Debate over LLMs relying on proprietary data rather than proprietary technology

A discussion on Reddit highlights the argument that large language models (LLMs) do not rely on proprietary technology, but are instead built by training on massive amounts of proprietary and potentially copyrighted data. This touches on the ongoing legal and ethical debates surrounding AI training, where the core value of LLMs is argued to come from scraped intellectual property rather than novel algorithms. It could influence future copyright regulations and data licensing models for AI companies. The argument points out that the underlying Transformer architecture is widely open and shared, meaning the competitive advantage of commercial LLMs lies almost entirely in the scale and quality of their ingested training data.