When an AI system hallucinates for content generation on a piece of text, it’s not an ideal situation, but it’s also not necessarily catastrophic. If an AI powering a piece of military technology hallucinates, the outcome could likely have more severe consequences.

Jaxon AI got its start by building out AI systems for the U.S. Air Force with requirements for the highest levels of reliability and accuracy. The startup is now expanding into the broader enterprise market with a developed technology called Domain-Specific AI Language (DSAIL) that seeks to address a major challenge in artificial intelligence: hallucinations and inaccuracies in large language models (LLMs). The technology incorporates IBM watsonx foundation models and represents a novel approach to developing more reliable AI solutions.

“Our tagline is AI for AI because we’re using Jaxon to help users create custom AI,” Scott Cohen, CEO of Jaxon AI told VentureBeat

Hallucination occurs when an AI system generates an inaccurate response to a query. The inaccuracy can be caused by several different factors, such as incomplete training data and a lack of verification.

The DSAIL approach aims to help mitigate the risk of hallucination. Cohen explained that DSAIL takes natural language inputs and converts them into a binary language format that can then be run through a gauntlet of checks and balances, like a boolean satisfier, to ensure the AI response meets all constraints before being returned. This is done to limit non-determinism and increase the trustworthiness of the AI system for applications.

A commonly used approach by multiple vendors to help reduce hallucination is Retrieval Augmented Generation (RAG). In the RAG model, the LLM will also have access to a knowledge base to help get accurate answers.

Cohen explained that RAG is one technique that DSAIL uses to address the hallucination problem, but that it is only part of the approach. He noted that with DSAIL the output from the RAG technique would still need to be run through the gauntlet of checks, before being returned to the user as a result, to further limit hallucination.

Jaxon uses models from IBM‘s watsonx foundation library as building blocks for its AI systems.

Cohen explained that the IBM StarCoder model is used specifically for the code generation step in Jaxon AI. Jaxon uses StarCoder’s capabilities to automatically generate initial code for AI projects based on the design and requirements collected, as one step in Jaxon’s overall methodology for building custom AI systems.

The StarCoder LLM is an open-source effort that was originally launched back in May, with support from ServiceNow and Hugging Face. Savio Rodrigues, VP, of ecosystem engineering and developer advocacy at IBM told VentureBeat that IBM was actually one of the founding contributors to the StarCoder project. He also noted that IBM partners closely with Hugging Face to help bring open models to enterprise users.

To be clear, IBM has multiple code-generation LLM tools in its watsonx library. While StarCoder has broad capabilities, IBM’s own models are focused on specific use cases.  IBM has used its own code generation LLM to help with COBOL code migration and build quantum computing applications.

The market for generative AI and LLM technology is a competitive one, with big players including OpenAI, Microsoft, Google and Amazon Web Services (AWS).

IBM is looking for its slice of the market, specifically looking to help out developers and independent software vendors (ISVs) like Jaxon AI through a program it calls IBM Build. 

Rodrigues explained that IBM Build provides partners with access to watsonx, technical assistance, and go-to-market support. The overall goal is to provide organizations with reliable trusted AI foundation models, with consistent pricing, performance and availability.

“We know our customers trust the approach that IBM has taken with AI from the standpoint of how we train our models and the legal checks we go through,” Rodrigues said.