GPT-4 Turbo and Assistants API: The Release That Pushed AI Apps Mainstream

Going back through https://openai.com/index/new-models-and-developer-products-announced-at-devday/ two years later is useful for understanding which of the ideas introduced then became load-bearing for the current AI product landscape.

GPT-4 Turbo with the 128k context window and lower pricing was the accessibility change that made building on top of GPT economically viable at scale. The previous pricing model was a meaningful ceiling for startups and independent developers. Removing that ceiling changed who could build and what they could build.

The Assistants API being the more structurally significant announcement is the argument I would make looking back. Persistent threads, file retrieval, code execution, and tool calling as managed infrastructure rather than custom implementation is the pattern that most AI applications still use. The pattern of a user interacting with an AI assistant that has access to relevant documents and can take actions on their behalf was not invented at DevDay but it was productised there.

What is worth tracking is which of the patterns introduced then have become standard practice and which have been superseded. Retrieval-augmented generation from uploaded files has largely been replaced by better context windows for many use cases. The agent pattern has evolved significantly. The Assistants API itself has been updated multiple times.

For builders: which AI app pattern has been most durable for you, chatbots, retrieval assistants, agents, or something else?