Вакансии

Location: Москва Данные — главный рычаг качества фронтирной модели. Станьте тем, кто его держит ------------------------------------------------------------------------------ GigaChat обучает одна из крупнейших ML-команд страны. Мы знаем то же, что знают OpenAI и Anthropic: качество следующей версии модели определяют претрейн-данные. Поэтому мы ищем не «инженера по датасетам», а исследователя, который будет владеть данными как рычагом роста модели — от стратегии сбора до измеримого прироста на бенчмарках. Контур влияния короткий и прямой: ваша гипотеза → эксперимент → обучающий прогон → прирост метрик модели, которой пользуются миллионы людей. Обязанности * Вести программу экспериментов с данными. * Планировать, запускать и анализировать ablation-эксперименты: сравнивать версии смеси данных, фильтров и порядка обучения при фиксированном вычислительном бюджете. Вы автор эксперимента, а не исполнитель. * Оптимизировать смесь и порядок использования данных. * Отвечать на главный вопрос претрейна: какие данные, в какой пропорции и в каком порядке должна видеть модель. Применять интуицию scaling laws к распределению токенов. * Измерять качество данных. * Превращать расплывчатые «хорошие данные» в конкретные измеримые сигналы: классификаторы для отбора документов, насыщенных рассуждениями, кодом и знаниями. * Курировать корпус на масштабе триллионов токенов. * Проводить точную и fuzzy-дедупликацию, фильтрацию, очистку от утечек бенчмарков, html2text. * Собирать данные для agentic-pretrain. * Проводить поиск, сбор и генерацию траекторий использования инструментов, многошаговых рассуждений и взаимодействия со средой — чтобы агентные способности закладывались в модель уже на этапе претрейна, а не только в пост-обучении. * Создавать рецепты подготовки синтетических данных. * Генерировать синтетику под дефицитные способности модели (длинный контекст, рассуждения, код) — с количественной проверкой через эксперименты. * Замыкать петлю: «данные → обучение → оценка». * Развивать контур бенчмарков и сетапов экспериментов для измерений, чтобы каждое изменение данных имело измеримый ответ. Требования * Нам нужны сильные исследователи, а не исполнители. Главное — умение самостоятельно задавать правильные вопросы, проектировать эксперименты и доводить гипотезы до измеримого результата. * Опыт именно с LLM не обязателен, если у вас сильная экспериментальная практика в любой области ML — компьютерное зрение, рекомендательные системы, классический ML, наука — специфике больших языковых моделей мы научим. * Сильные основы ML и статистики, культура контролируемого эксперимента: гипотеза → дизайн → измерение → вывод. * Уверенный Python и опыт обучения моделей на PyTorch или аналогах. * Самостоятельность в исследованиях: умение вести направление от вопроса до результата, а не ждать постановки задачи. * Понимание цикла обучения LLM хотя бы на уровне теории: токенизация, scaling laws, влияние состава данных на качество. * Знакомство с распределённой обработкой данных (Spark, Dask, Airflow, Kubernetes) и открытыми data-проектами (FineWeb, OLMo, Dolma) — как минимум представление о том, как это устроено. Не обязательно закрывать все вышеперечисленные пункты. Если вы сильный экспериментатор и хотите дорасти до владения всем циклом данных фронтирной модели — давайте поговорим. Условия * Вычислительные ресурсы под ваши эксперименты. * Один из крупнейших GPU-кластеров в стране. * ablation-эксперименты — это обучающие прогоны, и вы сможете запускать их регулярно, а не «раз в квартал, если повезёт». * Претрейн с нуля, а не дообучение чужих весов. * Мы обучаем собственную модель с первого токена — это значит, что ваши решения по данным определяют фундамент модели, а не косметику поверх чужого претрейна. Сейчас мы учим модели размером 400–700B параметров, и у нас есть амбиции выйти на масштаб 1.5–2T. * Опенсорс и мировая планка. Мы выкладываем модели в открытый доступ, и наша цель — достичь мирового уровня среди опенсорс-моделей. Ваша работа будет видна всему сообществу и будет сравниваться с лучшими открытыми моделями мира. * Полная ориентация на результат. * Минимум бюрократии, максимум ответственности и свободы. Работу меряем не количеством задач, а приростом модели на бенчмарках при том же бюджете. Подтверждённая гипотеза попадает в обучение следующей модели. * Развитая инфраструктура. * Зрелый стек для работы с данными и экспериментами: S3, YTsaurus, Airflow, DataLens, GitLab CI, wandb. * Вы приходите не строить платформу с нуля, а исследовать — пайплайны, оркестрация, трекинг экспериментов и визуализация уже работают.

Location: Москва Pretrain - является самым масштабным и затратным этапом обучения VLM. На нем закладывается огромный фундамент возможностей модели, который позволяет в дальнейшем ей быть полезной пользователю и решать задачи CV. Нам требуется эксперт, который понимает какие данные двигают качество, весь скоуп задач компьютерного зрения и подходы к их решению, будет собирать процессы вокруг обогащения претрейна самыми качественными и необходимыми данными для множества доменов, оценивать качество и сложность датасетов, создавать полезные дашборды, заботиться о воспроизводимости. Нужен хороший организатор дата-команды и дата-пайплайнов с сильными техническими навыками. _Первый этап отбора на эту вакансию – общение с AI-рекрутером. После отклика вам на почту и в чат на платформе HeadHunter придет приглашение пройти первичное интервью с ГигаРекрутером в Telegram. Диалог займёт примерно 10 минут. Его задача — уточнить недостающие детали и ускорить рассмотрение вашей кандидатуры. ГигаРекрутер только начинает свой путь, поэтому просим относиться с пониманием. Ваш опыт и участие помогут сделать его удобным и полезным!_ Обязанности * строить data engine для pretrain и stage 1.5; * отвечать за sourcing, curation, dedup, filtering, balancing и mixture design; * выявлять слабые места в данных, влияющие на качество модели, составлять план по улучшению и реализовывать план в срок; * управлять качеством данных и воспроизводимостью; * организовывать статистику, срезы и мониторинг по данным; * работать с training/eval командами; * приоритизировать ручную и полуавтоматическую обработку там, где это эффективнее полной автоматизации; * нанимать и вести дата-инженеров. Требования * опыт работы в production команде (сборка окружения, выкатка моделей как оффлайн, так и онлайн, большие нагрузки как преимущество); * понимание, как данные влияют на качество VLM; * опыт построения дата пайплайнов и data quality процессов; * умение ставить задачи дата-инженерам, девопсам/ML-опсам, ML разработчикам; * умение приоритизировать задачи, распределять имеющиеся ресурсы; Условия * крупнейшее DS&AI community — более 600 DS-специалистов банка * дайджест о самых последних разработках в области DS&AI и отчеты с крупнейших конференций мира * возможность быть соавтором НИРов и статей для международных конференций * возможность выбрать удобный формат работы: гибрид или офис * ежегодный пересмотр зарплаты, годовая премия * корпоративный спортзал и зоны отдыха * более 400 образовательных программ СберУниверситета для профессионального и карьерного развития * расширенный ДМС, льготное страхование для семьи и корпоративная пенсионная программа * ипотека выгоднее до 7% для каждого сотрудника * бесплатная подписка СберПрайм+, скидки на продукты компаний-партнеров * вознаграждение за рекомендацию друзей в команду Сбера.

Location: Москва Мы развиваем GigaChat и ищем сильного руководителя направления ML pretrain больших языковых моделей. Недавно мы обучили MoE-модель на 700 миллиардов параметров и не собираемся останавливаться — обучение идёт на кластерах H100 и B200, а pretrain является ядром самого быстрорастущего AI-проекта Сбера. Это роль для человека, который возьмёт на себя архитектурную часть pretrain: design моделей, законы масштабирования, выбор и адаптация attention / MoE / позиционных схем, а также все архитектурные решения, которые определяют, какой именно будет следующая флагманская модель GigaChat. Обязанности * развивать архитектуру моделей GigaChat * определять, как должна развиваться архитектура pretrain-моделей: какие направления наиболее важны, как измерять прогресс и что в первую очередь ограничивает рост качества модели * проектировать архитектуру флагманских моделей и отвечать за ключевые архитектурные решения: attention, позиционные схемы (RoPE и варианты), нормализации, активации, инициализация * развивать MoE-архитектуру: маршрутизация, балансировка экспертов, устойчивость маршрутизатора, влияние на качество и производительность * работать с long-context и мультимодальностью на уровне архитектуры: что именно нужно менять в модели, чтобы эти возможности работали стабильно * изучать и применять законы масштабирования * проводить scaling-эксперименты и на их основе принимать решения по размеру модели, ширине / глубине, числу и размеру экспертов, размерам батча и длительности обучения * предсказывать, как архитектурные изменения поведут себя при переходе с небольших абляций на полный масштаб * определять и развивать метрики, которые корректно отражают архитектурные изменения в обучении моделей * определять, в каких случаях архитектурные изменения действительно дают прирост качества по сравнению с более простыми baseline'ами, а в каких — нет * анализировать стабильность архитектурных решений * разбираться с нестабильностью на больших прогонах со стороны архитектуры: почему конкретная конфигурация расходится, где проявляются артефакты маршрутизации, коллапс энтропии, неустойчивости в attention * предлагать изменения в нормализациях, клиппинге, точности вычислений и структуре блоков, которые делают обучение более предсказуемым * обеспечивать безопасное масштабирование при внедрении крупных архитектурных изменений в основной трейн * писать ключевой код и оставаться сильным исследователем * самостоятельно писать и дорабатывать архитектурные компоненты и абляции. * Делать надёжные и воспроизводимые эксперименты: понятные версии данных, конфиги, сравнение запусков, контроль деградаций * читать статьи, воспроизводить ключевые результаты и адаптировать лучшие идеи под наши задачи и инфраструктуру * руководить сильной технической командой * руководить командой исследователей и инженеров, работающих над архитектурой, задавать высокую планку по качеству решений, скорости работы и глубине проработки * Паомогать команде превращать архитектурные идеи в работающие решения, которые можно встроить в основной цикл обучения * удерживать баланс между глубиной исследований, инженерной надёжностью и практическим результатом для модели. Требования * отличное владение Python и PyTorch * глубокое понимание устройства обучения нейросетей: не на уровне обзоров, а на уровне, где вы можете объяснить, почему конкретная архитектурная конфигурация расходится, глядя на кривые функции потерь, нормы градиентов и энтропии * глубокое понимание архитектуры LLM: Transformer, attention (MHA/GQA/MLA), RoPE и варианты позиционных эмбеддингов, нормализации, инициализации, long-context, MoE * практический опыт с обучением больших моделей (а не только инференсом) и проведением архитектурных абляций * способность самостоятельно взять направление и довести его до результата: от чтения статей и постановки гипотез до внедрения в основной трейн * умение ставить гипотезы, проектировать эксперименты и принимать решения на основе результатов * опыт руководства сильной технической командой и готовность лично писать важные части системы руками. Условия * возможность выбрать удобный формат работы: гибрид или офис * ежегодный пересмотр зарплаты, годовая премия * корпоративный спортзал и зоны отдыха * более 400 образовательных программ СберУниверситета для профессионального и карьерного развития * расширенный ДМС, льготное страхование для семьи и корпоративная пенсионная программа * ипотека выгоднее до 7% для каждого сотрудника * бесплатная подписка СберПрайм+, скидки на продукты компаний-партнеров * вознаграждение за рекомендацию друзей в команду Сбера.

Location: London, UK About Anthropic 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. Anthropic is at the forefront of AI research, dedicated to developing safe, ethical, and powerful artificial intelligence. Our mission is to ensure that transformative AI systems are aligned with human interests. We are seeking a Research Engineer to join our Pretraining team, responsible for developing the next generation of large language models. In this role, you will work at the intersection of cutting-edge research and practical engineering, contributing to the development of safe, steerable, and trustworthy AI systems. Key Responsibilities: Conduct research and implement solutions in areas such as model architecture, algorithms, data processing, and optimizer development Independently lead small research projects while collaborating with team members on larger initiatives Design, run, and analyze scientific experiments to advance our understanding of large language models Optimize and scale our training infrastructure to improve efficiency and reliability Develop and improve dev tooling to enhance team productivity Contribute to the entire stack, from low-level optimizations to high-level model design Qualifications: Advanced degree (MS or PhD) in Computer Science, Machine Learning, or a related field Strong software engineering skills with a proven track record of building complex systems Expertise in Python and experience with deep learning frameworks (PyTorch preferred) Familiarity with large-scale machine learning, particularly in the context of language models Ability to balance research goals with practical engineering constraints Strong problem-solving skills and a results-oriented mindset Excellent communication skills and ability to work in a collaborative environment Care about the societal impacts of your work Preferred Experience: Work on high-performance, large-scale ML systems Familiarity with GPUs, Kubernetes, and OS internals Experience with language modeling using transformer architectures Knowledge of reinforcement learning techniques Background in large-scale ETL processes You'll thrive in this role if you: Have significant software engineering experience Are results-oriented with a bias towards flexibility and impact Willingly take on tasks outside your job description to support the team Enjoy pair programming and collaborative work Are eager to learn more about machine learning research Are enthusiastic to work at an organization that functions as a single, cohesive team pursuing large-scale AI research projects Are working to align state of the art models with human values and preferences, understand and interpret deep neural networks, or develop new models to support these areas of research View research and engineering as two sides of the same coin, and seek to understand all aspects of our research program as well as possible, to maximize the impact of your insights Have ambitious goals for AI safety and general progress in the next few years, and you’re working to create the best outcomes over the long-term. Sample Projects: Optimizing the throughput of novel attention mechanisms Comparing compute efficiency of different Transformer variants Preparing large-scale datasets for efficient model consumption Scaling distributed training jobs to thousands of GPUs Designing fault tolerance strategies for our training infrastructure Creating interactive visualizations of model internals, such as attention patterns At Anthropic, we are committed to fostering a diverse and inclusive workplace. We strongly encourage applications from candidates of all backgrounds, including those from underrepresented groups in tech. If you're excited about pushing the boundaries of AI while prioritizing safety and ethics, we want to hear from you! 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. Annual Salary: £260,000 — £630,000 GBP 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 encourage you to apply even if you do not believe you meet every single qualification. Not all strong candidates will meet every single qualification as listed.  Research shows that people who identify as being from underrepresented groups are more prone to experiencing imposter syndrome and doubting the strength of their candidacy, so we urge you not to exclude yourself prematurely and to submit an application if you're interested in this work. We think AI systems like the ones we're building have enormous social and ethical implications. We think this makes representation even more important, and we strive to include a range of diverse perspectives on our team. Your safety matters to us. To protect yourself from potential scams, remember that Anthropic recruiters only contact you from @anthropic.com email addresses. In some cases, we may partner with vetted recruiting agencies who will identify themselves as working on behalf of Anthropic. Be cautious of emails from other domains. Legitimate Anthropic recruiters will never ask for money, fees, or banking information before your first day. If you're ever unsure about a communication, don't click any links—visit anthropic.com/careers directly for confirmed position openings. 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.

Location: London, UK About Anthropic 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: Anthropic's ML Performance and Scaling team trains our production pretrained models, work that directly shapes the company's future and our mission to build safe, beneficial AI systems. As a Research Engineer on this team, you'll ensure our frontier models train reliably, efficiently, and at scale. This is demanding, high-impact work that requires both deep technical expertise and a genuine passion for the craft of large-scale ML systems. This role lives at the boundary between research and engineering. You'll work across our entire production training stack: performance optimization, hardware debugging, experimental design, and launch coordination. During launches, the team works in tight lockstep, responding to production issues that can't wait for tomorrow. Responsibilities: Own critical aspects of our production pretraining pipeline, including model operations, performance optimization, observability, and reliability Debug and resolve complex issues across the full stack—from hardware errors and networking to training dynamics and evaluation infrastructure Design and run experiments to improve training efficiency, reduce step time, increase uptime, and enhance model performance Respond to on-call incidents during model launches, diagnosing problems quickly and coordinating solutions across teams Build and maintain production logging, monitoring dashboards, and evaluation infrastructure Add new capabilities to the training codebase, such as long context support or novel architectures Collaborate closely with teammates across SF and London, as well as with Tokens, Architectures, and Systems teams Contribute to the team's institutional knowledge by documenting systems, debugging approaches, and lessons learned You May Be a Good Fit If You: Have hands-on experience training large language models, or deep expertise with JAX, TPU, PyTorch, or large-scale distributed systems Genuinely enjoy both research and engineering work—you'd describe your ideal split as roughly 50/50 rather than heavily weighted toward one or the other Are excited about being on-call for production systems, working long days during launches, and solving hard problems under pressure Thrive when working on whatever is most impactful, even if that changes day-to-day based on what the production model needs Excel at debugging complex, ambiguous problems across multiple layers of the stack Communicate clearly and collaborate effectively, especially when coordinating across time zones or during high-stress incidents Are passionate about the work itself and want to refine your craft as a research engineer Care about the societal impacts of AI and responsible scaling Strong Candidates May Also Have: Previous experience training LLM’s or working extensively with JAX/TPU, PyTorch, or other ML frameworks at scale Contributed to open-source LLM frameworks (e.g., open_lm, llm-foundry, mesh-transformer-jax) Published research on model training, scaling laws, or ML systems Experience with production ML systems, observability tools, or evaluation infrastructure Background as a systems engineer, quant, or in other roles requiring both technical depth and operational excellence What Makes This Role Unique: This is not a typical research engineering role. The work is highly operational—you'll be deeply involved in keeping our production models training smoothly, which means being responsive to incidents, flexible about priorities, and comfortable with uncertainty. During launches, the team often works extended hours and may need to respond to issues on evenings and weekends. However, this operational intensity comes with extraordinary learning opportunities. You'll gain hands-on experience with some of the largest, most sophisticated training runs in the industry. You'll work alongside world-class researchers and engineers, and the institutional knowledge you build will compound in ways that can't be easily transferred. For people who thrive on this type of work, it's uniquely rewarding. We're building a close-knit team of people who genuinely care about doing excellent work together. If you're someone who wants to be part of training the models that will define the future of AI—and you're excited about the full reality of what that entails—we'd love to hear from you. Location: This role requires working in-office 5 days per week in London. Deadline to apply: None. Applications will be reviewed on a rolling 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. Annual Salary: £260,000 — £630,000 GBP 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 encourage you to apply even if you do not believe you meet every single qualification. Not all strong candidates will meet every single qualification as listed.  Research shows that people who identify as being from underrepresented groups are more prone to experiencing imposter syndrome and doubting the strength of their candidacy, so we urge you not to exclude yourself prematurely and to submit an application if you're interested in this work. We think AI systems like the ones we're building have enormous social and ethical implications. We think this makes representation even more important, and we strive to include a range of diverse perspectives on our team. Your safety matters to us. To protect yourself from potential scams, remember that Anthropic recruiters only contact you from @anthropic.com email addresses. In some cases, we may partner with vetted recruiting agencies who will identify themselves as working on behalf of Anthropic. Be cautious of emails from other domains. Legitimate Anthropic recruiters will never ask for money, fees, or banking information before your first day. If you're ever unsure about a communication, don't click any links—visit anthropic.com/careers directly for confirmed position openings. 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.

Location: San Francisco, CA About Anthropic 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: Anthropic's ML Performance and Scaling team trains our production pretrained models, work that directly shapes the company's future and our mission to build safe, beneficial AI systems. As a Research Engineer on this team, you'll ensure our frontier models train reliably, efficiently, and at scale. This is demanding, high-impact work that requires both deep technical expertise and a genuine passion for the craft of large-scale ML systems. This role lives at the boundary between research and engineering. You'll work across our entire production training stack: performance optimization, hardware debugging, experimental design, and launch coordination. During launches, the team works in tight lockstep, responding to production issues that can't wait for tomorrow. Responsibilities: Own critical aspects of our production pretraining pipeline, including model operations, performance optimization, observability, and reliability Debug and resolve complex issues across the full stack—from hardware errors and networking to training dynamics and evaluation infrastructure Design and run experiments to improve training efficiency, reduce step time, increase uptime, and enhance model performance Respond to on-call incidents during model launches, diagnosing problems quickly and coordinating solutions across teams Build and maintain production logging, monitoring dashboards, and evaluation infrastructure Add new capabilities to the training codebase, such as long context support or novel architectures Collaborate closely with teammates across SF and London, as well as with Tokens, Architectures, and Systems teams Contribute to the team's institutional knowledge by documenting systems, debugging approaches, and lessons learned You May Be a Good Fit If You: Have hands-on experience training large language models, or deep expertise with JAX, TPU, PyTorch, or large-scale distributed systems Genuinely enjoy both research and engineering work—you'd describe your ideal split as roughly 50/50 rather than heavily weighted toward one or the other Are excited about being on-call for production systems, working long days during launches, and solving hard problems under pressure Thrive when working on whatever is most impactful, even if that changes day-to-day based on what the production model needs Excel at debugging complex, ambiguous problems across multiple layers of the stack Communicate clearly and collaborate effectively, especially when coordinating across time zones or during high-stress incidents Are passionate about the work itself and want to refine your craft as a research engineer Care about the societal impacts of AI and responsible scaling Strong Candidates May Also Have: Previous experience training LLM’s or working extensively with JAX/TPU, PyTorch, or other ML frameworks at scale Contributed to open-source LLM frameworks (e.g., open_lm, llm-foundry, mesh-transformer-jax) Published research on model training, scaling laws, or ML systems Experience with production ML systems, observability tools, or evaluation infrastructure Background as a systems engineer, quant, or in other roles requiring both technical depth and operational excellence What Makes This Role Unique: This is not a typical research engineering role. The work is highly operational—you'll be deeply involved in keeping our production models training smoothly, which means being responsive to incidents, flexible about priorities, and comfortable with uncertainty. During launches, the team often works extended hours and may need to respond to issues on evenings and weekends. However, this operational intensity comes with extraordinary learning opportunities. You'll gain hands-on experience with some of the largest, most sophisticated training runs in the industry. You'll work alongside world-class researchers and engineers, and the institutional knowledge you build will compound in ways that can't be easily transferred. For people who thrive on this type of work, it's uniquely rewarding. We're building a close-knit team of people who genuinely care about doing excellent work together. If you're someone who wants to be part of training the models that will define the future of AI—and you're excited about the full reality of what that entails—we'd love to hear from you. Location: This role requires working in-office 5 days per week in San Francisco. Deadline to apply: None. Applications will be reviewed on a rolling 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. Annual Salary: $350,000 — $850,000 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 encourage you to apply even if you do not believe you meet every single qualification. Not all strong candidates will meet every single qualification as listed.  Research shows that people who identify as being from underrepresented groups are more prone to experiencing imposter syndrome and doubting the strength of their candidacy, so we urge you not to exclude yourself prematurely and to submit an application if you're interested in this work. We think AI systems like the ones we're building have enormous social and ethical implications. We think this makes representation even more important, and we strive to include a range of diverse perspectives on our team. Your safety matters to us. To protect yourself from potential scams, remember that Anthropic recruiters only contact you from @anthropic.com email addresses. In some cases, we may partner with vetted recruiting agencies who will identify themselves as working on behalf of Anthropic. Be cautious of emails from other domains. Legitimate Anthropic recruiters will never ask for money, fees, or banking information before your first day. If you're ever unsure about a communication, don't click any links—visit anthropic.com/careers directly for confirmed position openings. 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.

Location: Москва Мы команда исследователей и инженеров, разрабатывающих технологии post-training (SFT, RL) и контролируемой генерации для моделей генерации изображений, видео и омни-модальных. Мы сфокусированы на повышение качества, управляемости и устойчивости генеративных моделей с помощью методов RLHF / RLAIF и инструктивного редактирования. Помимо фундаментальных исследований, мы создаём прикладные решения для B2B-партнёров от системы генерации интерьера до системы генерации персонализированных аватаров. Обязанности * проектирование и обучение высокопроизводительных диффузионных моделей (R2V, R2I, R2V+A) для редактирования и генерации видео / изображений * разработка новых подходов и архитектурных решений для post-training фазы диффузионных моделей: RLHF (PPO, DPO и др.), RLAIF, SFT * разработка, масштабирование и поддержание RL пайплайнов * руководство экспериментами: постановка гипотез, разработка протоколов, анализ результатов * проведение code review, менторинг младших исследователей и инженеров, участие в планировании исследовательской дорожной карты * взаимодействие с командами Pretraining, Data, Infrastructure, Production для масштаб пайплайнов и повышения стабильности моделей в продакшене * поддержание и внедрение SOTA-подходов: мониторинг литературы (arxiv, NeurIPS, ICML, CVPR, ICLR), инициирование внутренних исследований. Требования * бакалавриат/магистратура по направлению компьютерных наук/прикладной математике/ машинному обучению или близких к ним * 3+ лет релевантного опыта исследований и разработки в области глубокого обучения/компьютерного зрения/генеративного ИИ * глубокие знания в Computer Vision и Generative Modeling: Diffusion Models, GANs, VAEs, Flow/Rectified Flow Matching * опыт работы с современными диффузионными фреймворками (Diffusers) и моделями FLUX, Wan 2.X и др.) * уверенное владение PyTorch и навыки распределённого обучения (DDP/FSDP) * понимание и практическое применение RL и RLHF ( PPO/DPO и др.) * умение проектировать архитектуры, планировать эксперименты и интерпретировать результаты. Будет плюсом: * наличие PhD * опыт публикаций или активное участие в исследованиях уровня NeurIPS / ICLR/ CVPR. Условия * конкурентный уровень заработной платы, годовые премии по результатам работы * участие в развитии и создании OpenSource продуктов * корпоративная пенсионная программа, страхование от несчастных случаев, социальные гарантии, ДМС * комфортный офис Sbergile Home (метро Кутузовская).

Location: Москва В Центре робототехники Сбера мы разрабатываем модели, которые позволяют роботам понимать окружающий мир, выполнять инструкции и решать задачи в реальных условиях. Мы работаем над Vision-Language-Action (VLA) и World Action Models (WAM), объединяя обучение на изображениях, видео, языке и робототехнических данных. Мы ищем Research Engineer, который будет заниматься предобучением и дообучением этих моделей: исследовать архитектуры и методы обучения, работать с большими мультимодальными датасетами, масштабировать обучение и адаптировать модели под конкретные задачи и робототехнические платформы. Работа охватывает полный цикл экспериментов — от постановки гипотезы и подготовки данных до оценки качества и проверки результатов на роботах совместно со смежными командами. Первый этап отбора на эту вакансию — общение с AI-рекрутером. После отклика ждите сообщение от него, диалог займёт примерно 10 минут. Задача AI-рекрутера — уточнить недостающие детали и ускорить рассмотрение вашей кандидатуры. AI-рекрутер только начинает свой путь, поэтому просим относиться с пониманием. Ваш опыт и участие помогут сделать его удобным и полезным для всех! Обязанности * разрабатывать и обучать VLA/WAM-модели для понимания сцен и инструкций, предсказания динамики окружающего мира и генерации действий робота * развивать методы предобучения на разнородных данных: изображениях, видео, текстах, демонстрациях и траекториях роботов * дообучать и адаптировать модели под новые задачи, объекты, условия и робототехнические платформы * исследовать архитектуры, представления действий и функции потерь, включая transformer-, autoregressive- и diffusion/flow matching-подходы * формировать смеси обучающих данных, выбирать стратегии семплирования и исследовать влияние качества, объёма и состава данных на способности моделей * разрабатывать и оптимизировать пайплайны распределённого обучения на GPU-кластерах: повышать стабильность обучения, эффективность вычислений и использования памяти * проводить воспроизводимые эксперименты и абляционные исследования, анализировать динамику обучения и причины ошибок * разрабатывать методики оценки моделей: измерять качество предсказаний и действий, успешность выполнения задач, устойчивость и перенос на новые условия * совместно с командами робототехники и инфраструктуры интегрировать модели в контур управления и проверять их на реальных роботах * формировать рекомендации по сбору дополнительных данных и улучшению обучающих выборок на основе результатов экспериментов и испытаний. Требования * опыт разработки и обучения современных DL-моделей в одной или нескольких областях: computer vision, video understanding/generation, multimodal learning, LLM/VLM или robot learning * уверенное владение Python и PyTorch * хорошее понимание современных нейросетевых архитектур, прежде всего transformers, методов оптимизации и принципов обучения на последовательностях * практический опыт pretraining или finetuning моделей: подготовка данных, выбор функций потерь, настройка обучения, оценка качества и анализ ошибок * опыт работы с большими датасетами и эффективными пайплайнами загрузки и обработки данных * опыт обучения моделей на нескольких GPU и понимание принципов распределённого обучения * умение ставить исследовательские гипотезы, планировать эксперименты и делать обоснованные выводы по их результатам * умение писать поддерживаемый и воспроизводимый код, работать с Git и инструментами отслеживания экспериментов * способность читать научные статьи и реализовывать перспективные подходы * самостоятельность и готовность доводить исследовательскую идею до работающей модели совместно со смежными командами. Будет плюсом * опыт работы с VLA, world models, action-conditioned video models или другими моделями для embodied AI * опыт обучения политик по демонстрациям: imitation learning, behavior cloning, diffusion policies * опыт с diffusion, flow matching, авторегрессионной генерацией видео и моделированием пространственно-временной динамики * опыт мультимодального предобучения, self-supervised learning или обучения на эгоцентрическом видео * опыт переноса моделей между робототехническими платформами и работы с разными представлениями действий * опыт reinforcement learning, offline RL или улучшения политик на основе обратной связи и результатов выполнения задач * опыт масштабирования обучения на нескольких узлах, работы с FSDP, DeepSpeed, Megatron или аналогичными инструментами * опыт исследования scaling laws, оптимизации смесей данных и планирования экспериментов при ограниченном вычислительном бюджете. * опыт запуска моделей на реальных роботах, анализа неудачных действий и оптимизации инференса * публикации, открытые проекты или воспроизводимые исследовательские результаты в ML, computer vision или робототехнике. * навыки работы с генеративными AI-моделями * опыт создания AI-агентов и использования их в работе будет преимуществом * опыт использования GigaChat, Kandinsky и аналогов в продуктах, навыки создания и использования AI-агентов * инструментальное владение AI для анализа, генерации и автоматизации. Условия * офисный формат работы, адрес офиса Автозаводская 23а к2 * ежегодный пересмотр зарплаты, квартальная и годовая премия * корпоративный спортзал и зоны отдыха * более 400 образовательных программ СберУниверситета для профессионального и карьерного развития * программа адаптации и помощь руководителя на старте * расширенный ДМС, льготное страхование для семьи и корпоративная пенсионная программа * гибкий дисконт по ипотечному кредиту, равный 1/3 ключевой ставки ЦБ * подписка Прайм с возможностью совместного использования на трёх близких * вознаграждение за рекомендацию друзей в команду Сбера.

Location: Singapore Workplace: Hybrid About Cantina: Cantina Labs is a social AI company, developing a suite of advanced real-time models that push the boundaries of expression, personality, and realism. We bring characters to life, transforming how people tell stories, connect, and create. We build and power ecosystems. Cantina, our flagship social AI platform, is just the beginning. About the Role: Cantina is expanding, and we're looking for a Research Scientist to join our growing Singapore team! In this role, you will drive foundational research on video generation models, taking ownership across the full research cycle and driving post-training research. Furthermore, you'll collaborate closely with data, infrastructure, and adjacent modeling teams to translate research findings into durable model improvements. What You’ll Do: - Build and maintain scalable systems for ingesting, preprocessing, and delivering large-scale video data for model training - Design and scale distributed data pipelines for preprocessing, dataset generation, and repeated dataset refreshes - Own workflow orchestration, job scheduling, monitoring, and failure recovery for large-scale data processing jobs - Implement and maintain containerized pipeline infrastructure using Kubernetes or equivalent orchestration systems - Optimize cloud-based data storage and movement across providers (AWS, GCS, or Azure) for cost, throughput, and operational efficiency - Define and implement best practices for dataset storage layout, versioning, caching, retention, and access patterns - Build tooling to support deduplication workflows at scale, including near-dedup pipelines over large video corpora - Research and develop distillation methods for large-scale diffusion and flow-based video generation models, including guidance distillation and adversarial distillation, with a focus on preserving or improving generation quality while reducing inference cost - Develop reward models and preference-based fine-tuning pipelines that align video generation quality with human judgments across dimensions such as aesthetics, motion quality, and prompt adherence - Analyze the relationship between base model behavior and post-training outcomes, and work with the foundation model team to inform pretraining decisions accordingly What You’ll Bring: - Strong hands-on experience building or scaling large-scale data systems or pipelines for machine learning workflows - Experience with distributed data processing frameworks such as PySpark or Ray, and orchestration tools such as Airflow or equivalent - Familiarity with containerization and container orchestration, including Docker and Kubernetes - Experience working with cloud-based data storage and compute (AWS, GCS, and/or Azure), including tradeoffs around cost, throughput, storage layout, and access patterns - Familiarity with video and media processing tools such as FFmpeg, PyAV, DALI, or OpenCV - Familiarity with multimodal or media data, including video, image, text, and audio - Strong research background in post-training methods for large-scale diffusion or flow-based generative models, with deep hands-on experience in distillation across both inference efficiency and quality preservation - Experience with reward modeling or preference-based fine-tuning for generative models, including RLHF, DPO or equivalent alignment approaches - Solid understanding of the interplay between pretraining and post-training, and how base model properties affect distillation and fine-tuning outcomes - Proficiency in Python and modern machine learning frameworks, with a strong preference for PyTorch or JAX - Track record of independent research, with the ability to drive projects from initial idea through experimental validation - Publications at top-tier venues (NeurIPS, ICML, ICLR, CVPR, ICCV, ECCV) preferred - Good understanding of the practical challenges involved in building reliable, scalable, and reproducible data workflows for machine learning systems Benefits We Offer: - Competitive salary and generous company equity - Personal time off and paid holidays - Health insurance - Global travel insurance: Covers you when traveling internationally - Monthly spending stipend: $500 (~S$635) - Equipment: All equipment needed for your home office

Location: Paris; London ABOUT MISTRAL Mistral provides full-stack AI solutions: from frontier models to developer tools, applications, and compute. We partner with enterprises tackling the hardest problems across high-stakes industries like finance, manufacturing, defense, healthcare, and the public sector, co-creating customized AI systems that they can run on their terms. We are a dynamic, collaborative team passionate about AI and its potential to transform society. Our diverse workforce thrives in competitive environments and is committed to driving innovation. Our teams are distributed between Europe, North America, Asia and the Middle East. We are creative, low-ego and team-spirited. THE ROLE As a Research Engineer on the Latency & Throughput team, you will push the frontier of LLM inference by researching and prototyping techniques that reduce latency and increase throughput — from speculative decoding and quantization to KV compression and new attention variants. The team sits at the intersection of research and production: you will take ideas from recent literature, form hypotheses, run experiments, and ship the techniques that prove out into our serving stack. This role sits between our inference engineers, who own the production serving stack, and our research scientists, who focus on paper-level contributions. You will be more science-focused than a pure engineer: you are expected to generate new ideas, keep up with the state of the art, and implement proof-of-concepts that bridge the gap between academic results and production impact. Some techniques you explore will be lossy or require integration into training — retraining models, tuning architectures, or training auxiliary models to improve inference. You will work closely with the Inference Foundation and Kernels teams, who turn your validated techniques into production recipes, and with pretraining and RL teams when your work touches model training. Success is measured by demonstrated latency or throughput gains, first at small scale, then adopted in production. What You Will Do - Research and prototype inference techniques — speculative decoding, quantization, KV compression, watermarking, and new model architectures — by reading recent literature and forming testable hypotheses. - Design and run experiments to validate or refute ideas, iterating quickly on small-scale setups before proposing production integration. Demonstrate clear latency or throughput gains without unacceptable quality loss. - Collaborate with the Inference Foundation team to integrate validated techniques into the vLLM-based serving stack. - Partner with pretraining and RL teams when techniques require co-training, retraining, or training auxiliary models to improve inference. Build the right tooling and validation platforms to facilitate AI-driven inference research. What We're Looking For - A strong ML background with solid knowledge of LLM training techniques — architecture, tokenization, fine-tuning, and evaluation. - Hands-on experience with Python and PyTorch; you write clean, experimental code and iterate fast. JAX is acceptable with strong Python proficiency. C++ and Rust are not required, CUDA and GPU knowledge is expected but hands-on experience is not required. - Research experience of some form: a thesis, publications, a dedicated research project, or efficiency-oriented work in a lab or industry research setting. A PhD is not required, but we are looking for genuine research depth. - A track record of conducting experiments: forming hypotheses, designing controlled comparisons, and drawing clear conclusions from results. - Familiarity with the state of the art in LLM inference — you know what has been tried, what works, and where the open problems are. - Self-starter, low-ego, and collaborative. You are comfortable with ambiguity, negative results, and are eager to take matters into your own hands. - Inference or serving experience is a plus, not a requirement. Prior work with vLLM, production LLM deployments, or serving systems will help you ramp up faster. Location and Remote This role is based in our Paris or London offices (hybrid). We will prioritize candidates who reside in the Paris area or are open to relocating. We strongly believe in the value of in-person collaboration to foster strong relationships and seamless communication within our team. For senior profiles, we will also consider remote candidates based in France. In that case, we ask all new hires to visit our Paris office: - For the first week of their onboarding (accommodation and travel covered) - Then at least 3 days per month WHAT WE OFFER We offer a comprehensive benefits package designed to support your well-being, growth, and work-life balance. Benefits vary by country and may include healthcare coverage, parental leave, retirement plans, relocation support, wellness programs, meal and transportation allowances, and other location-specific perks. For the most up-to-date details on benefits available in your location, please refer to our Benefits page https://app.notion.com/p/mistralai/Benefits-at-Mistral-36e6ba59a7fe836b93dd01737fcc27ef?source=copy_link. PRIVACY POLICY Your privacy matters to us. You can learn more about how we handle your personal data in our Applicant Privacy Policy https://legal.mistral.ai/terms/applicant-privacy-policy.

Location: Paris; London Workplace: Hybrid ABOUT MISTRAL Mistral provides full-stack AI solutions: from frontier models to developer tools, applications, and compute. We partner with enterprises tackling the hardest problems across high-stakes industries like finance, manufacturing, defense, healthcare, and the public sector, co-creating customized AI systems that they can run on their terms. We are a dynamic, collaborative team passionate about AI and its potential to transform society. Our diverse workforce thrives in competitive environments and is committed to driving innovation. Our teams are distributed between Europe, North America, Asia and the Middle East. We are creative, low-ego and team-spirited. Role Summary We are seeking a talented and experienced Research Engineer to join our newly forming Search and Services Platform team within Research Platform. You will build and operate the large-scale search and services that our models rely on as tools — from self-hosted, web-scale, search indexes to knowledge graphs and databases served through MCP. You will work closely with researchers and with the pretraining data, sandbox and data platform teams to make these services reliable, scalable and fast, on infrastructure measured in hundreds of terabytes of NVMe storage and thousands of CPU cores to support both data generation and critical training runs. What you will do As a Research Engineer on the Search and Services team, your responsibilities will include (but may not be limited to): - Designing, building and operating large-scale search infrastructure: self-hosted Vespa clusters indexing web-scale corpora, managed indexes for smaller corpora, and vendor search endpoints. - Self-hosting and maintaining a wide range of structured data sources — knowledge graphs (Wikidata, UniProt, Materials Project), scholarly data (OpenAlex, OpenAIRE), statistical databases (Eurostat, World Bank, IMF, OECD), open data portals and more. - Designing interfaces and tools to expose to the models and agents to use the services in different ways (direct YQL or SPARQL queries, or more refined and constrained MCPs to mimic external services) to support data generation and agentic training. - Building the data pipelines that ingest, index and serve massive datasets, including indexing our own pretraining corpora for exploration, deduplication and eval-leakage checks. - Building strong relationships with researchers and adjacent teams (pretraining data,, sandbox, data platform), understanding their workflows and enabling them to achieve more by leveraging your expertise. - Communicating and producing documentation or any content that will help teams make the most of the services you'll build. About you - 5+ years of successful experience in search engineering, data infrastructure, or a similar SRE / platform role. - Proficiency in software development (Python, Go...) and programming best practices. - Hands-on experience with large-scale search or data systems (e.g. Vespa, Elasticsearch / OpenSearch, ClickHouse or similar). - Exposure to site reliability engineering: root cause analysis, in-production troubleshooting, on-call rotations...) - Exposure to infrastructure management: containerization, orchestration (Kubernetes), infra-as-code, monitoring, logging, alerting, observability...). - Experience designing and operating data pipelines that address large amounts of data. - Excellent problem-solving and communication skills (ability to contextualizing, gauging risks and getting buy-in for high stakes and impactful solutions). - Ownership, high agency and constantly seeking to learn and improving things for others. - Autonomous, self-driven and able to work well in a fast-paced startup environment. - Low ego and team spirit mindset. - Your application will be all the more interesting if you also have: - First-hand Vespa (or equivalent large-scale information retrieval) experience. - Experience building or operating MCP servers / tool-serving infrastructure for LLMs. - Familiarity with knowledge graphs, RDF / SPARQL, or large structured databases (PostgreSQL, ClickHouse, DuckDB...). - Experience with large bare-metal infrastructure or HPC environments. WHAT WE OFFER We offer a comprehensive benefits package designed to support your well-being, growth, and work-life balance. Benefits vary by country and may include healthcare coverage, parental leave, retirement plans, relocation support, wellness programs, meal and transportation allowances, and other location-specific perks. For the most up-to-date details on benefits available in your location, please refer to our Benefits page https://app.notion.com/p/mistralai/Benefits-at-Mistral-36e6ba59a7fe836b93dd01737fcc27ef?source=copy_link. PRIVACY POLICY Your privacy matters to us. You can learn more about how we handle your personal data in our Applicant Privacy Policy https://legal.mistral.ai/terms/applicant-privacy-policy.

Location: US, San Diego, CA Workplace: OnSite Here at Humanoid, we believe in a future where robots amplify human potential. That’s why we’ve set out on a mission to build the world’s most capable, commercially-scalable, and safe humanoid robots. We’re bringing that mission to life with HMND‑01 - our rapidly developed humanoid platform being deployed in real industrial environments - and we’re growing the team to take it even further. ABOUT THE ROLE As an Autonomy Engineer focused in VLA Pre-training, you will work on all aspects of training capable policies. You'll pre-train base models on a diverse, multi-embodiment corpus of trajectories, fine-tune policies to excel at specific tasks, shape data collection processes, and explore effective ways to generate and use synthetic data. This is primarily a deep learning role, so we're looking for experience solving real-world problems with modern neural networks. Robotics experience isn't strictly required, but if you're coming from outside the field, be prepared to get up to speed on a new domain quickly.   WHAT YOU'LL DO - Post-train policies via behavior cloning and RL; own the full loop from data to deployment. - Partner with the Data Collection team to drive collecting new data: specify what good data looks like, identify failure modes, ensure diversity and coverage. - Work closely with external partners to ensure steady supply of high-quality pretraining-scale data. - Run pre-/mid-/post-training on VLA stack; explore new modalities and architecture changes. - Build and maintain continuous pipelines: ingest synthetic data and teleop logs, version them, apply weak‑supervision labelling, curate balanced datasets, and auto‑surface fresh failure cases into retraining. - Work with MLOps & Data Platform teams to scale distributed training and optimize models for real‑time edge inference. WHAT WE'RE LOOKING FOR - 3+ years building deep‑learning systems (industry or research) with shipped models or published artifacts to show for it. - Deep hands‑on experience with at least one of: LLMs, VLMs, or image/video generative models — architecture, training, and inference. - Experience with deep learning infrastructure: streaming datasets, checkpointing & state management, distributed training strategies. - Strong Python + PyTorch/JAX; you can profile, debug numerics, and write maintainable research code. - Familiarity with modern software engineering practices. - You document experiments clearly and communicate trade‑offs crisply. NICE TO HAVE - Robotics or autonomous driving experience. - Experience applying RL to LLMs or robotics. - Experience with VLA (vision-language-action) models. - Proven productization of deep nets (latency/throughput constraints, telemetry, on‑device optimization). - Publications at top-tier deep learning conferences or equivalent open‑source contributions. - Familiarity with OpenVLA, Physical Intelligence (π) models, or similar open source VLA frameworks. WHAT WE OFFER - Comprehensive health coverage for US‑based employees, including fully paid medical, dental, and vision insurance, with virtual care and employee assistance resources. - Meaningful time off to rest and recharge: 23 days of PTO (accrued), separate sick leave, and paid company holidays. - 401(k) retirement plan with employer match. - Equity included–we believe builders should share in what they build. - Free daily catered lunch, snacks, and drinks in‑office. - Collaboration with top‑tier engineers, researchers, and product experts in AI and robotics. - Freedom to influence the product and own key initiatives. For this role in California, the expected base salary range is $180,000–$300,000 USD per year; your placement in that range depends on how your experience maps to our internal leveling.

Location: US, San Diego, CA Workplace: OnSite Here at Humanoid, we believe in a future where robots amplify human potential. That’s why we’ve set out on a mission to build the world’s most capable, commercially-scalable, and safe humanoid robots. We’re bringing that mission to life with HMND‑01 - our rapidly developed humanoid platform being deployed in real industrial environments - and we’re growing the team to take it even further. About The Role As Director of Autonomy at Humanoid, you will lead the team that teaches our bipedal platform to perform locomanipulation tasks in homes and service environments: tidying a room, loading and unloading a dishwasher, putting groceries away, folding laundry, and similar tasks. The output of your team is capabilities that run reliably on real robots and pass our evals, not papers or one-off demos. This is a player-coach role: you own the roadmap, hiring, and delivery, while also training policies, curating data, and debugging failures on the robot yourself. That hands-on work is how you lead: it's how you find what slows the team down and fix it for everyone. This role sits at the intersection of applied deep learning, robotics, and people leadership, working closely with the Controls team on whole-body teleoperation and combining RL controllers with learned policies, and with the Core Technology team on foundation model pre-training for five-finger dexterity and RL methods. If you have hands-on experience with real robot hardware, a track record in applied deep learning, and have led engineers doing this kind of work, we want to hear from you. What You'll Do - Own The Capability Roadmap: Work with product to identify use-case priority, choose the execution strategy, define the evals, and deliver on the agreed timeline. - Lead The Team: Hire, develop, and manage a team of robot learning engineers; set engineering and experimental standards; run day-to-day prioritization. - Co-Own The System 0/1 Interface: Agree with the Controls team on what the policy commands (end-effector targets, root velocity, pelvis height, foot targets, hand joint targets), debug failures across that boundary, and push requirements into the whole-body control roadmap. - Partner On Data Strategy: Specify what good data looks like for bipedal locomanipulation with data collection and external acquisition organizations, ensure diversity and coverage, iterate on operator instructions, and supervise collection on the biped. - Drive Sim-To-Real Transfer: Work with the Simulation team to set up RL training on the digital twin for bipedal tasks, then iterate on reward and simulation quality to ensure transfer to the real world. - Shape Base Model Development: Feed bipedal requirements — embodiment coverage in the pretraining mixture, action representations, latency budgets — into the VLA Core Technology team's base model development, and adopt new base models as they land. - Inform Hardware Design: Interface with the Hardware Design team so findings about hands, arms, and legs on the current generation are reflected in future designs. - Report On Progress: Communicate progress to leadership in terms of capabilities working on the robot, eval results, and failure analysis. WHAT WE'RE LOOKING FOR - 5+ years working on robots (industry or research) with shipped artifacts to show for it, including learned manipulation or locomanipulation policies deployed on real hardware. - Experience leading a team of 5+ engineers doing applied robot learning, including hiring and people management, with capabilities delivered to a deadline. - Hands-on today: you've trained and debugged policies within the last year and intend to keep doing so, and can curate data, post-train a policy, and diagnose failures without delegating that work. - Experience post-training VLA-class policies (autoregressive, diffusion, or flow-matching based) via behavior cloning and/or RL. - Solid understanding of the modern teleoperation and low-level control stack — whole-body control, inverse kinematics, RL locomotion controllers — and the ability to tell whether a failure lives in the policy, the controller, or the data. - Familiarity with deep learning infrastructure: streaming datasets, checkpointing and state management, distributed training, PyTorch or JAX. - Strong grounding in modern software engineering practices, with the ability to document experiments clearly and communicate trade-offs crisply to engineers and leadership. NICE TO HAVE - Experience with multi-fingered dexterous hands: teleoperation and retargeting, tactile sensing, in-hand manipulation. - Experience with legged platforms: bipedal locomotion, whole-body control, humanoid teleoperation. - Experience applying RL to robotics problems in simulation or on real robots. - Familiarity with Physical Intelligence (π) models, NVIDIA GR00T, or similar open VLA frameworks. - Publications at top-tier robotics or deep learning conferences (CoRL, RSS, ICRA, NeurIPS) or equivalent open-source contributions. WHAT WE OFFER - Comprehensive health coverage for US‑based employees, including fully paid medical, dental, and vision insurance, with virtual care and employee assistance resources. - Meaningful time off to rest and recharge: 23 days of PTO (accrued), separate sick leave, and paid company holidays. - 401(k) retirement plan with 4% employer match. - Competitive equity: stock options with meaningful upside as we scale. - Free daily catered lunch, snacks, and drinks in‑office. - Collaboration with top‑tier engineers, researchers, and product experts in AI and robotics. - Freedom to influence the product and own key initiatives. For this role in California, the expected base salary range is $300,000-$450,000 USD per year; your placement in that range depends on how experience maps to our internal leveling.

Location: San Francisco, CA Workplace: Hybrid About Niantic Spatial AI has learned from almost everything online, but it still struggles with the physical world, where 99.5% of economic activity happens. It doesn’t know where things are, how spaces fit together, or what changes over time. That’s one of the biggest unsolved problems in AI, and it’s the one we’re solving at Niantic Spatial. We build spatial AI models and services so people and machines can reason, learn and act safely in the physical world. We turn camera imagery, video, sensor data and archives into accurate, AI-readable representations of real places that robots can train in and machines can use to know exactly where they are. In physical AI, mistakes have real consequences. That’s why we build AI grounded in measured reality, not guesswork, so the people and machines relying on it can trust it to get things right. Our team built some of the world’s most widely used maps. Now we’re building something bigger: a living, measured record of the real world that machines can query, learn from and use to understand how places change over time. If you want to help AI understand the world we actually live in, come build with us. About the Team Localization is where the map meets the moment. Our localization team builds the living geospatial world model and the centimeter-level Visual Positioning System (VPS) that lets people and robots know exactly where they are, from any camera, almost anywhere on Earth. That system is learned, not hand-built. Our localization stack has moved from classical geometry to visual models trained on billions of posed images — models that have to be accurate to the centimeter, robust to lighting and seasons and clutter, and fast enough to run in production for every query. This team sits at the junction of research and production: we take what works in a paper, make it work on our data, and make it work for customers. About the Role We're hiring an AI Engineer, Computer Vision, to contribute to the deep learning models at the heart of our localization system. Working alongside experienced researchers and engineers, you'll help train and iterate on the visual models — feature extractors, matchers, pose regressors, and retrieval models — that turn a single image into a precise position in the world. This is a full-loop environment. You'll get exposure to the entire pipeline: the data the model learns on, training, evaluation, and the service that runs it in production. You'll be surrounded by people who care deeply about correctness — where a loss curve going down is not the same as a system getting better — and you'll grow into that rigor quickly. What You'll Do - Train and Evaluate Models — Implement, train, and evaluate visual localization models in PyTorch: local feature extraction and matching, image retrieval, pose regression, and scene-coordinate regression. Iterate on architectures, loss functions, and training loops to improve accuracy on real production captures. - Build Data Pipelines — Contribute to high-throughput ingestion and preprocessing pipelines that turn raw imagery, poses, and 3D point clouds into training-ready datasets, including curation, sampling, and augmentation. - Run Experiments at Scale — Run training jobs across GPU environments on our 30-billion-image corpus, working with senior engineers and ML Infrastructure to keep experiments reproducible and results interpretable. - Extend the Benchmark — Help build and maintain the evaluation suite for localization — pose accuracy, recall across lighting and season changes, and latency under production load — so results are comparable across model versions. - Support Production Deployment — Work with senior engineers to take models from checkpoint toward serving: export, inference optimization, and monitoring once live. Learn what it means to own a model end-to-end, not just hand off a checkpoint. - Bridge Research and Product — Work with the Research team and Real-World Test Lab to understand how field failures translate into training data and benchmark improvements. What You'll Bring - Experience training deep learning models for computer vision tasks in PyTorch — through coursework, research, internships, or early industry work. - Foundational understanding of at least one of: visual localization, learned feature matching, image retrieval, pose estimation, SLAM, or structure-from-motion, and the 3D geometry underneath them (camera models, epipolar geometry, pose parameterizations). - Comfort working with large image or 3D datasets — preprocessing, curation, and understanding how data quality affects model behavior. - Familiarity with GPU-based training and the ability to read and reason about training runs — identifying whether a bottleneck is in data loading, compute, or model architecture. - Strong Python skills and a habit of writing code that others can rerun and build on. - A bachelor's or master's degree in computer science, computer vision, robotics, or a related field, or equivalent practical experience. Nice to Have - Published or contributed to work on learned localization, visual place recognition, scene-coordinate regression, or vision foundation models (CVPR, ECCV, ICCV, NeurIPS, 3DV, or equivalent). - Experience with large-scale self-supervised or multi-view pretraining, or fine-tuning vision foundation models for geometric tasks. - Exposure to inference optimization for edge or mobile hardware (TensorRT, ONNX, CoreML, or similar). - Experience with petabyte-scale spatial data: point clouds, meshes, posed image collections, or geospatial indexing. - Background in robotics, augmented reality, 3D reconstruction, digital twins, or autonomous systems. - Prior industry or research internship experience shipping or contributing to a production CV system. Compensation & Benefits The expected salary range for this role is $165,600 - $221,000 per year. Compensation also includes an annual bonus, equity, and a comprehensive benefits package including medical, dental, and vision coverage, 401(k), and more. Location & Work Model This role is based in our San Francisco, CA office on a hybrid schedule with a minimum of 3 days per week in office. Inclusive Application We know the strongest candidates don't always tick every box. If you're excited about this role and believe you could do it well, we encourage you to apply even if your experience doesn't match every qualification listed — you may be exactly who we're looking for. Equal Opportunity Niantic Spatial is an equal opportunity employer. Individuals seeking employment at Niantic Spatial are considered without regard to race, color, ancestry, national origin, religion, creed, age, gender (including pregnancy, childbirth, breastfeeding or related medical conditions), marital status, physical or mental disability, medical condition, genetic information, military or veteran status, gender identity, gender expression, sexual orientation, or any other protected category under applicable laws. Niantic Spatial will also consider qualified applicants with criminal histories in accordance with applicable laws. Please contact your recruiter if you want to request an accommodation for the job application or interview process. Candidate Privacy I understand that by submitting my job application, the information I provide as part of that application will be used in accordance with Niantic Spatial's Privacy Notice for Job Applicants and Candidates https://www.nianticspatial.com/applicant-privacy-notice.

Location: San Francisco, CA | New York City, NY | Seattle, WA About Anthropic 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 Frontier AI is on track to be among the most consequential and most adversarially-targeted technology in the world. The capability curve is steep, the adversaries who want these systems are extremely well-resourced, and the security bar this will eventually require is well beyond where the industry operates today. Incremental hardening alone is not going to close that gap, so we need breakthroughs and a group of people to go find them. Security Labs is that team. We run a portfolio of high-risk, high-expected-value security projects: the work that seems impractical until someone optimistic and stubborn enough actually tries it. Projects run on the order of weeks rather than quarters, and each one is either handed off to the Anthropic team that will own it in production or wound down with a writeup of what we learned. We expect a meaningful fraction of our bets not to land. This team is intentionally scrappy and ambitious. We're intentionally iterating fast on high-risk research bets and we expect a meaningful fraction not to land. If a 30% project success rate sounds uncomfortable, or uncharted and chaotic territory isn’t both frightening and exciting, this probably isn't the right fit. There are other places in Anthropic Security doing important work with more structure, less risk, and more certain paths to positive outcomes. The questions we're trying to answer include: Can our core research workflows survive extreme isolation? Can we replace trust with cryptographic guarantees? Can the models themselves become our most effective security control? What would it actually take to defend against a tier-1 state adversary, and how much of that can we build now? Who we're looking for. We're hiring generalists with rare depth. You're a strong software engineer as a baseline, and on top of that you've gone deep in at least one area most engineers don't get near: firmware or hardware security, applied cryptography, OS / kernel / hypervisor internals, formal methods, reverse engineering, or high-assurance and cross-domain systems. You've built things under your own direction, you're comfortable jumping layers when the problem demands it, and you'd rather take a swing at something that might not work than ship the safe incremental thing. You think the trajectory of AI matters a great deal, you're not comfortable with how the security side of it is going by default, and you'd rather be on the inside building the answer than watching from outside. Current Project Areas The portfolio changes as we learn. The kinds of bets currently in flight or queued: Standing up a prototype high-assurance research cluster: running real Anthropic training and research workloads under extreme isolation and physical security controls, and finding out exactly where productivity breaks and what we'd need to invent to get it back Provable inference: cryptographic verification (zero-knowledge proofs, attestation chains) that a given output came from a specific model running specific code, replacing "trust us" with math Build hypervisor-isolated microVM workload isolation across the fleet Compiling an ML kernel through a formally verified pipeline where every lowering step carries a machine-checked proof of equivalence, making compilation-layer sabotage mathematically detectable Regenerating clusters: automation that spins up a purpose-built cell, runs a workload, and tears the whole thing down on a TTL measured in hours, so attacker persistence has an expiry date Using Claude itself to drive security work end to end: threat modeling new compute platforms, rewriting critical services to zero external dependencies, running the test equipment that validates what hardware datasheets claim Part of your job is deciding what comes next. We hire people we trust to pick good bets, and project selection is owned by the engineers doing the work. Key responsibilities Own Security Labs projects end to end. You'll scope the bet, build the prototype, run it against real workloads, and bring it to either a hand-off or a documented exit Stand up novel security infrastructure fast (isolated clusters, attestation chains, hypervisor and runtime work, verification tooling) optimizing for what we learn rather than for permanence Find the receiving team early, bring them along while you build, and hand them something they actually want to own Work embedded with research and infrastructure teams (Pretraining, RL, Inference, Compute) to test whether their workflows survive what you're proposing, and document precisely where they don't Turn experimental results into short writeups that shape Anthropic's long-term security architecture, and into costed contingency plans we could execute on short notice Help pick the next round of bets and influence the industry to get better along the way Minimum qualifications Genuinely care about where AI is heading and want to work on the security problems that determine whether it goes well. This is the most important thing on this list Have real depth in at least one area most software engineers don't touch (e.g. firmware or hardware security, applied cryptography, OS / kernel / hypervisor internals, formal methods and verification, reverse engineering and exploit development, or high-assurance / cross-domain systems) Have built and shipped things under your own direction. Maybe you founded a company or research group, maintained an open-source project other people depend on, or shipped research that changed how people in your field work. We weight this far more than where you've worked or for how long Have a track record of choosing the problem yourself and seeing it through, rather than only executing a plan someone else handed you Are comfortable jumping between domains and layers of the stack when the problem calls for it, and have actually done so Have run prototypes or experiments where the goal was answering a hard question rather than shipping a permanent system, including ones that didn't pan out Write clearly enough to turn weeks of work into a couple of pages someone can act on Change your mind when the evidence says to, and are fine being the least-expert person in a room full of specialists Care about defense. Plenty of folks here come from offense and that background is valuable, but what you actually want to spend your time on now is making systems hold up Are a strong programmer (Python plus at least one of Rust, Go, or C/C++) and can stand up real infrastructure without that being the interesting part of your week Preferred qualifications Experience inside airgapped or high-side environments (classified networks, cross-domain solutions, ICS/SCADA, financial trading infrastructure) and the operational realities of working in them A background in offensive security, red teaming, or vulnerability research, with calibrated intuitions for which threats actually matter Familiarity with ML infrastructure (training pipelines, distributed schedulers, inference serving, accelerator hardware) sufficient for grounded conversations with researchers about what their workloads actually need A history of working in environments built around rapid iteration rather than rigid change control: startups, applied research groups, independent consulting, small security shops 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. Annual Salary: $320,000 — $485,000 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 encourage you to apply even if you do not believe you meet every single qualification. Not all strong candidates will meet every single qualification as listed.  Research shows that people who identify as being from underrepresented groups are more prone to experiencing imposter syndrome and doubting the strength of their candidacy, so we urge you not to exclude yourself prematurely and to submit an application if you're interested in this work. We think AI systems like the ones we're building have enormous social and ethical implications. We think this makes representation even more important, and we strive to include a range of diverse perspectives on our team. Your safety matters to us. To protect yourself from potential scams, remember that Anthropic recruiters only contact you from @anthropic.com email addresses. In some cases, we may partner with vetted recruiting agencies who will identify themselves as working on behalf of Anthropic. Be cautious of emails from other domains. Legitimate Anthropic recruiters will never ask for money, fees, or banking information before your first day. If you're ever unsure about a communication, don't click any links—visit anthropic.com/careers directly for confirmed position openings. 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.

Location: Москва Привет! Это GigaChat Vision — команда, которая делает полный цикл обучения VLM-моделей. Ищем сильного инженера в RL-направление: человека, который будет драйвить ключевые исследования и разработки, влиять на качество моделей и доводить сложные идеи до работающих решений. _Первый этап отбора на эту вакансию – общение с AI-рекрутером. После отклика вам на почту и в чат на платформе HeadHunter придет приглашение пройти первичное интервью с ГигаРекрутером в Telegram. Диалог займёт примерно 10 минут. Его задача — уточнить недостающие детали и ускорить рассмотрение вашей кандидатуры. ГигаРекрутер только начинает свой путь, поэтому просим относиться с пониманием. Ваш опыт и участие помогут сделать его удобным и полезным!_ Обязанности * Разрабатывать и улучшать RL-подходы для обучения VLM/LLM-моделей: выбор алгоритмов, постановка экспериментов и анализ результатов; * Проектировать reward-функции и пайплайны обучения, выстраивать стратегии масштабирования под разные домены и сценарии, напрямую влияя на метрики качества; * Определять требования к данным для RL: участвовать в построении пайплайнов сбора, фильтрации и подготовки датасетов; * Развивать систему оценки reasoning-качества: внедрять и улучшать метрики в существующем eval-фреймворке, предлагать новые способы измерения качества под новые возможности модели; * Работать на стыке с Pretrain / SFT / Infra: обеспечивать согласованность решений и перенос экспериментов в продакшн-пайплайн; * Следить за состоянием области и переводить свежие идеи из статей в проверяемые эксперименты и инженерные решения; * Глубоко погружаться в технические задачи: дебажить эксперименты, разбирать аномалии в обучении, находить узкие места и доводить решения до результата; * Делиться экспертизой с командой: участвовать в код-ревью, помогать улучшать качество решений и подходы к экспериментам; * Развивать инфраструктуру для RL. Требования * Глубокое понимание RL для LLM/VLM (RLHF, GRPO, PPO) и практический опыт; * Понимание полного цикла обучения VLM/LLM (pretrain → SFT → RL) и того, как решения на каждом этапе влияют на финальные метрики; * Опыт с распределённым обучением (DeepSpeed, FSDP) и inference-фреймворками. * Сильный практический опыт постановки, проведения и анализа RL-экспериментов; * Умение работать в условиях неопределённости и самостоятельно двигать сложные технические задачи; * Умение выстраивать процессы в условиях неопределённости; * Опыт взаимодействия со смежными командами и стейкхолдерами; * Системное мышление: способность видеть картину целиком — от данных и reward-дизайна до eval и продакшн-метрик. **Будет плюсом:** * Опыт менторинга или технического лидерства в рамках отдельных проектов; * Публикации или open-source вклад в области RL/LLM/VLM; * Опыт вывода RL-обученных моделей в продакшн и поддержки их качества. Условия * крупнейшее DS&AI community — более 600 DS-специалистов банка * дайджест о самых последних разработках в области DS&AI и отчеты с крупнейших конференций мира * возможность быть соавтором НИРов и статей для международных конференций * возможность выбрать удобный формат работы: гибрид или офис * ежегодный пересмотр зарплаты, годовая премия * корпоративный спортзал и зоны отдыха * более 400 образовательных программ СберУниверситета для профессионального и карьерного развития * расширенный ДМС, льготное страхование для семьи и корпоративная пенсионная программа * ипотека выгоднее до 7% для каждого сотрудника * бесплатная подписка СберПрайм+, скидки на продукты компаний-партнеров * вознаграждение за рекомендацию друзей в команду Сбера.

Location: Москва Привет! Это GigaChat Vision — команда, которая делает полный цикл обучения VLM-моделей. Ищем сильного senior-инженера в RL-направление: человека, который будет драйвить ключевые исследования и разработки, влиять на качество моделей и доводить сложные идеи до работающих решений. _Первый этап отбора на эту вакансию – общение с AI-рекрутером. После отклика вам на почту и в чат на платформе HeadHunter придет приглашение пройти первичное интервью с ГигаРекрутером в Telegram. Диалог займёт примерно 10 минут. Его задача — уточнить недостающие детали и ускорить рассмотрение вашей кандидатуры. ГигаРекрутер только начинает свой путь, поэтому просим относиться с пониманием. Ваш опыт и участие помогут сделать его удобным и полезным!_ Обязанности * Разрабатывать и улучшать RL-подходы для обучения VLM/LLM-моделей: выбор алгоритмов, постановка экспериментов и анализ результатов; * Проектировать reward-функции и пайплайны обучения, выстраивать стратегии масштабирования под разные домены и сценарии, напрямую влияя на метрики качества; * Определять требования к данным для RL: участвовать в построении пайплайнов сбора, фильтрации и подготовки датасетов; * Развивать систему оценки reasoning-качества: внедрять и улучшать метрики в существующем eval-фреймворке, предлагать новые способы измерения качества под новые возможности модели; * Работать на стыке с Pretrain / SFT / Infra: обеспечивать согласованность решений и перенос экспериментов в продакшн-пайплайн; * Следить за состоянием области и переводить свежие идеи из статей в проверяемые эксперименты и инженерные решения; * Глубоко погружаться в технические задачи: дебажить эксперименты, разбирать аномалии в обучении, находить узкие места и доводить решения до результата; * Делиться экспертизой с командой: участвовать в код-ревью, помогать улучшать качество решений и подходы к экспериментам; * Развивать инфраструктуру для RL. Требования * Глубокое понимание RL для LLM/VLM (RLHF, GRPO, PPO) и практический опыт; * Понимание полного цикла обучения VLM/LLM (pretrain → SFT → RL) и того, как решения на каждом этапе влияют на финальные метрики; * Опыт с распределённым обучением (DeepSpeed, FSDP) и inference-фреймворками. * Сильный практический опыт постановки, проведения и анализа RL-экспериментов; * Умение работать в условиях неопределённости и самостоятельно двигать сложные технические задачи; * Умение выстраивать процессы в условиях неопределённости; * Опыт взаимодействия со смежными командами и стейкхолдерами; * Системное мышление: способность видеть картину целиком — от данных и reward-дизайна до eval и продакшн-метрик. **Будет плюсом:** * Опыт менторинга или технического лидерства в рамках отдельных проектов; * Публикации или open-source вклад в области RL/LLM/VLM; * Опыт вывода RL-обученных моделей в продакшн и поддержки их качества. Условия * крупнейшее DS&AI community — более 600 DS-специалистов банка * дайджест о самых последних разработках в области DS&AI и отчеты с крупнейших конференций мира * возможность быть соавтором НИРов и статей для международных конференций * возможность выбрать удобный формат работы: гибрид или офис * ежегодный пересмотр зарплаты, годовая премия * корпоративный спортзал и зоны отдыха * более 400 образовательных программ СберУниверситета для профессионального и карьерного развития * расширенный ДМС, льготное страхование для семьи и корпоративная пенсионная программа * ипотека выгоднее до 7% для каждого сотрудника * бесплатная подписка СберПрайм+, скидки на продукты компаний-партнеров * вознаграждение за рекомендацию друзей в команду Сбера.

Location: Москва Мы - команда GigaChat Pretrain Data, готовим pretrain данные для GigaChat и GigaChat Vision. Pretrain данные - это фундамент, с которого начинается путь современной LLM модели и то, от чего наиболее зависит ее итоговое качество. Сырых данных более 40Пб и основная задача заключается в том, чтобы из этого хаоса сделать датасет, на котором будет обучена лучшая LLM в России. Обязанности * генерировать синтетические данные: математика, код, произвольная синтетика с сидами - документами из Web * исследовать токенизацию и ее влияние на качество модели (возможно написание статей) * решать задачи кластеризации миллиардов документов * исследовать разные факторы, которыми обладают текстовые данные * генерировать Vision данные для прокачки VLM * разрабатывать новые алгоритмы парсинга HTML и исследовать его влияние на качество модели * исследовать зависимости между pretrain данными и agentic capabilities итоговой модели * разрабатывать стабильную инфраструктуру, которая будет поддерживать проведение сотен и тысяч экспериментов над данными. Требования * имеешь коммерческий релевантный опыт связанный с NLP или построением инфраструктуры для данных от двух лет. Будет плюсом: * навыки работы с генеративными AI-моделями; опыт создания AI-агентов и использования их в работе будет преимуществом * опыт использования GigaChat, Kandinsky и аналогов в продуктах, навыки создания и использования AI-агентов * инструментальное владение AI для анализа, генерации и автоматизации * опыт с MapReduce системами. Условия * комфортный современный офис рядом с м. Кутузовская * гибридный формат работы (2 дня в офисе, 3 дня на удалёнке) * ежегодный пересмотр зарплаты, годовая премия * корпоративный спортзал и зоны отдыха * система обучения для профессионального и карьерного развития * расширенный полис ДМС с первого дня работы и страхование для семьи * льготная программа ипотеки для сотрудников * бесплатная подписка СберПрайм+, скидки на продукты компаний-партнеров * вознаграждение за рекомендацию друзей в команду Сбера.