Member of Technical Staff, Mechanistic Interpretability

Radical Numerics Inc.; San Francisco (CA); Híbrido; USD 120 000 - 160 000; Tempo integral

Resumo da oferta

Radical Numerics Inc. in San Francisco is seeking a Member of Technical Staff to explore multimodal genome language models. The role involves researching how these models process information and developing tools for mechanistic interpretability, with key responsibilities that include designing experiments and collaborating across various research disciplines.

The ideal candidate comes with a solid background in machine learning and a passion for scientific discovery. Join us in advancing the understanding of biological models and their clinical applications.

Qualificações

• Strong background in machine learning and representation learning.
• Deep understanding of modern models and training methods.
• Excellent written and verbal communication skills.

Responsabilidades

• Study how multilodal genome language models work.
• Develop experiments to uncover model behavior.
• Develop infrastructure for mechanistic interpretability.

Conhecimentos

Machine Learning Research

Large Language Models

Model Analysis

Python

PyTorch

Experimental Skills

Ferramentas

Activation Analysis

Causal Interventions

Sparse Representations

Descrição da oferta de emprego

About Us

Radical Numerics is an AI research lab building general biological intelligence. Our mission is to master the code of life, and our purpose is to reduce human suffering.

Our team created Evo, and started the field of generative genomics. Our work was featured on the cover of Science, and presented by our CEO on the main stage of TED2025. Evo was used to create the first AI gene therapy tool CRISPR-Cas9, and the first AI whole genome from scratch. Evo 2, featured in Nature, is the largest fully open source AI project across any domain.

Radical Numerics is bringing the rigor of distributed systems, model architecture, and numerics research to the challenges of biology. We’ve redesigned the foundation model training stack to turn the world’s raw scientific data (e.g. biological sequences, experiments, and physical processes), into intelligible, generative models that can expand and accelerate what humanity can understand, design, and cure.

The same generative breakthroughs that enable life-saving cures also lower the barrier to creating engineered threats and AI-generated bioweapons. We believe these forces are inseparable. Radical Numerics was founded to develop both the power to design and the responsibility to defend.

About the Role

As a Member of Technical Staff, Mechanistic Interpretability at Radical Numerics, you will study how multimodal genome language models represent, process, and reason about information internally. Your work will focus on opening the black box: developing tools, experiments, and theories that help us understand model behavior, uncover internal mechanisms, and drive scientific discovery with real-world clinical applications. Interpretability doesn't just start and end at off-the-shelf models. It's critical to our model understanding and model development itself, and pushes the boundaries of what these biological language models can do.

This is a research-oriented role for someone excited by the intersection of deep learning, scientific discovery, and model understanding. You should be motivated by questions such as: What concepts emerge during training? Which circuits drive specific behaviors? How do we extract biological insights for understanding the fundamental pathways for disease? How do we control generative design for life saving treatments?

We believe that understanding frontier models will become increasingly important as they are applied to consequential domains in science and biology. Mechanistic interpretability offers a path toward deeper scientific insight into learning systems, improved model evaluations, and ultimately, mastery over the code of life. This role sits at the intersection of AI research, systems engineering, and scientific discovery.

What You’ll Do

Understand how frontier biological models work. Design and execute experiments to uncover the features, circuits, representations, and mechanisms that drive model behavior. Study how models learn, store, retrieve, and manipulate information across scales.

Build the tooling for model understanding. Develop infrastructure and methods for mechanistic interpretability, including activation analysis, causal interventions, probing, feature discovery, sparse representations, circuit tracing, and large-scale interpretability workflows.

Connect mechanisms to capabilities and failures. Investigate how internal representations relate to downstream performance, reasoning, robustness, controllability, and scientific usefulness. Identify the mechanisms underlying both desirable capabilities and failure modes.

Advance interpretability research. Develop new techniques for understanding large models and use them to improve evaluation, reliability, safety, and model design.

Study multimodal genome language models. Investigate how models represent biological concepts, mechanisms, and abstractions, and use those insights to better understand both the models themselves and the systems they are modeling.

Collaborate across research disciplines. Work closely with teams spanning model architecture, training, systems, safety, and biology to turn interpretability insights into better models, evaluations, and scientific outcomes.

What We’re Looking For

• Strong background in machine learning research, particularly in large language models, representation learning, mechanistic interpretability, model analysis, AI safety, or related areas.
• Deep understanding of modern deep learning architectures and training methods, including transformers, representation learning, optimization, and large-scale model systems.
• Proficiency in Python and PyTorch, with experience building research tooling and conducting rigorous empirical investigations.
• Strong experimental skills and scientific judgment. You are comfortable designing studies, evaluating competing hypotheses, and distinguishing meaningful findings from artifacts.
• Excellent written and verbal communication skills, including the ability to clearly explain technical insights and research results.
• Curiosity about how models work internally and a desire to develop deeper scientific understanding rather than treating models as black boxes.

Nice to Have

• Experience with mechanistic interpretability techniques such as activation patching, causal tracing, probing, sparse autoencoders, feature analysis, circuit discovery, or representation analysis.
• Research experience in frontier AI systems, AI safety, alignment, or model evaluations.
• Experience working with large-scale training systems, distributed computing, or model infrastructure.
• Background in computational biology, genomics, neuroscience, complex systems, or another scientific field involving high-dimensional data.
• Contributions to open-source ML research, interpretability tooling, or model analysis frameworks.
• Publications or demonstrated research contributions in machine learning, interpretability, AI safety, or related areas.

Why Radical Numerics

Help build the next generation of biological foundation models while advancing our understanding of how frontier AI systems work internally. At Radical Numerics, interpretability is not an afterthought. We view model understanding as a core part of model development itself, driving better capabilities, stronger reliability, and deeper scientific discovery across both AI and biology.

Radical Numerics is committed to equal employment opportunity and does not discriminate in any employment opportunities or practices based on an individual's race, color, creed, gender (including gender identity and gender expression), religion (all aspects of religious beliefs, observance or practice, including religious dress or grooming practices), marital status, registered domestic partner status, age, national origin or ancestry (including language use restrictions and possession of a driver’s license issued under California Vehicle Code section 12801.9), natural hair, physical or mental disability, political affiliation, medical condition (including cancer or a record or history of cancer, and genetic characteristics), sex (including pregnancy, childbirth, breastfeeding or related medical condition), genetic information, sexual orientation, military and veteran status or any other consideration made unlawful by federal, state, or local laws. It also prohibits unlawful discrimination based on the perception that anyone has any of those characteristics, or is associated with a person who has or is perceived as having any of those characteristics.