John Schulman (OpenAI Cofounder) - Reasoning, RLHF, & Plan for 2027 AGI
Dwarkesh Podcast
Post training as a company moat
Post training data enables companies to understand user needs and preferences, helping in developing complex models aligned with specific styles and moralities. Creating models post training requires skilled individuals, tacit knowledge, and organizational understanding. It involves a significant accumulation of R&D effort, making it non-trivial to replicate instantly. Companies focusing on serious pre-training efforts usually invest in post-training as well. However, the ability to distill or replicate models from others reduces the uniqueness of this post-training effort, although this method may be used by smaller players. This practice, while possibly violating terms of service and denting pride, can be a shortcut for newcomers to establish themselves in the field.
Remember Everything You Learn from Podcasts
