About this role
To qualify you must work as an ML research engineer, research scientist or applied scientist, and have personally trained, fine-tuned or optimized models in PyTorch or JAX.
What you'll do
- Come up with ideas for hard machine learning engineering tasks ("plans") for a long-horizon AI research project with a leading AI lab
- Design tasks hard enough to stump frontier models: a strong task is solved in no more than 2 out of 8 attempts
- Lean toward inference, evaluation and systems work rather than heavy model training. About half of the tasks run on GPU and half are CPU-only
- Explain clearly why each problem is hard for a frontier model
- You mainly contribute ideas. You do not need to build or review the tasks yourself
Must-haves
- Current or recent role as an ML research engineer, research scientist or applied scientist
- Hands-on experience personally training, fine-tuning or optimizing models in PyTorch or JAX
- Deep knowledge of at least one of these areas:
- LLMs: pretraining, alignment, evals
- GPU inference and performance: vLLM, Triton, CUDA, profiling
- Reinforcement learning (RL)
- Scientific ML: weather and climate, physics, biology
- Able to explain why a problem is hard for a frontier model
Nice-to-haves
- Experience writing benchmark or evaluation questions (a big plus)
- PhD in ML, CS or physics, or an MSc plus strong research output
- Publications
- Prior our partner contracts with strong performance
Not a fit
- Software or full-stack engineers whose AI experience is mainly building LangChain or RAG applications
- Backgrounds limited to AI trainer, rater or annotator work
Location: United States.
