By applying to this role, you will be considered for Research Scientist roles across all teams at OpenAI.
About the Role
As a Research Scientist here, you will develop innovative machine learning techniques and advance the research agenda of the team you work on, while also collaborating with peers across the organization. We are looking for people who want to discover simple, generalizable ideas that work well even at large scale, and form part of a broader research vision that unifies the entire company.
We expect you to:
- Have a track record of coming up with new ideas or improving upon existing ideas in machine learning, demonstrated by accomplishments such as first author publications or projects
- Possess the ability to own and pursue a research agenda, including choosing impactful research problems and autonomously carrying out long-running projects
- Be excited about OpenAI’s approach to research
Nice to have:
- Interested in and thoughtful about the impacts of AI technology
- Past experience in creating high-performance implementations of deep learning algorithms
Below are some of our current teams and the work you may do on them.
- Algorithms: Conduct exploratory research and drive algorithmic and architectural advances on the critical path to AGI. Algos has been responsible for a number of high profile OpenAI releases including GPT-2, Image GPT, Jukebox, DALL-E.
- Alignment: Fine-tune large language models to perform tasks in accordance with the user's intentions. Explore learning from human feedback and assisting humans evaluating AI. Recent projects: InstructGPT, Book summarization
- Applied AI Research: Develop innovative solutions and open-ended research to solve real-world problems. Care about applications driven by user feedback and long-term research with significant impact on products, while keeping a high safety product standard. Recent project: Text and Code Embeddings in OpenAI API
- Code Generation: Research and develop AI programmers, the neural models that write, debug and improve computer programs. Creating AI which can solve hard symbolic reasoning problems is one of the most difficult problems in modern deep learning and you can attack it head-on. Seek new ways for increasingly powerful AI systems to interact with the world through a very general interface of computer code. Recent project: Codex that powered the creation of GitHub Copilot.
- Language: Develop GPT advancements by exploring and forecasting future capabilities and resource needs. Seek a deep understanding of the scaling of our models across multiple orders of magnitude to optimize for best-case performance at the largest scale. Value impact over novelty – Team has found outstanding results from solid engineering with good design, implementation, and benchmarking.
- Mathgen: Help models learn to solve problems in informal natural text and prove theorems in formal languages like Lean. Large language models are known to make false claims that sound plausible, while mathematics requires the utmost rigor so training models that can reason robustly is a critical step on the path to AGI. Recent projects: Training Verifiers and Solving Olympiad Problems.
- Scaling: Build the model training software stack, solving problems at all layers of the stack including iteration speed, observability, compute efficiency, correctness, and fault detection and recovery. Scaling owns the engineering and research required to harness custom-built hyperscale supercomputers, the latest algorithmic improvements, and massive datasets to train AI models of unprecedented capability.
- Science of Deep Learning: Explore and understand the dynamics of training large models to guide both the training of current models and the trajectory of the next ones.