Personae-Driven A.I.

Personae-Driven AIPersonae-Driven AI seeks to reduce hallucinations, guard against malicious use, and generally align generative AI (genAI) systems by layering on a ôpersonalityö on top of an Large Language Model's (LLMs) áinteractions with the user.ááKeeping an LLM's interactions helpful, truthful, and aligned to the mission are often done with a mixture of LLM training, guard rails, system prompts, and carefully constructed interactions via prompt engineering techniques.ááAn effective ôprompting techniqueö when seeking aligned LLM responses to ask the model to take on the role of something or someone, i.e. ôas an accounting professor, help me understand the difference between debits and credits.ö We assert this approach can be enhanced, scaled, packaged, and made as a configurable layer to install above LLMs.ááAlignment Via PersonalityOur Personae-Driven A.I. system is package of sophisticatedly articulated personalities, roles, and intents that deliver mission aligned behavior. Our vision is to drive a use-case's desired behavior by making the LLM ôwantö to be aligned to the mission. We anticipate this is will shortcut needs to multi-shot prompt or heavily example prompt corner cases - as the vector-richness of a well-developed personality will enhance the next-token selection results beyond typical anodyne system prompts.ááThese personalities are a suite of:proprietary system prompt modules áretrieval augmented generative (RAG) datasetsOur objective is to craft wholistic personalities áôout of the boxö, as well as fine-grained personality trait modules that can be configured by the user that results in a custom persona that can align to different genAI use cases.ááTechnology StackThe stack of technologies we will leverage open source tools including:áPythonOpen Source LLMs such as META's LlamaAutoGenLangChaináBlue & Red Teaming RefinmentOur personae will be developed via Blue and Red Teaming.ááBlue team personae will be focused on broad, generally mission aligned use cases (i.e. SAM proposal assistant that's a stickler for requirements), as well our configurable personality modules (i.e. ôpithyö with direct, spartan responses vs ôdiscursiveö which may assist in more meandering ideation.)ááThe Red team will be developing adversarial personae to serve as internal tooling: to generate synthetic test data, test the alignment of the Blue personae, and investigate possible malicious corner cases that may be exploitable in the personae offerings.ááVisionThe vision is to develop these configurable modules of system prompts and RAG datasets, wrap them in a configurable toolset, and have them transform LLMs into aligned systems. Along the way we wish to mature our adversarial toolset to improve quality of our blue team personae.áWe also wish to mature our testing suite to capture performance metrics of our personae and personality modules.ááThe end result is a modular toolset that installs above LLMs that results in more aligned behaviors.á