Research, Mid-Training

📋 Description

• • Research, Mid-Training role at Cognition focused on sharpening base model capabilities through data mix, synthetic data, annealing schedules, and context length extension to build foundational model performance for AI software engineers like Devin.
• • Day-to-day responsibilities include designing high-quality data mixtures, driving capability injection in coding and reasoning, developing synthetic data pipelines, optimizing learning rate schedules, extending context length, building evaluation metrics, and measuring scaling behavior of mid-training interventions.
• • Cognition is an applied AI lab building end-to-end software agents, known for Devin (the first AI software engineer) and Windsurf (an AI-native IDE), with a small, talent-dense team of competitive programmers, former founders, and researchers from top AI organizations.
• • The role offers the opportunity to make original contributions at the frontier of LLM training, work with large-scale GPU clusters, and directly impact deployed AI systems through rigorous empiricism and innovative research in a fast-moving, autonomous environment.