I was reading through my daily email newsletters on AI, InfoSec, and Big Data, and found a different approach that may benefit Logos. In my understanding, most AI approaches train all-purpose LLMs for specific purposes and use retrieval-augmented generation (RAG) to ensure currency and applicability. The RAG and specific training data set improves accuracy but still leaves room for hallucinations because the LLM design and development include training data from across the Internet and lack the specific context of the enterprise. It's also very power and time intensive, making it expensive to operate. A new approach being explored in highly regulated industries like banking, where accuracy must be virtually perfect, uses purpose-built small language models (SLMs) trained exclusively on the target data set. The approach is domain-specific, faster, and virtually eliminates hallucinations because the SLM understands the domain language natively. Here's the link to the article: https://ragaboutit.com/the-small-language-model-awakening-why-enterprise-rag-is-abandoning-foundation-models-for-domain-specific-precision/. If done well, the SLM approach would eliminate the need to check outputs across multiple LLMs to minimize hallucinations as well, further increasing output speed and reducing licensing fees.
If the article is correct, taking an SLM approach could make Logos AI faster, more accurate, and less expensive to operate. As the article points out, it may also require change in the architecture, but my suspicion based on Mark's regular updates is that Logos' existing architecture isn't far off.
