Interview, Fireside Chat
How OpenAI Builds for 800 Million Weekly Users: Model Specialization and Fine-Tuning
- ChatGPT's user base, currently at 800 million monthly active users, is projected to grow to even higher levels beyond the present.
- By 2030 or the near future, the industry is expected to transition away from a single dominant model toward a proliferation of specialized models and varied model types.
- OpenAI anticipates benefiting significantly from the expansion of the AI ecosystem, including increased activity and solutions provided by other companies, operating on a "rising tide" philosophy rather than winner-take-all dynamics.
- The company plans to maintain usage-based pricing on its API for quite a long time due to infrastructure complexity, though outcome-based pricing is being monitored as a potential future shift.
- Future advancements in instruction following are expected to render prompt engineering obsolete, enabling deterministic execution of multi-step processes.
- While fine-tuning currently relies on offline data, the reinforcement fine-tuning API is expected to drive paradigm shifts toward "sofa level" improvements for specific use cases, with potential benefits including discounted inference and free training for data contributors.
- The industry is projected to increasingly treat work as procedural and SOP-oriented, necessitating the deployment of deterministic agents.
- Priorities are shifting toward context engineering, focusing on tool integration and data retrieval, while the market evolves to include a product portfolio of models tailored to specific use cases.
- OpenAI maintains that even with the open sourcing of GPT-5, competitors would face significant difficulties in matching its influence and optimization capabilities due to inference constraints.
- Infrastructure is expected to eventually converge, bridging the gap between text and image model operations, despite current practices treating them as distinct systems.