#15: Digging into explainable AI

All right, well today is going to be special. We're talking about neuro symbolic AI and all sorts of bleeding edge, state of the art AIs and finding out where the world is going with it. Who are you?

Hi, I'm Angela Dali. I co founded a company called Oomni and my background is in AI. I've been in AI for more than 20 years, applying AI to different industries, different applications and finally discovered some of the limits of AI and the way forward to overcome those limits.

Yeah, like all of us, we're all figuring out, oh, it's hallucinating, oh, it doesn't do this, it doesn't do this or it can't think of on its own very well. Right?

Yeah, exactly.

I'm happy to announce that I am now taking sponsors for unaligned. I'm proud to have AI Top Tools as our first official sponsor. One of the most important things for me running my podcast is staying on top of the industry, knowing all the players and what they can do. I made a complete list of companies in the AI industry on x.com but no one has as comprehensive a library of tools as AI Top Tools. They break down products by use case like productivity, media, chatbots, customer service and more. A lot of companies are looking for the exact right tools for their needs and there's no better place to find it than AI Top Tools. From your point of view, where are we today? In March of 2024, where are we and where are we about to go?

Well, as a tech entrepreneur, I've set up lots of different companies and saw how AI has affected different industries. For example transportation industry, the retail, big data analytics. A lot of different indices have changed as a result of new forms of AI. I think some of the applications are quite unsexy, like, you know, getting better transactions, get, getting better loan decisions made. But you know, obviously generative AI right now is the buzzword. Spatial computing is something that is going to be something of the future. I think that's going to be much more mainstream and I think that where we are right now, everyone is now finding out, okay, AI is useful. We can now start using it in everyday applications, we can actually start creating new jobs about it or change different things in our daily lives. So I think that where we're going to go from here is that AI is going to become much more pervasive, more integrated into everyday life and I think it's going to be very exciting. I think that it's going to be here to stay. Yeah.

As we get, you know, If I start looking like, what's a decade from now? Well, we, we want a humanoid robot in our house, folding our laundry, right? Interacting with us, playing games with us, high fiving us, teaching us to dance, stuff like that. We need new AI to get to that world, don't we?

We do. As everyone is discovering, there are limits to what AI can do. For example, um, GPT starts making stuff up. We weren't used to that. We were used to going into a search engine, typing in something and getting an accurate answer. Um, you also get this problem of hallucinations. AI didn't hallucinate before. Um, I first programmed my first neural net way back in the 90s. And back then neural nets were seen as a toy thing. No one really took them seriously. And that all changed because of hardware, because of hardware advances and availability of more data. And I think the same is going to be now. I think we're at the very start of large language models and also of generative AI. And I think these problems are going to be ironed out and figured out. I do think that yes, we are going to have robots in our daily lives. I think that AI is going to be our companion and helping us in absolutely everything. And I hope that we'll get more free time as a result and also time for us to be able to bring about the potential, the potential in people so that he can be creative and do more stuff.

Yeah. You're an expert in this new field of neurosymbolic AI, right?

Yes.

Tell us a little bit about what is different about that kind of AI compared to like what's running ChatGPT, a large language model, or Google's Gemini or whatever. Tell us where, where, where is the bleeding edge going and what, what is that going to mean to humans? What, what, what is neuro symbolic AI going to be able to do that our current AI can't do.

So basically most AI, the current AI that we have, including the ones that are running Gemini, ChatGPT, Llama and all that, all of that runs on neural networks, so called deep neural networks that are very large neural networks that learn things in form of weights, they work. But unfortunately, in order to examine how these networks work, it's almost impossible. Not even the creators of, you know, all the large language models actually know how they work. So the symbolic part is related to logic, rules, equations, you know, saying, you know, A happens because of B, because of C, the way that we, people actually, we're humans, actually reason out. So neuro symbolic is the combination of the Neural network, the, you know, the pattern matching capabilities, the prediction capabilities of neural networks together with the logic that comes from the symbolic part. So being able to deduce something, being able to come to a conclusion about something, being able to also generalize about new things that you've never seen before and also in diagnostic mode, say seeing for example a few symptoms and, and being able to assign it to a particular cause and also, you know, perform some kind of action based upon it. So these three capabilities are things that current AI cannot really do readily. In the past there have been attempts at doing this but unfortunately the so called maybe remember expert systems, they failed because you had to have people actually typing in this route. You had to maintain it all the time actively. So while the idea was right, it couldn't scale up. So what neuro symbolic AI does is take the good bits from deep learning AI, taking the good bits from those old ideas and making them scale up and work in large scale that we're now used to.

I was just flashing on stoplights and while you were talking and you know, humans know what a stoplight looks like because it's usually a common shape, right? It has three lights usually and we know when it goes green to yellow, yellow that red is about to come on, right? You know, there's some rules there, right? Is that what you're talking about?

Not just that, I mean that would be one example in, in, in which the system will learn, okay, just about if it's, if it's red and now it's yellow, it's going to go to green. So those kind of things it will learn reliably. Now if you've got like thousands and thousands of video of footage of you know, traffic lights going from red, yellow to green or the other way around, obviously you're going to learn that even from a neural network. However, if it's something that is much more, less common, a new situation happens. This is where systems like ChatGPT suffer. This is where large language models actually suffer in adapting to new situations. And that is something where people excel. In fact, if you were to give, you know, if I were to give you an object, I just draw something at random and show it to you. You can just copy it with just seeing one example. That is something that is currently beyond what these systems can do. And imagine we are going to encounter these situations as AI goes more out into the world. If you're wearing your, you know, your spatial computer and you encounter some strange scenario, some room that you've never seen before that maybe it doesn't look like other stuff, you know, like doesn't have a standard design or layout. You need that kind of flexibility. So the symbolic part gives the flexibility that, you know, to new things.

When I talk to the guy who's building the little rabbit device that was hot at the Consumer Electronics show back in January and sold, I don't know, 50,000 units, he has a camera and a microphone on that device. And he's really trying to get to this new symbolic AI where it can learn about a new application very quickly, like a human being does. Right. If I, if I give you a new app, like if you've never seen Uber before and I give you Uber, you figure out how it works pretty quick. And he's trying to get an AI that works that way. Is, is that sort of the same work you're trying to do?

So basically what we're doing is to provide a framework, a total architecture that can, that can power neural networks, that you can have explainable transformers, for example, so that the next step would be applying this to large language models that can eliminate hallucination. So this is actually very exciting because if we manage to eliminate hallucination totally, so the large language model knows, is this something plausible or is this something that is a total parenthesis, gibberish basically, then it gives it awesome new powers. Maybe you've seen the example about Tom Cruise's mother, for example, if you ask it who Tom Cruise's mother is, but then if it's a merely pfeiffer and say who, who is her son? The LLM answers correctly about the mother, but doesn't answer the other way around. No human is going to make that kind of mistake because you've got the symbolic logic. So these things that we take for granted, common sense, it's not so easy to actually program into AI. So the nice thing is that once you've got the symbolic part, the AI can explain itself, can explain actually how it is working, how it is doing these decisions. So this is very interesting because you get an introspection into how the AI is working. What is the decision making process of the AI? How does it reason out? Does it make sense? Because this is the only way that you're actually going to check whether the AI is making sense or not by actually understanding how it's working.

Yeah, what a new world we have to build, isn't it? This is going to lead to all sorts of new personal assistants and finish off autonomous cars. All sorts of things are in play because of the AI of the Day but because of people like you who are trying to fix some of these problems. Go back to the hallucination because you mentioned that a few times. What causes like a ChatGPT to give you a wrong answer?

Yeah, so. So basically when you large language models like ChatGPT, all they're doing, not all, but they're being fancily doing next words prediction. So you give them a question and they're trying to predict okay, what word should I put in next in my answer? You know, as it goes often gives you the answer. Now unfortunately, the way that they work, you have many different spots. Imagine, you know, you're going down a branch like a big big tree and you're selecting each branch with every token, with every word that you're outputting. Obviously a lot of those output trees are going to be fantastically inaccurate. There is going to be like a branch of kind of okay ish answers and hopefully the GPT goes into that branch and there are going to those branches that are actually real accurate answers. Now unfortunately there is no mechanism to check the plausibility of what the output is saying. So if the most probable output is gibberish or it is something that is wrong, you need to correct it. Now there have been mitigation techniques put into LLMs. For example, for instance, you could spit.

The answer out and bring it back into the LLM and have the LLM rate the answer for accuracy and it doesn't fit. Fairly good job at that basically.

And that's basically how a lot of this guardrails are done. Literally having a lot of people asking the questions and also giving it the answer. And effectively the LLM tries to imitate that. That's kind of the simple explanation. Also there are techniques called RAG where the LLM goes out to the Internet, performs a search and tries to reconstruct the output using stuff that has found from search engine material. And that's kind of.

Yeah, I asked, I asked chat GPT, what's the hours on a super cuts a play retail, a retail establishment in America that cuts your hair, right? And it went to three separate websites. It went to Supercuts website and then went to the chamber of commerce website which has hours information and went to third website, grabbed all that data, then came back and decided oh, what's the hours on all these? Who has the best set of hours for this question? And then give me the answer. Right?

Yeah. So symbolically I would do, for example would take that one step further. It would also be able to explain why it gave you that particular answer. So it will actually give you the reasoning. If you just want a quick answer, you don't need the explanation. But if you're interested in the explanation, it can actually explain to you why. Why did it tell you know that it opens at 909 or 10 and.

We checked these three websites, we got ours, but this is the most accurate. And here's the reason it's the most accurate. Okay, which is?

Which is the way that you would explain it to me, for example, if I were to ask you so. And I would ask you why. So it has more human like qualities, it reasons more like a person would. And I think this is very important because if we are going to trust AI, we need to have it think more like a human being. And this is where neurosymbolic AI comes into play. And the way that it generalizes the way that it actually can apply itself to things that it has never seen before and also understands how the world works. Because one of the main problems currently is that LLMs do not have what is called a world model. They literally imagine someone who's blind, who actually has no senses, has learned about the world just by reading about it. They don't actually know the word looks like. So even when you see things like Sora, for example, like the video generation stuff, what happens there is that all those videos, they're turned into small pixels and they are kind of coded as text. So even though you get the illusion that they do know how the word looks like, they actually don't. And this is a very important problem, in fact that needs to be solved because with symbolic methods you would know that, for example, a person has five, you know, has five fingers and not six, like you know, some of the generative AI does. And this is a way actually of solving that. You know, you actually make rules that, okay, you have, you need to have five fingers and only one thumb out of those five fingers.

Time I checked.

Verified.

Yeah, that's how we're going to prove we're not AIs to people hold up your hand. That'll work for a couple more months, right?

Imagine these rules working in an autonomous car. If in a self driving car, for example, if you're backing down your driveway and sometimes cars tend to stop all of a sudden for no apparent reason. It will actually I, you know, I imagine a world where we'll have an explain button on the dashboard and the car actually explains to you why it took a particular action. So you can actually direct it and also teach it not to Follow, you know, not to do something wrong. It may be right. It may be that it saw something, you know, and, but in some cases it may be wrong.

A lot of times it sees stuff you don't even see. Like a, a child is behind your rear bumper. You can't see them in your mirrors. The camera can see it, right? And they can say, oh, stop. You know, and you're not knowing why it's stopping. Tesla. A guy who has a Tesla got in an accident because he overrode the AI going through an intersection and it turned out he got hit by somebody. The car saw the, the guy coming, he didn't because he was looking straight ahead, right?

Yeah, yeah. I mean, these things will happen and I, they will happen more frequently. For example, you know, cars that then would, you know, even, even if you get a green light, but you see someone drunk driving down the road and I'm sure you've seen some of these videos, you know, and the car just rolls just in straight in front of that person. As a person, you'd never do that. Even though you have the right of way, you're not going to do that. So, you know, this little common sense, it's actually very difficult to input into AI, to actually teach it. And I'm not saying, you know, that neurosymbolic AI is going to learn all of this overnight, but it has a chance of actually progressing beyond the current limits. And I believe that it is going to be the way of the future. Basically how AI is going to.

It's crazy. When do you think this is really going to start showing up for everyday people? Right? We're getting whiffs of it with the Rabbit, you know, but the Rabbit's only selling 50,000 copies. When does this get to the scale of a chat GPT or a Google or an Apple or something like that?

So what we've done, we've been actually working on this for a number of years, steadily, like creating the new architecture. We're calling it hybrid intelligence. And it has been, you know, more than four years of testing actually to try to get it right. And we did actually manage to crack it. It's incredibly difficult because when you're actually creating a new foundational model, totally new, it takes quite a lot of time and effort to actually get everything lined up. So we've created actually a number of pocs of proof of concepts in different industries and we've had fantastic results. So on average, in different industries that we've tried, they outperformed Deep Learning by more than 17, 20%. And also an interesting statistic is that when it comes to the power that they use, even in training and also during normal operation, it reduces the amount of power needed by 90%. And this is a huge, huge reduction. And 90% reduction means literally like millions saved when training large AI models. The reason about it is that the nice thing, when you learn about how a model works, you know also what you can take out. Because if I know exactly, you know, like this wire leads to this, this connection leads to this, I know how to actually take out things that are irrelevant. So by compressing the model, you know, to a much smaller size, you get these benefits. And when you think about it, you know, we've got proof that we can have a very advanced model, which is our brain, which takes much less space, much less power, and is still arguably much more powerful than the powerful models that we have.

This sounds like it could speed up the models as well, which is another major problem with chat with LLMs. You know, you ask chat GPT a question and it sometimes it takes 12 seconds to answer back. Well, if I'm trying to build a personal digital assistant, 12 seconds is a long time, right. It interrupts your flow of talking with the AI and interacting with it. Sounds like this could really bring some speed advantages, is that right?

Yes, especially, I mean, obviously once you have a smaller model, it's going to work faster. And also what's exciting is that you.

Can put it and work on shitty computers too, right?

Exactly. And you can use less.

I work on phones or Raspberry PIs or small, little, tiny computers.

In fact, we are testing it on small Raspberry Pisces to see if it works. And it works actually quite well. I mean, it's still early days actually when it comes to the hardware part, but you can already see the glimpses of the future basically of how this thing can actually fit into a very small, low power hardware. And you literally get this advanced models running absolutely everywhere. There have been advances, for example, people running LLMs on like small smartphones. But usually there are a lot of techniques, you know, like to try to approximate it, compress it a lot. You're not getting the full model with this. You actually can get the full model, no compromises. You get an actual full model running on your smartphone, for example, and this can open up totally new applications that maybe we haven't even thought about.

Wow.

One of the things, what a new world.

Tell me about the company you're building and what kind of effort you're building. Because We've been talking about technology, but not, not necessarily what you're building.

So, so the company is called UMNI for Universal Machine Network AI. And basically with what we've done, we're focusing right now on fintech applications. Just because people who are working in regulated industries get immediately the value of having an AI that can actually explain itself, that can explain also its decision making process. You can audit those results. You can also verify that the thought process, the, the reasoning process by the machine is actually cor. And just to quote some, you know, like some successes that we've had, for example, when it comes to false positives in compliance reporting, we reduce, we reduce the error by more than 70% and this is huge savings. Like literally that translates to billions of dollars in saving for banks, financial institutions and, you know, other, other companies who are regulated. So I think that regulated industries are going to be one of the early adopters actually. And we've had a good uptake so far. We're also focusing, we're based in Europe, we're based in London and we have also offices in Malta and Italy and other countries in Europe. And they passed recently in Europe the EU AI act, which I think is going to have a lot of impact into how AI models are run in Europe. I think the US will follow suit, maybe with a slightly different set of regulations, but you can see the writing is on the wall that everyone wants to have advanced AI that actually works in a, you know, controlled way. So this kind of systems that we're building are perfectly positioned to actually be that solution, to be able to control. So I think that being able to trust what you do is important. And this is what we're focusing. So we've built, we've built this model, this hybrid intelligence framework as a library and you can, you know, either download it using Python or you can integrate it with other systems. So it's kind of agnostic. It's like an engine that you can put into any application and by enriching the applications, it's very obvious because instead of just getting an answer from the AI, you get the answer and an explanation of how it got to that answer together with a justification. And the justification is literally in code. So you can actually run a piece of code and check, okay, the AI is right, or maybe the AI is not right. And for people, you can actually examine the output in form of rules so that it's human readable rules. I mean, some of the models may be a little bit more complex than others, but you can actually understand what is going on. And so the nice advantage of this is that you can actually correct it directly rather than having to find an example. You can actually write it down and you get the right answer.

This is going to really matter when we have billions of people using AI because it can learn from us. If we see something wrong, we can say, oh, that's wrong, here's more information, or here's a better source of information for this and retrain the model. Do you, are you also working on that, on systems that bring back, you know, this kind of feedback into models really quick? Because right now, yeah, that, that happens, but it takes a year for it to show up, right? OpenAI is still working on models that were built a year ago, right?

Yeah. Or, you know, or you can correct, for example, like a session, an LLM in that particular session with a prompt, but it's kind of forget what it had done previously. So unlike those systems, what we've built actually can take rules directly from people and you can retrain them in kind of a more traditional sense, but you can even put a rule that is incorporated in real time into that model. So if you've got a real time system that needs to constantly update itself, this is an almost perfect system. If you can actually correct it in real time, it's going to adapt instantly. And this is again one of the powers of symbolic systems that they can adapt much faster in fact, than the neural patterns. Neural networks are good for this, for, you know, like large scale processing large scale information of things that they've seen before. But for quick adaptation, symbolic are just unbeaten. And when you think about it, say for example, the autopilot on a plane is a symbolic system. All the software that we've written, most of it are symbolic systems. And you know, when you look at all the mission critical systems, they are written in symbolic code. So it's kind of pretty obvious when you think about it with hindsight that, you know, the symbolic systems are going to perform much better.

What else do you want people to know about your work or what you're seeing happen in the world? You know, where you think AI is going?

I think that AI is going to be integrated into everyday life much more than we're seeing now. I mean, you know, you get people who are doing long chat sessions with ChatGPT already. I mean, it's quite amazing. One of the nice things that happened is that, you know, in the last years I didn't need to explain what AI is. People used to ask me, you know, like 15 years ago, what AI is all about. I had to actually explain what it is. Now most people actually do know what it is, so I think that we will have more something about it.

You know, they, they still don't, still don't understand how it works or what's going on underneath. But at least we have some examples now like chat GPT or Google Card or whatever.

Exactly. You know, I, I knew that AI made it. When I heard Snoop Dogg talk about AI, I was like, okay, it's mainstream. And you know, when you think about it, I think that we will all have AI in our pocket. AI that is also locally driven, not just connected to the Internet. If you're going on a hike with your vision, whatever, pro something, I don't know what number it would be back and then you'd be able to.

People at Apple say it's version three, the glasses, maybe the lightweight blades. That'll be lightweight, yeah, Fairly affordable, portable and have a lot of application because not just the AI is coming along, but the developer community has built a lot of things for them.

Right.

They say it's about three to five.

Years away and, and obviously one of the problems with that is that the headsets can be quite heavy. But one of the reasons why is because of the processing power, the batteries for example, that you need. So being able to run the AI model more efficiently will have a knock on effect because you'll get lightweight gear, you'll be able to, you know, use charge for much longer periods of time. So AI will be able to accompany us in more places than it can actually do right now. So I do see also delegation, intelligent delegation to AI. I do not like the fact that people just rely on AIs, you know, output without checking. I want to have a reliable AI that I can actually trust. So I see trustworthy AI as being very important. I also see that, you know, apart from trusting the AI, you, you need to be able to understand it and be able to say okay, for this kind of cases now I trusted the AI, you can just handle it on your own. But if you're not sure, I want you to be able to selectively de automate, actually alert the person and ask for advice. Listen, there is this case, I don't know how to handle it. Give me advice. And I think this is the way that actually AI is going to learn and be trusted basically more often. So I do see actually AI applied in robotics also in lots of different hardware. I think cars will benefit from neuro symbolic AI because finally we'll be able to Overcome some of the common sense problems that actually plague it from going up more levels in autonomy. And I think in general we are going to be, you know, having something that is closer to the starter computer that is all annoying, that is actually going to give you a reliable answer. And we're seeing those kind of things, but I think it's going to get there. Another development is also goal driven AI. So one of the problems currently with LLMs is that they cannot plan, they cannot make a plan like an agent would do. So ideally you'd be able to give a task to an AI. It's bring it breaks it down into subtasks and then it actually starts executing them one by one or creating multiple copies and actually doing that. That's something that LLMs currently actually cannot do very well. They can do it to some extent, but not that well. So by having the symbolic component, it can actually do that. It's kind of goal planning, it can execute things. And having this agent computing basically at scale is going to revolutionize how AI works and how it interacts with us.

Because it's going to dig into that agent world. Because today we're seeing companies and people trying to do agents, right, using ChatGPT to spin up a whole bunch of agents that go and do different work. But it's not perfect, it doesn't work on its own very well. Right, Dig into that. What is the agent of the future and how is. Yeah, dig in on that because that's really interesting, this idea that you're going to be talking to it like a Siri and then Siri is going to go out to many different AIs to do different things and have them, you know, analyzing things or planning things, doing things. Tell me a little bit about agents, because that's really.

Yeah, so, so imagine in hybrid intelligence, the, the intelligence will actually check. Okay, this is a pattern that we've got, we've got this new task and this is the plan that we need to do in order to meet a particular goal or objective. Once it does that, it goes kind of into orchestration mode because he needs to create multiple agents, see how to actually solve each of those tasks and see what is the best way of doing it. For example, ask a website, ask another agent, you know, create a GPT, maybe doing, you know, write an email or even call a person and this kind of stuff. And obviously this thing needs to actually know what it's doing. I mean, I'm sure you've been frustrated talking to like chatbots of airlines or you Know different stuff detection sometimes give you really rigid responses. And this is part of the problem that in some cases, if you're giving it something that is repetitive, say for example, you want to scrape a stock price and then take some kind of action based on that, that's something that is very limited, like in, in domain and scope. So you can get that very done very well. Even with current systems, if you want to do something that deals with the messy world, you need the power of neural nets in order to be able to find out the exact pattern, the exact, you know, pattern of pixels that they need to interact with, which represents a person. They need to be able to speak, to generate the proper voice detection order to communicate with people. But you also need then something that really understands what it's doing, which is the symbolic part. And that is what is missing in all these agents that are based on GPT because again, they don't actually have a world model. And this is super important. For example, when you see all the spacecraft, for example, that land on different planets and moons, they all have a world model. They know exactly where the, you know, where the surface is. So yeah, I'm sure you've seen latest videos on video generation. For example, when you see that glass, you know, like falling on a tape and you see like liquid spilling out of it, but the glass stays intact. Symbolic systems will know that that is impossible to actually happen in the real world, that the glass is going to shatter. So this kind of expectation of expecting how someone reacts, you could say that that is actually one of the main things in intelligence, that is one of the main features that make something intelligent. Anticipating the results of your actions, what's going to be the impact. So we're going to see AI that is predictive, reactive. You've mentioned first. Yeah, we want the AI to teach us how to dance. But that would be, I think, one of the ultimate goals in robotics. It's incredibly difficult, but in order to do that, you need to know, to predict how your dance partner in that case is actually going to react. So you step back or you step forward and that is incredibly difficult for current AI based on LLMs to do. And this is what neurosymbolic actually solves, the antecedent of the future.

So you're, you're the guy who's going to teach my robot how to, how to teach me to dance.

Well, I need to know how to dance myself first.

But yeah, you can go to Houston and go to a nightclub. They do line dances every night. You can study that it probably will learn how to. Why haven't, why, why is the AIs that we've seen so far not this Neuro Symbolic, is it? A technical problem is takes more computer time and we don't have very much GPU yet to build these new models. Why, why hasn't this happened from the beginning?

So people have been trying to solve this for over 30 years. So there have been a small, actually quite small group of people who knew this all along. But it is actually technically very challenging to create an architecture that works. So along the way, in this four years that we've been through, we had to solve quite. We literally had to invent new maths and actually new frameworks, new theoretical frameworks to actually be able to support this. So the solution was not obvious. It involves quite kind of a lot of rethink on how things work. So I think this is actually why people have not managed to go and model a Neuro Symbolic AI that works at scale. Once you've cracked it, obviously it's kind of easy to see with hindsight, as with all things that actually have an elegant solution, but it's not an obvious solution. So one of the, you know, I can give you a little bit of information about that. The basis of the solution is based on graph computing on graph networks, but a particular type of graph networks called hypergraphs that generalize their kind of more complicated version of graph networks. So to explain it simply, normally a graph represents the relationship between one object and another object, but hypergraphs enable you to represent the relationship between one object and a group, and a group and a group. So yeah, this is like the simple version of it, but you can imagine that this is how it is and we don't think of things as just one object, but also of a collection. And this is how we'll be able to actually reason. The car needs to reason. You know, okay, I'm in a group of cars that is driving along the highway. If I'm gonna, you know, make a left and there's some car blocking me, I need, you know, to maneuver in advance in order to be able to hit the off ramp for up and this kind of stuff, which for us it comes out naturally, but for AI it doesn't. If you're walking, you know, and if you're hiking somewhere and you see a group of trees, you're not going to count each and every tree. This is kind of how current AI works. It will like try to obsessively count each pixel. You don't care, you just Want to know that there's, you know, some trees and that there's a stone, you know, just in front, that if you're going to hit it, you're going to fall down and stumble. Um, so one of the advantages of neurosymbolic AI is that it can focus on what matters and discard a lot of information. And you may think, okay, but why is he talking about discarding and forgetting information rather than learning it? This is kind of the converse of learning again. Another sign of intelligence, apart from anticipation is being able to forget and also ignore stuff that is not relevant. And this is where what we've created for them, explainable neural networks, they excel at it, at just focusing at what is most important.

Humans don't see everything, right? We don't see atoms, we don't see atoms vibrating on the table in front of us, right? We throw away a lot of information and say, oh, it's just a table, right?

Oh, yeah.

And same thing with AI and same thing with a lot of computer systems that they get better by throwing away a lot of information, right? If you go to cern, they, the first row of computers is all about throwing away information, right? To get to the thing that they're trying to find, right?

Yeah, yeah, exactly. And being able to focus on what is important, being able to notice what is different in real time, this is important in order to give AI autonomy. But autonomy, you know, you have unaligned to achieve alignment, actually. And this is very important because if you cannot actually understand and predict the consequences and behavior of something, you cannot align it in a controlled way. So while we do want to give AI autonomy, and this is going to be very powerful AI, much more powerful than GPT4 and all the kind of systems that we currently have. So we do have a responsibility to be able to control it. And by being able to have it explain itself, it's much easier to control something that you can directly, you know, put a rule to if it's, you know, if it's a system that you don't even understand how it's working, you know, there may be stuff in there that's, that's, you know, that's hidden inside that we don't even know about. So by being able to examine actually the AI, you'll be able to certify it, you'll be able to use it in mission critical systems. And I think it's, it gives a part a solution to alignment. And this is super important because, you know, I'm obviously on for AI development, but I do see that, you know, the people who raise objections about AI being out of control. I think there's a lot of hype, by the way, but I think there are some valid points and we do need to be able to control it in a safe way.

We need to understand why it made a mistake. Right. Google's Gemini puts different races, people onto leaders in the past, right? That was a mistake and we need to understand why it made that mistake. So we can fix the system to not do that, right?

Yes. So basically one of the ways that that are used today in order to kind of align in a way or put guardrails against AI, is what I said, you know, like before, is that you have a question, you have a query and you have an answer. So obviously there are division, there's the visual version of that, and then she gets, you know, sub query and you give a picture basically and the AI is trying to imitate what people are giving it to. Obviously if you overdo it and I think as we've seen in Gemini, it was overdone, then you get stupid results basically. So, you know, obviously even neurosymbolic AI, if like someone goes overboard and gives it a lot of rules, it is going to influence it and bias it, but at least you will know, okay, this rule came from a person and you'll know why it is biased. And if something happens, you'll be able to undo it much faster. So for example, if Google were using neurosymbolic AI, they would be able to much undo those rules much quicker than, you know, having to retrain the model all over again because this is what's happening now. They'll have to retrain the whole model all over again. So that's why there's going to be some time before we actually see Gemini and direction producing pictures.

Yep, it's a tough problem for a lot of people. A lot of computer science being being worked on right now.

Yeah, I think it's one of the problem of alignment, of control, of being to able to trust AI. I think it's one of the hardest problems that there is. It's one of the most important because you can get commercially useful AI in narrow problems. I mean, if you're doing some simple action, you know, like that doesn't require anything, you can actually do that without very advanced AI. And I think that has its place. But once you go into something that's complicated, that's actually truly useful, you get all sorts of these issues basically almost instantly.

Thank you so much for spending some time with me this really important. And it's a new area of A.I. that's coming. Right. And I'm honored to talk to you. Where do we learn more about what you're doing?

So we have a website, www.umni.com umnai.com and, you know, we've got also a lot of articles, blog posts, and also our LinkedIn page, which we regularly do.