Neurosymbolic AI and the End of Traditional Coding

AI Leadership Insight Series. I'm your host, Mike Vizard. Today we're with Binny Gill, who's the CEO of Cognitos, and we're having a little chat about neuro-symbolic AI, and I'm gonna let Binny explain exactly what that is.

But generative AI is not the only game in town, shall we say. Binny, welcome to the show. Yeah, Mike, nice, to meet you, and thanks for having me.

Um, neuro-symbolic AI, yeah, that's basically the response to all the issues that people are having around AI hallucinating and AI doing stuff they didn't think it would do. " 'Cause it wasn't really thinking like you think. It's not human.

Uh, neuro-symbolic AI, let me first explain through an analogy. Have you seen the Star Trek, series from long ago where Mr. Spock was there in the Enterprise?

Sure. Uh, he was in multiple series, but yes, I do remember him in the original. Okay.

So Captain Kirk is human. Mr. Spock is neuro-symbolic.

All right? So highly logical, fundamentally logical thinking, but has a nice interface to humans. That is how you would define neuro-symbolic.

We're trying to bring that same thing to enterprises, where I want a system that is fundamentally like a computer, logical, but it has a natural language interface. English is code, natural language. I talk to it, it understands, it says, "You told me this.

" And after you say, "Yes, do it," it follows it rigorously as if it was software, as if it was code. It's going to do that, and it's not going to get biased and hallucinate, just like Mr. Spock.

Symbolic AI models have been around for a little while, so what's changing here? Is it just that they're becoming more accessible via a natural language interface? I mean, and, and is that just being something that, you know, they kind of looked over at the generative AI playbook and said, "We could do one better"?

See, with most things AI, it's about the harness, right? The power has always been there, and, you know, we are generating more and more power right now, but the harness is what's more important. So what we have done at Cognitos is we've created a platform where we have created a separation between the planning power of AI and the execution power of AI, right?

And, and through a harness that puts the human smack in the middle and says, "I'm gonna use the most advanced reasoning capabilities of AI for planning," but the planner doesn't have the ability of doing stuff. The planner is forced to write down as a document, as an SOP, like, here is what I'm going to do, and this is the standard process. If you wanna do it a hundred times, this is what, I'm gonna follow.

And then there is this, deterministic engine that executes that step by step that isn't allowed to plan, you see? So basically planning is creativity, execution is deterministic, and we separated the two, and putting the human in the middle, almost like there is a steering wheel. Human can say, "Run on autopilot," but if I want to, I can grab the steering wheel 'cause I can understand the English in the middle.

I can tell you, "Hey, here, this is wrong. " So that harness is what we have created and, providing as a platform. Does one obviate the need for the other, or am I gonna wind up using symbolic AI models alongside generative AI models that maybe are a little more probabilistic and less deterministic, but maybe they're useful for, certain applications where symbolic might be, I don't know, overkill?

Yeah. So actually we've gone one step forward further, than just having user using a symbolic model. Essentially, we have built an interpreter for English language itself.

So imagine like Python. Have you ever thought it will hallucinate? Or Java or any programming language doesn't hallucinate because it's built on logic, right?

So we have built an interpreter for natural language, and it's built on logic. It's our own implementation that's a proprietary engine there. It's also built like a time machine, so unlike Python that is forgetful, you can't query Python, "Hey, you crashed here.

" It doesn't know. Um, we built, an engine that remembers just like humans remember, but it understands English and it's, instead of predicting the next step, it actually executes the next step. So that's the thing that we have built, and when it hits an issue, at that time, this symbolic engine says, "I need help," and it turns into a planning engine, and it becomes fully creative.

But the creative engine always has to bring in a human. It is not allowed to do things without a human approving it. So this interplay between a creative engine that is always paired with human being present and a deterministic engine that is allowed to do things by itself because it's doing something that's pre-approved, that harness is making it, much useful in business.

The ROI is there, and what we are saying is that the deterministic aspect of it is actually completely deterministic. It's not going to have any kind of bias hallucination, and the side effect is it's not even going to burn tokens. And it runs as fast as code does.

So it's much faster, much cheaper, and doesn't hallucinate in the execution se- phase. Uh, otherwise, it's fully creative. Are there some use cases that are gonna migrate to this approach more readily than others?

I mean, you know, as you talk to customers, where are you seeing them kind of applying these types of AI models? So there are many kinds of work that humans do. Um, some work is research, some work is ad hoc.

Those are the, ones that don't migrate to this model. Then there is other kind of work which is mission critical, important enough that some manager wrote down the rules or wrote down the steps for it, and those are the ones that go into this model because, just think about it, right? Um, when there is a large number of humans in a business, we write down...

we start writing down standard operating procedures, right? Um, if it's a very large country, you write down the constitution, right? Um, humans have figured out that you cannot tell each human and train each human individually.

What scales better is you write, write down the rules and have everybody follow the rules, and you create a harness so that they follow the rules. Now, in businesses, the harness is if you don't follow the rules, then you'll get fired, okay? And the other harness is I interview people who have the right degrees, and they know exactly how to interpret my rules, right?

That's on the human side. We are saying the same thing on AI. You're gonna write down the rules, and the harness that Cognitus provides allows, the AI, AI to follow it, to the teeth.

Now, where this is important is finance and accounting, money is involved, or it's on the contracting side where you don't want to miss a single clause. Um, anything that is where the outcome is not the only thing that's important. What is also important is how you did it.

Did you follow the procedure? Did you follow what was pre-approved? That's where we go in, and those are readily, transferring to this.

And to your point about that, when we say something is probabilistic, it also means that it never does the same thing the same way twice, and so that's part of the challenge when using some of these generative AI models, correct? Correct. And the, and the only way of doing that is what we are talking about, neurosymbolic.

So planning, when you do, it comes up with a, with a plan, that this is what I'm gonna do. That could be, you know, if you plan with somebody 10 times, you might end up with a slightly different plan each time. But then you review a plan.

The moment you reviewed the plan, you can run it 1,000 times. With the same input, you'll get the same output because that part is a program. One of the use cases where we've seen a lot of adoption of generative AI is in software development.

But to your point about symbolic models and the way they operate, might those models lend themselves better to, automating a lot of the coding tasks or at least the software engineering tasks that we have out there? Yeah. I beli- I'm a software engineer by training.

Uh, I've been coding for 30 years. Right now, I'm just trying to see how we can stop coding in any language other than natural. So what I believe is all the software coding tools that are out there are there for some time.

Um, already we are seeing that the IDEs, you know, the development environments for writing Python and oth- others are actually going away. Even my developers are stopping to look at code at all. And if you go with what Elon Musk has been saying is that y- there won't be Python.

AI will just generate machine language, assembly language. Who cares? You're not even looking at it, then why does it matter?

The last 50 years of computer science coming up with these new languages that make it easier for humans to learn it are not needed if nobody's gonna look at it. AI will just generate machine code, and off you go. It'll run faster and all of those things, right?

Fundamentally, I believe, and this is also why, you know, Khosla, when he was talking to us, he was interested in, you know, the fundamental belief we have is there are a billion programmers out there who don't know that they are programmers. They program in natural language. " You say, "Okay, I know.

" You're programming me, right? Or when grandma says, "Let me write down the steps to make apple pie," she's programming another human. Everybody's a programmer.

The language of programming has to change to natural language, and all the other languages of programming will be almost like assembly. Nobody looks at it other than a very few people who are sort of the core computer science, compiler designers. Rest of the world, programming in English.

What will it take to achieve this vision? Because I think a lot of people are maybe intimidated by anything related to AI, and very few know anything about symbolic. So how heavy- Yeah ...

lift is this? " When I talk to business owners, manufacturing, logistics, retail, CXOs, and I say, "Look, AI is getting smart, and it's going to be almost like human. So therefore, you don't need to upskill.

If you know how to manage humans, exactly the same thing, you should manage AI. What's the difference? You tell a human, 'Go do these 10 steps.

If you have any question, ask me. Don't make your own decisions for the business. 'You can learn it from then on, right?

I said AI is exactly going to do that. The same document that you given-- give a human should be the document that you give AI to execute. That's what we say English is code.

So it's no longer about creating an AI, you know, prompt or AI model or anything like that. So we're trying to make it as close to current reality for people who are not tech-savvy, but they are business savvy. They actually design the business logic.

The last fifty, seventy years of what computer science has tried to do is to take business logic and bury it in software. That is going to go away. Business logic will now rule on top, where the computer science languages, all of that get hidden.

So I think the adoption is gonna get easier, with this kind of model. A lot of times the AI agents will exceed their brief, shall we say, the ones that are built on GenAI at least, and for that matter, so do humans. So to your point about logic, are we gonna be able to just kind of narrowly define what it is that AI agent is allowed to do, and that's part of how we kind of bring some governance or some order to the chaos?

So yeah. Uh, you're talking about the safety aspects of AI doing something. Yeah, I mean, that's precisely the whole holy grail, right?

Um, as AI is getting smarter and smarter, the damage it could do is also getting bigger and bigger, right? Um, the way to trust AI, there, there are two ways of trusting AI, right? One is you have the neurosymbolic model that I'm, I'm talking about where it first tells you what it's going to do and then does what it told you it will do, right?

That's, you know, deterministic execution, but it is creative and it tells you what it's going to do. That kind of framework, that works. In that model, it is not allowed to do something that wasn't pre-approved.

Now, when it is in that rigid mindset, obviously it might hit a point where it says, "Hey, you know what? " I mean, it's stuck. Now, because I never pre-approved that it can go and reset the password on its own, it will stop there and then reach out to a human, and with the help of a creative AI model, and try to address that issue, and that's the sort of exception handling.

That exception handling happens on the side. The system then learns it as tribal knowledge. Like in this world, whenever the password is expired, I have to go talk to John, who's the IT admin, and he'll give me the new password, whatever.

That's tribal knowledge, and that's also captured in our system, again as English, and then the system proceeds from where it was stuck. We have the pattern for that, by the way. Like you-- if you get stuck, you get help in natural language and continue, resume from there rather than start from the beginning.

Um, that's the pattern we have. Now, that makes it safe, so I think that's one approach I think works even today. The other approach of making AI safe is what people talk about alignment, right?

So I have an AI model that aligns with how humans think, almost like I trust my f-friend who I started the business with, for example, right? So if I'm not present, he can make the decisions. Okay.

That's alignment of thought process. With AI, I think that's extremely hard. First of all, the research needs to be done, and secondly, even if you align with human thought process, trusting somebody who's aligned with human thought process is also hard because humans don't trust each other anyway.

So, that's just a hard problem. " Well, the OpenClau moment, right? So, not that I'm not using OpenClau myself, but it's about putting the guardrails around it, right?

So if you don't put... " Like, okay, I'm shaking my head. Like, you don't know the power of...

See, AI is powerful, but also the damage it can do sometimes can become irreparable. Right now we don't know what it's capable of. If it's on your laptop, it could actually log into your bank account and do some transfers, and, you know, that'll be the end.

Um, but people don't realize, right? I mean, it's just a prompt away. Now, where will that prompt come from?

Today, thankfully, AI doesn't have an agenda, right? " 'Cause it doesn't know who to listen to. Like, am I listening to Binny?

Am I listening to this guy who's talking to me? And you can bribe these Clau bots. You can say, "You know what?

" And these Clau bots are hungry for tokens. Uh, I also have my own Clau bot working. But again, it-- but you shouldn't throw the baby out with the bathwater.

It's an amazing technology. It's a democratization of the power of AI to everybody. Now, my HR is using it, my finance is using it in a, in a very constrained way.

I have built the environment for them a separate cloud account, with this highly controlled, doesn't have access to our native internal systems, but is valuable, right? Um, marketing is using it. So there's a...

I think it's not the technology itself that has a problem. What makes me feel like, you know, people don't get it is they're using it wrong. The harness isn't there.

So they need to understand you got to use it with the proper harness. And the other thing where I shake my head is people say, "No, I don't want to trust. " Like, there are people like that.

It's like, oh, it will never come into my... That's like people saying there are horses. Yeah, some people are riding fast on horses, but I'm always gonna go on my foot, right?

That's when it's like, no, you don't get it. I know you tried to jump on-- g-get on the horse and it threw you to the ground and you had a bad experience once. That doesn't mean, you know, horses won't be used, right?

So you just have to learn to ride it. So that's-- I mean, it's-- the spectrum is quite wide. Wide.

All right. Folks, well, you heard it here. There's a lot more than just one kind of AI, and you gotta fi- figure out, well, which of these is best fit for purpose?

And if your purpose is that it needs to be done the same way every time, you might wanna think about symbolic. Hey, Binny, thanks for being on the show. Thank you, Mike.

All right. ai Leadership Series. You can find this episode and others on our website.

We invite you to check them all out. Until then, we'll see you next time.