Keep pulling the thread on Sherman Wu.
ChatGPT has approximately 800 million weekly active users, which represents 10% of the global population.
The AI market is increasingly making room for a proliferation of specialized models, moving away from the idea that one model will dominate all use cases.
The recent introduction of reinforcement learning fine-tuning allows customers to improve a model to state-of-the-art performance on a specific use case, a significant advance over supervised fine-tuning which primarily adjusted tone and instruction following.
OpenAI offers customers discounted inference and potentially free training via its reinforcement fine-tuning API if they are willing to share their training data with the company.
OpenAI has not observed any cannibalization of its commercial products from the release of its open-source model, GPT-OSS.
OpenAI has a local, on-premise model deployment running inside a classified supercomputer at Los Alamos National Labs.
OpenAI's founders, Sam Altman and Greg Brockman, have maintained a principled strategy of developing both a first-party application (ChatGPT) and a third-party developer API simultaneously.
The retention rate for developers building on OpenAI's API is surprisingly high, indicating a strong stickiness for its models despite the availability of tools to swap between different model providers.
The initial internal belief at OpenAI was that a single, general model would eventually subsume all other specialized models, a view that has since completely changed.
OpenAI's evals product now allows users to bring in and test models from other providers, reflecting a strategy that a rising tide in the AI ecosystem benefits OpenAI.
The prevailing industry view in 2022 that prompt engineering would become obsolete as models improved has been proven wrong.
The focus of prompt engineering has shifted from simple instruction following to "context engineering," which involves determining the right data and tools to provide to a model.