Rafi Ayub

Rafi Ayub started in computational psychiatry at Stanford, training vision models on neuroimaging data, before specializing in post-training for large language models. At Meta he founded torchtune and torchmultimodal, the open-source PyTorch libraries for fine-tuning language and vision models, and built the post-training infrastructure behind Llama 2, 3 and 4. At Anthropic he led custom model training for enterprise customers and helped drive benchmark gains across the Claude model family, curating data from vendors in finance, legal, healthcare and software engineering. In 2026 he left to write and research independently at South Park Commons on how AI is changing us from the inside, working toward a book at the intersection of AI, philosophy, psychology and spirituality. He also serves as a research advisor to AI startups exploring the same questions.