02 Applied Research Scientist San Francisco, CA · Full-time

Softmax Inc.; San Francisco (CA); On-site; USD 150,000 - 210,000; Full time

Job summary

Softmax Inc. is seeking an Applied Research Scientist to own end-to-end applied research on organic AI alignment.

You will select questions, design environments and evaluations, implement experiments, analyze results, and convert findings into reusable systems. The role focuses on training new model families and requires reasoning from first principles about architectures, losses, curricula, scaling, and failure modes.

Qualifications

• Graduate work or equivalent research depth in ML, reinforcement learning, CS, physics, math, complex systems, or another quantitative field.
• Strong deep learning and ML/RL instincts, including neural architecture design, optimization, loss functions, training dynamics, evaluation, reward design, generalization, baselines, and failure modes.
• Experience training nontrivial models end-to-end, debugging unstable runs, and implementing architectures below the paper level when needed.

Responsibilities

• Choose and run high-leverage experiments. Translate questions into architectures, tasks, curricula, evaluations, and experiments; run them and iterate.
• Train new deep learning architectures from scratch using novel substrates, losses, data mixtures, and scaling recipes.
• Run rigorous experiments with ablations, baselines, controls, and statistical judgment.
• Diagnose training behavior by reading learning curves, gradients, activations, memory traces, and failure cases.
• Build research infrastructure with multi-agent tasks, curricula, metrics, dashboards, and analysis workflows.
• Turn research into working systems by writing production-quality Python and C++ and integrating with training/evaluation pipelines.
• Move fast with scientific discipline, learning from real runs, agents, and users.
• Leverage AI tools aggressively to augment research throughput.
• Collaborate with a small, strong team with direct feedback.

Skills

Deep analytical

ML/RL instincts

Deep learning

Systems fluency

Ownership mentality

AI-native workflow

Collaboration

Tools

Python

C++

Job description

We're looking for an Applied Research Scientist to own applied research on organic AI alignment end-to-end. You'll choose the question, design the environment configs and evaluations, implement experiments, analyze results, and turn what works into reusable systems. In this role, research and execution go hand in hand: you generate clarity from ambiguity and produce conclusions that survive contact with reality.

This role is focused on training new model families. You should be comfortable reasoning from first principles about architectures, losses, optimization, curricula, scaling behavior, and failure modes. You'll help train a new generation of RL approaches: systems that can stream, adapt, grow, and keep learning over time.

You're fluent in modern AI tools and use them across coding, literature review, experiment design, analysis, and writing. You use that leverage to take on harder problems, not just to iterate faster.

What You'll Do

• Choose and run high-leverage experiments. Translate model-training and alignment questions into concrete architectures, tasks, curricula, evaluations, and experiments; run them, interpret them, and iterate.
• Train new deep learning architectures from scratch. Work with novel model substrates, loss functions, data mixtures, recurrent computation, graph structure, and scaling recipes to discover what actually trains.
• Run rigorous experiments. Use ablations, baselines, controls, and statistical judgment. Understand what your data can and can't tell you.
• Diagnose training behavior. Read learning curves, gradients, activations, memory traces, ablations, and failure cases to understand whether a result is an architecture issue, optimization issue, data issue, evaluation issue, or implementation bug.
• Build research infrastructure that changes what we can know. Design multi-agent tasks, curricula, metrics, dashboards, and analysis workflows that surface real signal.
• Turn research into working systems. Write production-quality Python and C++ when needed, integrate with training and evaluation pipelines, and make your ideas reusable by the rest of the team.
• Move fast with scientific discipline. Learn from real runs, real agents, and real users. Know when to prototype quickly and when the foundation matters.
• Leverage AI tools aggressively. Use AI assistants, coding agents, and research tools to operate beyond traditional throughput.
• Collaborate with a small, strong team. Low overhead, high ownership, direct feedback, and rigorous technical discussion.

What We're Looking For

• Deep analytical and research background. Graduate work or equivalent research depth in ML, reinforcement learning, CS, physics, math, complex systems, or another quantitative field.
• Strong deep learning and ML/RL instincts. You understand neural architecture design, optimization, loss functions, training dynamics, evaluation, reward design, generalization, baselines, and failure modes.
• Deep learning implementation depth. You have trained nontrivial models end-to-end, debugged unstable runs, designed or modified architectures, and are comfortable working below the paper abstraction level when needed.
• Interest in architectures beyond standard transformer stacks. Experience with recurrent models, sequence modeling, graph neural networks, message passing, world models, memory systems, multimodal models, or reinforcement learning is especially relevant.
• Systems fluency. Comfortable building, debugging, and scaling the tooling your research needs. Python and C++.
• Scientific taste. You choose questions that matter, design experiments that can answer them, and notice when a result is beautiful but probably wrong.
• AI-native workflow. You already use AI tools daily and have opinions about how to supervise them effectively.
• Ownership mentality. You see problems, propose research programs, and drive them to completion.

Compensation

Level and compensation are calibrated to the scope and impact expected from the role. The range spans experience levels that larger organizations might label senior, staff, principal, or research scientist.

Softmax is an equal opportunity employer. We welcome applicants of all backgrounds, identities, and experience levels, and consider qualified applicants with arrest and conviction records in line with the San Francisco Fair Chance Ordinance.