Is there an alternative architecture? Can we build distributed open source rules bases? What is the learning layer? See the comments for a record of how this page was amended. Below, or overleaf are my notes and links …
Links on the science
- Google Gemini, on “What is the information systems architecture of an LLM? Please document sources?”
- In trying to get to the bottom of the architecture, I asked google, “Which neural network products are used by the top LLMs?“
- And followed that up by asking google gemini, “What is a transformer neural net? Please quote sources?“.
Additionally, this white paper came out recently.
- Google on how to attack an LLM, a white paper
Links
- https://www.instaclustr.com/education/open-source-ai/top-10-open-source-llms-for-2025/, from instacluster; I have skimmed this once, there are some interesting things here.
- https://www.prismetric.com/how-to-build-ai-agent-with-deepseek/
- https://www.astronomer.io/blog/ask-astro-open-source-llm-application-apache-airflow/
Google AI insights says, “The Apache Software Foundation (ASF) does not develop a single, proprietary large language model (LLM) itself. Instead, it provides a foundation of open-source projects and tools that are widely used to build, manage, operationalize, and deploy LLM-powered applications and workflows.”
Today, and I need to look more, it seems none of these are about the rules base, I wonder if that’s deliberate and caused by the search engines and new players enclosing their search indexes. We’ve already learnt Google are doing this.
I added some notes on architecture, and a reference to a white paper on attacks.