- Working with other alignment researchers to propose, run, analyse, and visualise experiments.
- Working on large-scale ML frameworks that train models in parallel across many machines.
- Implementing new models or optimisation techniques from research papers.
- Building large-scale datasets.
- Building internal tooling and infrastructure for model inference, visualisation, and interpretability.
- Implementing new models or techniques from research papers.
- Doing exploratory mechanistic interpretability research
You might be a good fit for this role if:
- You have a understanding of performance in HPC workloads, have worked with large GPU clusters, and ideally with some modern ML frameworks (e.g., PyTorch, Jax)
- You are proactive, driven, and creative. People with interesting past research and with github profiles full of open-source contributions stand out to us.
- You are able to solve both small, isolated problems like bugs in code, as well as grapple with large meta-level problems, such as epistemic strategies and research agendas.
- You have a broad knowledge of topics related (even tangentially!) to machine learning and alignment - e.g computer science, information theory, statistics, philosophy, neuroscience.
- You are good at collaboration and teamwork - many of our projects are large engineering efforts that involve most or all of the team.
- You care about the impact of your work on the longterm future of humanity and creating safe and beneficial AI.
- Large scale distributed computing + machine learning systems.
- Large models that need to be parallelized to fit in memory (think tensor / data / pipeline parallelism or surrounding techniques) including frameworks such as Deepspeed
- Deep expertise with machine learning frameworks (PyTorch, Jax etc)
- Academic publications in fields related (even tangentially) to AI safety and machine learning.
- University-level physics, mathematics, computer science or computational neuroscience.
- Non language-modelling aspects of ML such as reinforcement learning, bayesian graphical models, statistics etc.