research problem in applied AI today. You'll lead training for one of the model families that powers our on-device agents: pretraining recipe choices, post-training (SFT, RLHF, DPO, GRPO and whatever the next acronym ends up being), distillation, quantization, and the long tail of tricks that make a small model punch above its weight.
This is for the researcher who's tired of training models that go behind an API. You want your model on the device in your pocket, your mom's pocket, and a hundred million pockets you'll never meet.
🤩 Tasks you will own- One or more model capabilities end-to-end - from data mixture and training objective through eval and shipping into a production on-device runtime
- The experiment design and writeups that compound across the team - kill what doesn't move the metric, double down on what does
- A training workstream with a clear success metric and a checkpoint that ships
🤚 Areas where you will assist- Infra and product engineers, by turning research wins into shipped capabilities
- Partnerships, by telling them honestly what's possible at the next device refresh and what's not
- Other researchers, by reading their code and making theirs easier to read
Skills you'll be expected to teach- The training techniques that matter most for our regime - distillation from frontier teachers, MoE at small scale, speculative decoding, KV cache compression
- How to design experiments that move a number you actually care about
Skills you'll be expected to learn- What production model deployment looks like under hardware deadlines from OEM partners
- On-device tool use and agentic post-training at consumer scale
- The full stack from training run to phone
Timeline of successAfter 30 days - You've reproduced one of our recent training runs end-to-end. You've named the three highest-leverage research bets for the next quarter and have a take on which two to run.
After 60 days - You're leading a training workstream with a clear metric. You've shipped a checkpoint that beats the previous best on the eval that matters. People trust your read on what's working.
After 90 days - Your work has shipped into a partner build. You've made one non-obvious bet that paid off and one that didn't, and the team has learned from both. You're shaping the next training cycle.
CompensationCompetitive cash and meaningful equity. Top-tier relocation and immigration support. Permission to publish what's safe to publish. SF, in person.
How to applySend a link to your most interesting result - paper, blog, model card, GitHub - with one paragraph on why it matters. Plus your resume, Google Scholar, or LinkedIn. Every exceptional candidate hears back within 48 hours.
