Research Engineer, Interpretability

Anthropic Limited; San Francisco (CA); Hybrid; USD 315,000 - 560,000; Full time

Benefits offered by this job

Equity donation matching

Vacation and parental leave

Flexible working hours

Office space in SF

Job summary

Anthropic is seeking an experienced software engineer to advance the Interpretability team’s infrastructure for training and inference of safety-focused AI models. You will work with researchers and infra teams to optimize performance, instrument activations, and empower rapid experimentation.

Strong candidates will have 5–10+ years building software, expertise in Python and ML stacks (PyTorch/CUDA or JAX), and a passion for interpretability research and its societal impact.

Qualifications

• 5-10+ years of software development experience.
• Proficiency in at least one programming language (Python preferred).
• Interest in interpretability and AI safety.

Responsibilities

• Build and maintain inference and training infrastructure for interpretability research.
• Resolve scaling bottlenecks via profiling and optimization.
• Design tools and platforms enabling rapid researcher experimentation.
• Support production safety audits with reliable tooling.
• Collaborate with researchers and infrastructure teams across the stack.

Skills

Python

Software engineering

5-10 years

Research collaboration

Ambiguity handling

Interpretable AI

Education

Bachelor's degree

Tools

PyTorch

CUDA

JAX

Job description

Anthropic’s mission is to create reliable, interpretable, and steerable AI systems. We want AI to be safe and beneficial for our users and for society as a whole. Our team is a quickly growing group of committed researchers, engineers, policy experts, and business leaders working together to build beneficial AI systems.

About the role:

When you see what modern language models are capable of, do you wonder, "How do these things work? How can we trust them?"

The Interpretability team at Anthropic is working to reverse-engineer how trained models work because we believe that a mechanistic understanding is the most robust way to make advanced systems safe.

Think of us as doing "neuroscience" of neural networks using "microscopes" we build - or reverse-engineering neural networks like binary programs.

• 60 Minutes segment - Around 8:07, see a demo of tooling our team built
• New Yorker article - what it's like to work on one of AI's hardest open problems

Even if you haven’t worked on interpretability before, the infrastructure expertise is similar to what's needed across the lifecycle of a production language model:

• Pretraining: Training dictionary learning models looks a lot like model pretraining - creating stable, performant training jobs for massively parameterized models across thousands of chips
• Inference: Interp runs a customized inference stack. Day-to-day analysis requires services that allow editing a model's internal activations mid-forward-pass - for example, adding a "steering vector"
• Performance: Like all LLM work, we push up against the limits of hardware and software. Rather than squeezing the last 0.1%, we are focused on finding bottlenecks, fixing them and moving ahead given rapidly evolving research and safety mission

The science keeps scaling - and it's now applied directly in safety audits on frontier models, with real deadlines. As our research has matured, engineering and infrastructure have become a bottleneck. Your work will have a direct impact on one of the most important open problems in AI.

Responsibilities:

• Build and maintain the specialized inference and training infrastructure that powers interpretability research - including instrumented forward/backward passes, activation extraction, and steering vector application
• Resolve scaling and efficiency bottlenecks through profiling, optimization, and close collaboration with peer infrastructure teams
• Design tools, abstractions, and platforms that enable researchers to rapidly experiment without hitting engineering barriers
• Help bring interpretability research into production safety audits - with real deadlines and high reliability expectations
• Work across the stack - from model internals and accelerator-level optimization to user-facing research tooling

You may be a good fit if you:

• Have 5-10+ years of experience building software
• Are highly proficient in at least one programming language (e.g., Python, Rust, Go, Java) and productive with Python
• Are extremely curious about unfamiliar domains; can quickly learn and put that knowledge to work, e.g. diving into new layers of the stack to find bottlenecks
• Have a strong ability to prioritize the most impactful work and are comfortable operating with ambiguity and questioning assumptions
• Prefer fast-moving collaborative projects to extensive solo efforts
• Are curious about interpretability research and its role in AI safety (though no research experience is required!)
• Care about the societal impacts and ethics of your work
• Are comfortable working closely with researchers, translating research needs into engineering solutions.

Strong candidates may also have experience with:

• Optimizing the performance of large-scale distributed systems
• Language modeling fundamentals with transformers
• High Performance LLM optimization: memory management, compute efficiency, parallelism strategies, inference throughput optimization
• Working hands-on in a mainstream ML stack - PyTorch/CUDA on GPUs or JAX/XLA on TPUs
• Collaborating closely with researchers and building tooling to support research teams; or directly performed research with complex engineering challenges

Representative Projects:

• Building Garcon , a tool that allows researchers to easily instrument LLMs to extract internal activations
• Designing and optimizing a pipeline to efficiently collect petabytes of transformer activations and shuffle them
• Profiling and optimizing ML training jobs, including multi-GPU parallelism and memory optimization
• Building a steered inference system that applies targeted interventions to model internals at scale (conceptually similar to Golden Gate Claude but for safety research)

Role Specific Location Policy:

• This role is based in the San Francisco office; however, we are open to considering exceptional candidates for remote work on a case-by-case basis.

The annual compensation range for this role is listed below.

For sales roles, the range provided is the role’s On Target Earnings ("OTE") range, meaning that the range includes both the sales commissions/sales bonuses target and annual base salary for the role.

$315,000 - $560,000 USD

Logistics

Minimum education: Bachelor’s degree or an equivalent combination of education, training, and/or experience

Required field of study:A field relevant to the role as demonstrated through coursework, training, or professional experience

Minimum years of experience: Years of experience required will correlate with the internal job level requirements for the position

Location-based hybrid policy: Currently, we expect all staff to be in one of our offices at least 25% of the time. However, some roles may require more time in our offices.

Visa sponsorship:We do sponsor visas! However, we aren't able to successfully sponsor visas for every role and every candidate. But if we make you an offer, we will make every reasonable effort to get you a visa, and we retain an immigration lawyer to help with this.

We believe AI will have a transformative impact on the world, and we’re seeking exceptional candidates who collaborate thoughtfully with Claude to realize this vision. At the same time, we want to understand your unique skills, expertise, and perspective through our hiring process. We invite you to review our AI partnership guidelines for candidates and confirm your understanding by selecting “Yes.”

How we're different

We believe that the highest-impact AI research will be big science. At Anthropic we work as a single cohesive team on just a few large-scale research efforts. And we value impact — advancing our long-term goals of steerable, trustworthy AI — rather than work on smaller and more specific puzzles. We view AI research as an empirical science, which has as much in common with physics and biology as with traditional efforts in computer science. We’re an extremely collaborative group, and we host frequent research discussions to ensure that we are pursuing the highest-impact work at any given time. As such, we greatly value communication skills.

The easiest way to understand our research directions is to read our recent research. This research continues many of the directions our team worked on prior to Anthropic, including: GPT-3, Circuit-Based Interpretability, Multimodal Neurons, Scaling Laws, AI & Compute, Concrete Problems in AI Safety, and Learning from Human Preferences.

Come work with us!

Anthropic is a public benefit corporation headquartered in San Francisco. We offer competitive compensation and benefits, optional equity donation matching, generous vacation and parental leave, flexible working hours, and a lovely office space in which to collaborate with colleagues. Guidance on Candidates' AI Usage: Learn about our policy for using AI in our application process.

We are a federal contractor or subcontractor. The law requires us to provide equal employment opportunity to qualified people with disabilities. We have a goal of having at least 7% of our workers as people with disabilities. The law says we must measure our progress towards this goal. To do this, we must ask applicants and employees if they have a disability or have ever had one. People can become disabled, so we need to ask this question at least every five years.

Completing this form is voluntary, and we hope that you will choose to do so. Your answer is confidential. No one who makes hiring decisions will see it. Your decision to complete the form and your answer will not harm you in any way. If you want to learn more about the law or this form, visit www.dol.gov/ofccp .

How do you know if you have a disability?

A disability is a condition that substantially limits one or more of your “major life activities.” If you have or have ever had such a condition, you are a person with a disability. Disabilities include, but are not limited to:

• Alcohol or other substance use disorder (not currently using drugs illegally)
• Autoimmune disorder, for example, lupus, fibromyalgia, rheumatoid arthritis, HIV/AIDS
• Blind or low vision
• Cancer (past or present)
• Cardiovascular or heart disease
• Celiac disease
• Cerebral palsy
• Deaf or serious difficulty hearing
• Diabetes
• Disfigurement, for example, disfigurement caused by burns, wounds, accidents, or congenital disorders
• Epilepsy or other seizure disorder
• Gastrointestinal disorders, for example, Crohn's Disease, irritable bowel syndrome
• Intellectual or developmental disability
• Mental health conditions, for example, depression, bipolar disorder, anxiety disorder, schizophrenia, PTSD
• Missing limbs or partially missing limbs
• Mobility impairment, benefiting from the use of a wheelchair, scooter, walker, leg brace(s) and/or other supports
• Nervous system condition, for example, migraine headaches, Parkinson’s disease, multiple sclerosis (MS)
• Neurodivergence, for example, attention‑deficit/hyperactivity disorder (ADHD), autism spectrum disorder, dyslexia, dyspraxia, other learning disabilities
• Partial or complete paralysis (any cause)
• Pulmonary or respiratory conditions, for example, tuberculosis, asthma, emphysema
• Short stature (dwarfism)
• Traumatic brain injury

PUBLIC BURDEN STATEMENT: According to the Paperwork Reduction Act of 1995 no persons are required to respond to a collection of information unless such collection displays a valid OMB control number. This survey should take about 5 minutes to complete.