The steady stream of new LLM-based products has refined my thinking about how it will impact the technological landscape since my first thoughts.
My earlier thoughts about the role of LLMs have evolved to be more mind-bending. The early use of LLMs is against mostly static data. Chaining LLMs together provides a more comprehensive set of data for responses. Creating custom LLMs augments their general-purpose nature with domain-specific knowledge. The popular LLMs provide the capabilities to understand requests and phrase/format responses, while custom LLMs provide specific information to add relevancy and reduce hallucinations. I’m struck by the powerful information mapping capabilities of LLMs (e.g., language translation, text-to-image, text-to-code, etc.).
LLMs blur the line between input and output. Generated content can be translated and input to another LLM. For example, conversational text is used to generate an application. The application is useful, but the code base can be used by another LLM to create a variation. The recursive reuse of content will lead to an explosion of new content and in different forms. With such an explosion, I wonder what the value of the shared experience will be in the face of aggregated experiences.
I foresee a world where content is translated into other forms with little effort. This could get confusing because the content form has traditionally separated markets. Think about a book that is semi-automatically (a) translated into other languages, (b) has audio and video generated, (c) has a website generated, (d) and an app generated. In education, think about a set of standards/learning objectives and content and this information being translated (and magnified) into (a) interactive learning activities, (b) projects, (c) personalized content, (d) student-friendly rubrics, (e) projects to prove to mastery, and (e) a personalized website with resources for assistance.
My earlier thoughts about software development were that LLMs would supplant traditional development as the request for information would be made without the traditional software layers (presentation, logic, and data). LLMs will eliminate the need to code more commonplace types of applications as developers validate and debug automatically generated code. Application code that is translated into plain English and vice versa. In the case of more complex applications, LLMs will still play a role by generating specific code layers. Examples include: (a) generating SQL CRUD scripts from the database schema, (b) generating web interfaces from hand-drawn mockups, (c) generating data visualization from queries, (d) and generating business logic from a description of the validation rules and business goals.
