O Moat de Dados: RLVR, Dados Sintéticos e a Nova Economia do Pós-Treinamento

Esta palestra explora como os dados moldam as capacidades dos sistemas modernos de IA, com foco especial na etapa de pós-treinamento. Vamos discutir o avanço do Reinforcement Learning com Recompensas Verificáveis (RLVR), o papel dos dados sintéticos, dos humanos no loop e dos benchmarks na evolução da IA. Com base na experiência da Vetto AI na produção de dados para o desenvolvimento de modelos de fronteira, a sessão aborda tanto os fundamentos técnicos quanto as implicações estratégicas dessas transformações, oferecendo uma visão interna de onde está surgindo a próxima onda de capacidades em IA. CV. Resumido Roberta Antunes is an entrepreneur focused on building companies at the intersection of technology and new economic models. She is the co-founder of Vetto AI, a data intelligence company helping train and evaluate large language models. Before that, she was Partner and Chief Growth Officer at Hashdex, where she played a key role in building the institutional crypto market in Brazil and internationally. She also co-founded Hotel Urbano, one of the largest online travel platforms in Latin America. Rigel Bezerra de Melo is an AI Research Engineer at Vetto AI, where he works on improving how large language models are trained and evaluated through high-quality, structured data. Prior to Vetto, he gained experience at Google, where he worked on large-scale systems and developed a strong foundation in applied machine learning and software engineering. Rigel is particularly interested in how human expertise and real-world data can shape the next generation of AI systems.