Research Engineer, Science of Deep Learning

Underpinning Character.AI’s trajectory is a sequence of larger, more intelligent, and empathetic models. As these models get larger, it is increasingly important to make rigorous decisions when training and serving them: what architecture/data composition/hyperparameters? How do we optimize the model? How much ROI do we get from human input? This is especially the case when models are so large that we can only train them once.

Our team discovers the scientific understanding for building our models. We dig deep into modeling and optimization decisions, and use our empirical knowledge to inform the design and training, leaving no stone unturned in the quest for better and better models.

We are looking for candidates with an abiding interest in how large models work, and the technical skills to advance our understanding of them. Especially, we desire candidates with an unusual clarity of thought, and a capacity to think from first principles. We work closely with our data, pre-training, and post-training teams.

Here is a list of sample projects:

• Run thousands of ablation and scaling experiments to understand underlying mechanisms in models.

• Select model and architecture hyperparameters achieving best-possible tradeoffs in inference latency and model quality.

• Determine the incremental value of additional human input in post-training.

• A bachelor's degree in a quantitative field (Physics, Mathematics, Computer Science); PhD preferred.

• 2+ years experience in industry, training, evaluating, and meaningfully modifying models.

• Working experience with at least one of PyTorch, Jax, or TensorFlow (not just high level APIs like keras.train).

• An excellent understanding of probability theory, linear algebra, and stochastic processes.

• A healthy dose of skepticism about experimental results and interpretations, and a deep cynicism about published AI/ML research.

• O(1000) careful ablation experiments under your belt.

• Papers in Neurips/ICML/ICLR.

• Specialized knowledge in (stochastic) optimization, condensed matter physics

Founded in 2021 by AI pioneers Noam Shazeer and Daniel De Freitas, Character is a leading AI company offering personalized experiences through customizable AI 'Characters.' As one of the most widely used AI platforms worldwide, Character enables users to interact with AI tailored to their unique needs and preferences.

Noam co-invented core LLM tech and was recently honored as one of TIME's 100 Most Influential in AI. Daniel created LaMDA, the breakthrough conversational AI now powering Google's Bard.

In just two years, we achieved unicorn status and were named Google Play's AI App of the Year – a testament to our groundbreaking technology and vision.

Ready to shape the future of AGI? 🚀

At Character, we value diversity and welcome applicants from all backgrounds. As an equal opportunity employer, we firmly uphold a non-discrimination policy based on race, religion, national origin, gender, sexual orientation, age, veteran status, or disability. Your unique perspectives are vital to our success.