As a physicist by education, electronics and computer geek at heart, with 30 years of business and information technology experience, Andy is committed to deliver innovation at the edge of today’s technological frontier.
At OneForce, he leads a team of talented scientists and engineers developing cognitive technology and directly applying it to build intelligent software for marketing, sales and recruitment automation.
In today's presentation, we'll be sharing OneForce vision about the alternative path towards Artificial General Intelligence. My name is Andy Zhulenev. I'm the CEO at OneForce.
I will talk about the state of AI today and what different scientists think about how AI can evolve towards AGI. We will also share OneForce's vision about alternative paths toward AGI. We'll talk about the BrainCore system intelligence system that OneForce is building. And we'll also talk about SmartLeads intelligence software.
So about the state of AI today.
Let's take a holistic view of what has happened in information technology over the last 75 years. Computing, as we know it today, actually started around 1946 with a computer called ENIAC. For the purposes of this analysis, we would take the beginning of the Artificial Intelligence era as 2015. That's when image recognition by AI matched human level and opened up new opportunities such as self-driving cars. So, we are at the beginning of this Artificial Intelligence era, and people believe that in a certain number of years, probably more than 10, it may be less than 100, we will reach a point of Artificial General Intelligence (AGI) when capabilities of an intelligent machine will be comparable to a human brain. Now, before we go and talk about AI, let's look at what happened in computing in the last 75 years.
The evolution in computing was going across multiple layers of technology. It started with technology itself, like how central processing units were built using tubes initially, then transistors, then chips, and then 3D chips. Then in telecom space, I mean, we may still remembered analog days, but then switching to digital and optical cables with full spectrum use of optical cable capabilities. In storage, we went through an evolution from tape to hard disc to CD-ROMs to SSD. And a similar evolution has happened in hardware with data centers becoming public clouds. Endpoint devices, PCs becoming mobile devices, tablets. Now in the networking world, there has been an amazing evolution in terms of long-distance, as well as the local network from 1G to 100G, and in internet from DSL to satellite and in mobile phones to 5G over time. In software, we have seen a huge evolution in terms of programming languages, as well as in operating systems from Unix, Windows, Linux, Red Hat, and in middleware. And then in applications, when it started, there were mainframe applications, then client-server, web, and now zero-code applications. Now all of this happened over 75 years, and these technologies were built layer by layer. So, if we will be talking later today about the path for Artificial Intelligence going forward, then probably we can assume that this AGI technology should be built also layer by layer and not immediately.
Now the good news is projections about computing power affordability. In the coming years, by 2025, an affordable computer will have human brain-level processing capability in terms of computations per second. And very soon it will even exceed the computing capabilities of all humankind. Computing power is not the challenge here. So, then it is all about software for AI.
There are several scientists like Ray Kurzweil and Jeff Hawkins who have been working on a vision of creating a digital brain - a neural network that can do universal things. And there have been quite several interesting discoveries on this path, such as frames and how the brain works and analysis of human neurons, and trying to extrapolate and project how it can work in computing. Now, I was also very excited reading these books and imagining that at some point intime using an MRI to scan my brain and have the information uploaded to a computer. But as a practitioner who, at some point developed in assembler and C++, I'm just thinking, what if one line of code in the brain is misread? Will we get a blue screen? And so, building a universal system might be too big of a challenge right now. I think that while there are some amazing ideas and suggestions in terms of how such a digital brain can be built, we probably need to look for a more evolutionary layer-by-layer path.
There is also a set of scientists that already helped us understand that today's AI in itself will not develop into AGI. And that it is called AI is, to great extent, a misconception. Today's AI (such as machine learning, deep learning, or NLP) is a black box trained on big data. And it has certain input and certain output, but it cannot talk to a human expert. It cannot learn from a human expert in real-time. It doesn't have those knowledge transfer capabilities and it is good only for specific purposes. So, if a system has been built to play Go, then it will play Go, but it will not play chess. And a system that won against a world champion in chess will not be able to play go. These scientists also talk about the need to create a common sense for AI, because we as humans, when we read something, we fill the blanks with our common knowledge about how the world works, and we don't need to explain everything in every sentence. Now, the AGI system also has to accumulate this knowledge to be able to learn and make certain decisions in the context of common sense.
Now, what about the OneForce vision as an alternative path toward AGI?
If we step back and look at how homo sapiens emerged in the animal world and differentiated themselves from chimpanzees and other animals, well, that happened with language. This happened about 200,000 years ago. With language, humans first were able to communicate. They were able to form groups and that made them the most powerful animal on the planet. And homo sapiens moved up to the top of the food chain. Now, language eventually evolved into writing and books, and today, it is a digital age. But language allows us to store information in writing and one generation can learn from inventions that the previous generation has done. So, language is not only a method of communication, it's also a method of storing knowledge and transferring it between people, between generations.
One of the great discoveries was done by Jean Piaget, a Swiss psychologist who developed theconcept of a child's cognitive development. It talks about four stages. The first one is sensorimotor at the age of up to two years, the baby can do certain actions and the baby has senses. The next stage of this cognitive development comes when the child can speak and comprehend language. So, language is the second most important phase of a child's development after the senses have been there already for two years. And then another important discovery here is that language becomes the basis for a person to learn about the world and eventually, be able to do logical thinking at the age of seven. And at the age of 12, people develop hypothetical thinking, which is no longer about concrete objects but hypothetical objects, something that's used in science. So, looking at this, if human development goes through these stages, it would be probably logical to assume that language and mastering language comprehension is what artificial intelligence needs to do first before moving to more complex things.
Again, the good news is that today's AI has already mastered multiple senses. One of them is image recognition, and that happened in 2015 with the ImageNet Competition. Speech recognition with Switchboard in 2017 and language understanding got on par with a human level of capabilities around the 2019 - 2021 timeframe with GLUE and SuperGlue competitions. Now, these senses take up about 40% of a human's brain. About 30% of all neurons in the cortex are used for visual processing, 8% for touch, and 3% for hearing. Now, while these sensory capabilities of today’s AI are pretty capable, they are not still at the same level as humans. When we open our eyes and look around, we move our heads but the image that is in our heads about our surroundings stays. This means that our brain is processing the surrounding world, building the virtual representation of it in our head, and doing image stabilization. This is not what today's self-driving cars can do. They are creating for themselves a simplified virtualization world of the surroundings, not yet at the human level. So, it would be kind of safe to assume that today's AI capabilities can do about maybe 10% of what the human brain can do in sensory functions specifically.
Now, if we look ahead, then what we see is that today's AI already can-do sensors, it can do language understanding. And so, the next step on this journey should become language comprehension. There is a big difference between language comprehension and language understanding. Language understanding technologies today are still black boxes, they canpowerfully translate from one language to another. They can transform text from a long article into a short form. But all of this is language transformation without really understanding it, without comprehending it. So, this is not what's needed, let's say, for a machine to communicate with a human. There needs to be a level of language comprehension. And that's in our view, the first step for an intelligent system to master. Later on, the next step would be to develop logical thinking based on this language comprehension. And a certain level of capabilities of understanding the 3D world such as vision and creating a representation of it may also be required. We will focus in today’s presentation on the language aspects. And then, eventually, it would move to hypothetical thinking, which is the next level of abstraction.
Now, with image recognition getting on par with the human level, AI gave us self-driving cars. While the work on those self-driving cars started much earlier, we are at the point where multiple car companies are testing self-driving cars under different conditions. And eventually, they may get on the road, maybe not in all areas, maybe with certain conditions and certain levels of autonomy, but we made significant progress in space. Speech recognition gave us virtual assistants such as Amazon Alexa, Siri, and Cortana. And these tools are used by people everywhere and they made a significant impact. So, what we believe is that language understanding technologies are becoming that steppingstone in creating a set of new intelligent systems that can do knowledge work automation. Having language comprehension capabilities would enable these systems to take on many initial low-end tasks in the knowledge work. And that's the focus of OneForce.
We are talking about knowledge work, that's about 1 billion workers globally. There is also physical labor where there are about 2 billion workers globally today. The total working population on this planet is about 3 billion. And so, for intelligent systems in the area of knowledge work, they would need to have language capabilities before moving to logical thinking and hypothetical thinking. If we look at robotic systems, they, generally speaking, would be going mostly by senses, such as vision and touch to enable what they need to do. But at the same time, they need to develop logical thinking at some point in time and be able to also learn from people. And so, to do that, they may still require language capabilities and to build that level of logical thinking. So, it may still be needed even in that space.
OneForce is building a BrainCore intelligent system.
It is a cognitive technology that at a very high level consists of a layer of comprehension, a model of the world, and decision-making capabilities. The goal of the system is to read, understand and make logical decisions.
Now, the model of the world is the core component of such a system to accumulate human expert knowledge and be able to use that knowledge to comprehend what it reads in the text. It can initially be focused on specific business domains or technical domains. For example, it can have a knowledge graph about relationships in business, such as a company, a seller, or a buyer of the product. Accordingly, those other companies can become a customer or a vendor to this company or some companies can be competitors or partners. If a person is working at a certain company in a certain role, that can be represented in the system. And so, when such a system reads the sentences “person A works as a marketing manager at a company B”, it will be able to understand that information in the context of this model of the world.
Now, initially, this system will learn from experts. But we see a path where eventually, it can learn also from data. Creating new structures of the world and enriching them from information obtained from text, obtained from data.
Now, while this system brings certain universal capabilities, we are working on specific business applications of this technology.
And the very first of them is SmartLeads intelligent software, which is designed for sales development automation. It acts as intelligent software and can be used by people to build a much bigger database of all potential customers, find relevant people, and segregate unrelated people. And then, micro-segmenting those people that are relevant based on their specific roles in their organizations or the type of organization they work for. The system would support outreach to those people using multiple channels, such as email, social media, and cold calling.
And so, what BrainCore enables us to do in the system is to introduce a new method of search, which we call semantic search. Traditional search as we know it is keywords based. We enter a couple of keywords, and the system returns relevant results, let's say companies or people. But this search is not quite accurate because, what is there in the sentence, meaning can still differ in many ways. And so semantic search eliminates that uncertainty because it is looking at not just the keyword itself, but it is looking at the whole sentence, deriving the meaning of what is described as an action with that specific keyword.
This is a SmartLeads application and as part of it, there is a database of companies. Now, each of these companies has a description. We have over 10,000 companies in this database. Let's see how semantic search can help us in finding relevant companies.
To start with, we need to specify what we're looking for. And in this case, we are looking for companies doing social media management. What this request is telling us is that we are not just looking for “social media” keywords in sentences, but we're looking for those sentences that describe a company, doing something with social media.
And so, when we review the data, we discover that the data in our database has multiple different meanings. One meaning among all the sentences in this company description is about “company doing something with social media”. This is relevant for us. But if a sentence says that “someone is doing something with social media”, but it is not specifying that it is a company, then those sentences will not be relevant to us, and they are marked as irrelevant. Now, a company using social media is also relevant for us. And if a sentence says that the company is a social media management company, those cases will also be relevant. The company maximizes social media performance, for example, is also relevant. But if we find a sentence that has a keyword, such as “social media” in the company name, we've seen that it doesn't mean that this company is doing social media management. They may be just blogging, training, or a recruitment company. And so, we mark such meanings as irrelevant.
Now, with this selection of meanings, there will be a pretty limited set of meanings for a big database with different sentences. The system will be able to present to a user, all the companies that it found with those relevant meetings marked as “yes” here. And so, Celtic Marketing here, for example, is saying “we are a creative social media company," and that's the meaning “company is social media” in a normalized sense. "VI Media 2020 is using social media platforms. VI Media will help you expand your brand." - This sentence is also relevant because it's about a company using social media to achieve something. But here is another interesting example - HSBC, a well-known bank, in its company description has a statement "to view our social media terms and conditions, please visit the following webpage." Now it does have a keyword “social media” and with a normal keyword-based search, this sentence and this company will come up as a relevant result. But with this meaning analysis, we see that there is no action in this sentence done by the company. It is done by somebody else and alone by specifying that “someone” does social media, we can classify the sentence as not relevant.
And so, from here, we can go back to the database and see those companies. With this semantic search, it becomes about 8,000 companies, or about 30% less than we originally had. This is a functional prototype of what semantic search can give as capabilities and how powerful they are in managing databases of companies. And pretty much in a similar way, it can be used for a database of people.
With that, we are at the end of today's presentation.