11 дней назад
Research Engineer, Interpretability Systems (AI)
350 000$
Мэтч & Сопровод
Для мэтча с этой вакансией нужен Plus
Описание вакансии
Текст:
TL;DR
Research Engineer, Interpretability Systems (AI): Building experimental infrastructure, RL-style environments, and tooling for mechanistic interpretability and alignment research in large language models with an accent on activation tracing, latent concept detection, and activation-level steering. Focus on implementing probes for hidden objectives, defining robustness benchmarks, and rapidly converting research ideas into experiments.
Location: San Francisco Bay Area, on-site
Salary: $350,000 per year
Company
Early-stage AI research lab founded by former frontier-model researchers, focused on alignment and interpretability for large language models.
What you will do
- Build custom RL-style environments and experimental testbeds for interpretability research.
- Develop tooling for activation tracing, concept detection, mechanistic analysis, and internal representation probing.
- Implement probes for latent concepts such as deception, uncertainty, goals, and hidden objectives.
- Prototype activation-level steering methods beyond prompting and fine-tuning.
- Collaborate with researchers to iterate from research ideas through implementation, experiments, and results.
- Define benchmarks and measurement frameworks for model consistency, robustness, alignment, and interpretability.
Requirements
- Strong software engineering fundamentals and experience building experimental ML systems.
- Experience working with model internals, representations, or post-training systems.
- Strong Python and deep learning framework experience, preferably PyTorch.
- Ability to rapidly prototype and iterate in open-ended research environments.
- Experience in interpretability, alignment, or ML research is preferred.
- A PhD is ideal.
Будьте осторожны: если работодатель просит войти в их систему, используя iCloud/Google, прислать код/пароль, запустить код/ПО, не делайте этого - это мошенники. Обязательно жмите "Пожаловаться" или пишите в поддержку. Подробнее в гайде →