2024 MIT R&D Conference: Startup Exchange Lightning Talks - Jaxon

SCOTT COHEN: Large language models like ChatGPT, the open source varieties are rotten liars. Cannot trust them. The reason is they're all sequence-based.

They've been trained on the world's data. And they're looking at patterns, sequences of characters to determine what someone is saying, and then spits out the sequence of characters on the other side to try to placate you and give you the answer that you were looking for. Unfortunately, it's not based on logical reasoning. It's just based on these sequences. They are very much probabilistic.

One of the issues here is that when we're trying to use these in corporate settings, in government settings, we bring them into workflows where we have to or want to trust them. So what we have done is embraced various techniques that address this problem.

The most popular one today, all the rage in the AI community, is this so-called RAG technique, which stands for Retrieval-Augmented Generation. And what it's doing is bringing representative data into the mix. As opposed to using the vanilla large language model, it's giving you representative data as little nuggets of information relevant to the task at hand, relevant to the use case, relevant to the domain that this particular model is working within.

It's kind of like a clue toward the right answer, and it increases the probability that it will give you a good answer, but there are no assurances. It does not give you proof that you can trust this output.

So what we're focused on is number 4 here, symbolic reasoning. Tried and true. Been around for decades. We are the bridge between the very powerful large language models. Very creative. They can, as they describe it, hallucinate, come up with fabricated responses.

What we're doing is taking that natural language and turning it into a mathematical problem, something that can be solved for. Think back to geometry where you had a formal proof and you showed every step of your process, every claim that went into the answer, and there was only one answer at the end.

This is the same thing that we're doing here. We take the natural language. We turn it into a domain-specific, language-powered program that gets run through these off-the-shelf so-called solvers. The solvers are saying yes or no. It does or doesn't meet all the constraints and assertions and known facts that you're imposing upon it.

The way we do this is we bring it into a knowledge graph. We turn the natural language sequence of characters into a graph structure that aligns to the domain that we're working within. It's a process that has a human in the loop so we can get feedback and constantly improve the models, and has a path toward this more trusted AI.

And that's really the value, is bring it into use cases where people are reluctant to use it. Say, clinical decision support for health care, or in the claims adjudication process for insurance, or picking the right portfolio in financial services.

Or with us, we've started with the Department of Defense where they have lots of data that needs to be properly classified, meaning which security level can this particular piece of data fall into? What label should we apply to this particular email? What marking should we put on this particular report?

And everything is driven by these security classification guides, which are documents that were written by humans for humans, never intended for AI to interpret. So we're breaking that down into constraints and rules that we can apply in near real time to new data coming into our environment.

We have a number of projects with DOD. We're working across the intelligence community as well. We are intentionally now focused on bringing it to the commercial sector. We have use cases, as I've already mentioned, in financial services, insurance, life sciences, and health care.

We are eager to find pilots. We would love to work on collaborative development efforts with anyone that wants to perhaps white label the solution, and for partners that want to help us bring this into the industries that we feel make the most sense, which are usually regulated industries, ones that have laws and regulations that have to be adhered to that we can use as guidelines for the processing of this new information coming in. So you can find me at jaxon.ai or email me directly, scott@jaxon.ai. And I'll be outside. Thank you all.

SPEAKER 1: Thank you, Scott.

[APPLAUSE]