Astera is a private foundation on a mission to steer science and technology toward a better future. We believe the coming years will bring an era of unprecedented scientific and technological advancement as exponential progress in AI converges with central advances in other fields to accelerate innovation. This inflection point provides an unparalleled opportunity to fundamentally rethink the institutions, systems, and tools that drive progress. We have committed a $3.5B endowment to that work and fit the form of our support to the function it serves: grants, investments, and in-house research programs. We accept meaningful risk in the research we back and expect some of it to fail because transformative ideas are rarely the ones already proven. Our projects, and the foundation itself, operate like high-velocity startups, with a high performance bar and competitive compensation to match. We are searching for leaders who are compelled by the challenge of driving high-impact change in the way science is funded, conducted, and communicated. You can read more about our mission, vision, and programming at astera.org/vision.
Datasets in many areas, science in particular, are often small, heterogeneous, and expensive. Human scientists can take these datasets and generate models to describe them, but this process of model induction is labor-intensive and error-prone. Machine learning is a general and scalable solution, but it is not uniformly sample efficient.
The Astera Institute is seeking a Machine Learning Researcher to help surmount this barrier with new architectures for data-efficient and general model induction. This includes bootstrapped program synthesis, along with components for a system that synthesizes its own learning algorithms - a machine learning strange loop. This is a full time position that reports to Timothy Hanson.
Hypothesize, test, and refine means of improving generalization performance of common architectural elements, including different forms of attention. This includes devising controlled datasets to elucidate e.g. learning order & learned representations.
Think both mathematically and empirically about problems of runtime inference in gradient-trained networks, with an eye to the extensive literature on statistical learning and an open mind to the many forms of constrained optimization.
Contribute to a well-documented and well-instrumented code base that is performant where necessary yet expeditious where experimental throughput demands.
Masters or equivalent in machine learning, mathematics, or equivalent fields (strong candidates from neuroscience are encouraged to apply).
Fluency with Pytorch, and familiarity with JAX, CUDA, and/or Triton + their open-source ecosystems.
Demonstrated ability do fundamental research.
Demonstrated ability to work in teams.
This position is hybrid at our office in Emeryville, CA. Some travel may be required from time-to-time for in-person collaboration and work.
The posted salary range is based on location in the Bay Area. The successful candidate will receive a competitive compensation package, commensurate with their experience and location.
Own the development and deployment of machine-learned force fields for PFAS substitution, from generating first-principles training data and validating models to delivering reliable property predictions for molecular screening.
Design and deploy machine learning models for ad ranking, bidding, and optimization, owning the model lifecycle from research and feature engineering through scalable production systems, monitoring, and retraining.
Develop and own deep learning R&D for computer vision, designing and implementing algorithms, improving model performance, and applying advances in AI to 2D and 3D vision technologies.
Build and improve production machine learning systems, developing scalable data pipelines, deep learning and time-series models, agentic workflows, and monitoring processes for model performance.
Build end-to-end machine learning systems for real-time fraud detection, from large-scale feature engineering and model training to low-latency prediction serving across distributed infrastructure.
