I built a coding agent two months before ChatGPT existed (solmaz.io)

A developer recently reflected on their experience building a coding agent just two months prior to the launch of ChatGPT in late 2022. Utilizing OpenAI's code-davinci-002 model within a custom Jupyter kernel, they navigated limitations of early AI tools, generating and executing Python code without full access to system resources. Unlike today’s advanced coding agents, their initial tool operated with a Jupyter nbformat structure, primarily utilizing input/output pairs, which is now seen as less efficient than current chat models that employ a more sophisticated tree structure of messages. This recounting underscores the rapid evolution within the AI/ML community, particularly the integration of new features like tool calling and improved reasoning capabilities, which have significantly enhanced coding agents' functionality. The developer's journey highlights how foundational models have paved the way for the latest advancements and reflects on both the challenges and innovations that have emerged in just a short span of time. Their experience calls attention to the importance of adaptability in AI development as models continue to evolve, illustrating the creeping complexity and capabilities expected from future AI systems.