yanjin101/the-first-principle-of-agi

This repository is a short written essay rather than a piece of software. It lays out one person's personal idea about how artificial general intelligence, or AGI, might work in practice, and the author is upfront that it is meant as a reference point for what could be possible, not a recommendation to follow blindly without fully understanding it first. The central idea, which the author calls the first principle of AGI, is that any long, continuous task can be broken down into small, discrete steps that a language model can complete correctly one at a time, through a process of multi-step execution. The author compares this to how a human brain works, with the breaking down of tasks acting like rational logic, and the language model completing each small step acting like intuitive perception. From there, the essay sketches out three related ideas. First, it suggests that an AI agent learns tasks by preserving the accumulated experience of human experts, either by summarizing that experience into a file called skill.md, or by hard-coding it directly into program code, which the author again compares to sparse versus dense connections in a brain. Second, it suggests that an agent decomposes tasks because human experts have already refined the underlying logic, breaking work down into pieces small enough that a language model can reliably handle each one. Third, it suggests that an agent scales by continuously saving task experience from many different industries and professions, gradually shrinking the smallest unit of work that a language model needs to complete reliably. The essay closes with an invitation to put this way of thinking into practice, calling those who do so the AI Adventists. There is no code, no installation steps, and no license mentioned in the README, since the repository exists purely to share this line of thinking.