Introduction to Herbert Roitblat and Reliath AI

I wrote a book about artificial general intelligence called Algorithms are not enough: Creating general artificial intelligence (MIT Press). It lays out a roadmap for AGI. I had started to write another book on how to pursue that roadmap when my now co-founder, Chris LaCour, read a post I did on Substack, The self-curation challenge for the future of AI. We decided to expand that into Reliath AI.

Reliath AI provides a bespoke model for each customer that contains only information identified by that customer as a source of truth. It is not a language model, but uses facts as its atomic level of analysis. It represents each fact as a symbolic proposition and as semantic vector. By the combining the two, we get more power than can be supplied by either one alone. It is a deterministic model, that trains in one pass, so it is very quick to train and very efficient to query. It uses only a tiny amount of energy compared to an LLM.

Rather than including every piece of information that the LLMs can access, its focus on customer-supplied data means that there is no unauthorized data included in the model to block from coming out. Governance is greatly simplified. The symbolic representations can be organized into knowledge graphs without an ontology.

We have applied for patents. We have a few customers and a few applications. Thousands of applications can be built on top of Reliath models. One of our current applications is an AI harness governing AI agents.