Research Mid-Training (AI)

Location: San Francisco, United States; on-site

Company

Building Devin, an AI software engineer, and developing models that can reason about real-world tasks.

What you will do

• Design and iterate on high-quality data mixtures for late-stage and annealing training runs.
• Drive capability improvements in coding, mathematics, and long-horizon reasoning through data strategies and training interventions.
• Develop and evaluate synthetic data pipelines that generate training signals at scale.
• Research learning-rate schedules, warmup strategies, compute allocation, and context-length extension methods.
• Build evaluations that distinguish genuine capability improvements from benchmark overfitting.
• Measure how mid-training interventions scale with compute and data, developing new methods when existing approaches reach their limits.

Requirements

• Deep familiarity with the end-to-end LLM training pipeline, including pre-training data, optimization, architecture, and interactions between mid-training and post-training.
• Hands-on experience with continual pre-training, annealing, or late-stage data mixing for large models.
• Experience developing or evaluating synthetic data pipelines for capability improvement.
• Proficiency in Python and PyTorch, with the ability to debug distributed training at scale.
• Strong fundamentals in optimization, statistics, and machine learning theory, plus a track record of original contributions.
• Ability to work in ambiguous, fast-moving environments and prioritize demonstrated capability over credentials.

Culture & Benefits

• Small, highly selective team where research and product development move together.
• Research prototypes reach real deployment quickly.
• Large compute allocations with training jobs routinely running across thousands of GPUs.
• Environment emphasizing speed, autonomy, technical depth, and minimal process overhead.