Implemented GPT-2 in pure Postgres. Full forward pass in SQL. Multi-head attention, LayerNorm, sampling. Matmul hijacks pgvector's index operators. The KV cache is a table. Generates 0.112…

> LLM.sql great idea because > of the way it works. The problem is that the SQL Server database is not very good Tempting to go deep into improvements, curious about scaling properties, performance,…