A Letter from a Machine Learning Engineer

G, a Machine Learning Engineer at a frontier AI lab, reached out via email to express their concerns about the future of the current LLM craze and the eventual downfall of frontier labs. G explains that internally, they have hit limits on scaling models for programming tasks, leading to executives quitting and stress surrounding IPOs.

The competitive moat for labs is non-existent, as open-weight models have largely replaced proprietary diffusion models, rendering image and video generation obsolete. As major figures in the research community shift away from LLMs, G notes that the best-performing models may emerge from open-weight alternatives within a year. G asserts that AGI is mathematically impossible with LLMs, citing Cantor's diagonalization argument and a leaked Google researcher memo from 2023.

G warns that mainstream sources are unreliable and that programmers should be skeptical of CEO forecasts. Lastly, G shares their personal experience with LLMs, finding them useful as a secretary for fetching codebase information, but expressing concerns about their utility for development and their potential impact on the industry.