Willison also demonstrated the use of LLMs for structured data extraction and code interpretation. He showed how unstructured text or images can be converted into structured data using the datasette-extract plugin. He also showcased the capabilities of ChatGPT's Code Interpreter mode, where the model can generate and execute Python code as part of an ongoing conversation. Despite some challenges and limitations, Willison's talk illustrated the potential of LLMs in data journalism and other fields.
Key takeaways:
- The author gave a talk at the Story Discovery at Scale data journalism conference, discussing the use of Large Language Models (LLMs) in various applications, including data extraction, AI-assisted SQL queries, and more.
- Several live demos were conducted during the talk, demonstrating the use of different tools and models such as Claude 3 Haiku, Datasette Cloud, and Gemini Pro 1.5.
- LLMs can be used for a variety of tasks, including generating haikus from images, extracting structured data from unstructured text or images, and even executing Python code as part of an ongoing conversation.
- However, the author also highlighted the potential risks and challenges of using LLMs, such as the possibility of the model hallucinating extra details in the output, and the importance of considering the model's output length limit.