(You are reading the English original, translations are available here: Deutsch)
Things are extremely busy these days. The last twelve years I have been working on the fundamental principles of AI, what it is, how it works, and how solutions can be designed with it. But the time has come to tackle the next step: building a company that commercializes the AI tech and in the process shaping a venture that, possibly, can change the world.
But everything starts small and it means lots of work to build the business step by step. Since the beginning of the year I contacted potential clients, looked into ways to acquire funding, and actively participated in the networking some prosperous regions here in Germany have begun to foster with the aim to promote the application of AI.
I had two goals that I wanted to achieve in this initial step. First I needed to identify a concrete project that I will work on. This first project should showcase what the AI can do, ideally it should demonstrate features that wouldn't be possible without AI. More specifically, as the AI I have developed can do things the biggest competitors in the world can't do, the showcase should exactly demonstrate this. This project also needs to be commercially viable, which means the time it needs to completion, the work and investment required, and the return on investment all matter. Other big factors are how good the tech being implemented can be protected, it is the company's trade secret and its main capital after all, and to what extent I remain in control of the venture.
The second goal was to find an ideal place, a city, a region, a country, where the company will initially be set up. There are a lot of factors to consider, like the infrastructure, government, tax regime but also how open a region is to foster new technologies. What also matters is how easy it is to relocate there. The availability of public funds helps here. But there are also barriers to consider like moving to a different country puts additional legal constraints on you, the language may be different, and it could mean a shift to a different cultural context. As an employee you can usually tackle such issues in your free time, but as a self-employed entrepreneur you need to be capable to focus as much as possible on your venture.
In short it meant I had to identify a viable way by talking to people and learning what type of project is ideal and what each region offers to implement it.
Making a decision
There were a number of potential initial projects I looked into ranging from space-faring MMO that would benefit from an AI that is capable of making strategic decisions, i.e. taking all the context in, and automating features of the game, to scientific research that allows a more detailed and realistic assessment of climate change. If you look at this article about the tipping points you understand the challenge.
What won out in the end, basically because it could be implemented in twelve months, requires only modest investment, keeps the control of the tech firmly in my hands and offers a large scalable profit, is a strategy game for PC. It is a unique opportunity because the foundation for the game exists already and a lot of the AI tech I have developed for an earlier prototype can be reused and ported. It is also ideal in the sense that I can develop the game to a large extent in-house and keep the source code close. But the decisive factor is that the gain per time/work/money invested exceeds other projects. It allows me to turn an investment of 250,000 € into a profit of 5 Million € within twelve months. This is a realistic assessment of a reasonable case, in terms of market and demand, and not by far close to what could be best case sales, which could be 10x times higher. Such figures aid in the search for private investors, and eventually provide the company with a solid foundation to expand and meet its goals.
When I searched for a good location to set up the company, a press statement caught my attention. It was about a regional GamesHub being set up in association with a promotion of games that have dual use, i.e. as an entertainment and a tool to further education, medical treatment and other more serious uses. Given that AI is dual use per se, as it is universally applicable to a wide range of uses, I followed up and investigated. It turned out that the GamesHub was not established yet, and wouldn't help me all that much, but that the region was in the process of setting up a support framework for companies working on AI. Plus this region had already a good infrastructure with local technology centres involved in networking and promoting cooperations. While there weren't any funds available to support AI start-ups, the region has a program to provide funds to game development studios, which in my case could be used to finance 80% of the prototyping costs. As the location is in Germany as well, there aren't extra barriers that require negotiating. Essentially there were a good number of reasons to choose this location.
Investment in AI and games
When you look for private investors for your start-up, it is mandatory that you have a good proposal and can make a credible case that it will work.
A friend asked me why I believe that my AI is better than anything Google and other big corporations have, and of course why my project will be a success.
The answer turned out somewhat lengthy, but I think it is worth sharing here because many people asked questions that go in the same direction. It has a lot to do with what AI truly is, and what currently is hyped as AI.
When I studied computer science at the TU Braunschweig, in the late 80s, people were first excited about neural networks (NN), but then quickly came the disillusionment that they are highly inefficient and unreliable. When Google turned its gaze to AI twenty years ago they didn't let themselves be deterred by this scientific consensus and said, we now have supercomputers, and we "solve" the problem of inefficiency by investing massive computing power (about 100,000 times more than you would normally need). But the thing is, if you look closely, NN do not much more than substitute programmers, that is, programming is replaced to some extent by NN. It is an automation of software engineering. Of course, you can say programmers are intelligent too, so it's also an advance in AI. To their credit Google is also exploring this technology methodically by hiring many people and providing supercomputers.
But in reality, they don't know how AI actually works. I started working on AI at a similar time, but my approach was to understand how it works. I had max points in my A-levels in math and physics, worked for many years professionally in that direction to develop innovative simulation and abstraction techniques, and then eventually found an approach how to develop true AI and systematically followed that through. I did this for 12 years in a mode in which I practically worked 4,000 hours a year, twice what normal working people manage. So in the end, I have worked 24 years in real terms on the real progress of AI, while Google and others have had 10,000 and more people working on the wrong approach. If you like, the lead I have on them by now is beyond what they can imagine.
Google of course did one thing, they portrayed themselves as the market leader for AI, and told people that AI can do many things now. I've been following that, of course, and yes, all sorts of governments and big corporations believe this and are funding this approach of AI based on NN. But I know what they can't do (true artificial intelligence) and what they can do, which is really modest, in effect some select applications for which they need supercomputers plus limited know-how for machine learning.
Even if everyone else gets caught up in the hype, I don't! To me it's a strange situation, it's like Google has created a market for AI (by letting people know there's something coming) that they can't really serve. On the other hand when my products will eventually be ready, are a thousand times more efficient and can really do intelligence, i.e. they can understand problems most people fail at, Google will have to do a reality check while I say thanks.
The same goes for raising start capital, yes there are all kinds of venture capital funds now that love to invest in AI, but on "their" terms, which usually means they take control. Most entrepreneurs would probably say to themselves, well, to a few millions I won't say no even if it's actually a one-sided deal. But for me, that's no longer interesting, because profitability is now within reach, in a good twelve months. So I prefer to do it this way. And basically, I can only do it this way because the AI is sufficiently advanced that it can be used to produce a game that (a) would otherwise consume significantly more development costs and (b) has features that players won't find anywhere else. It's actually a sound bootstrapping process, evidence inclusive.
Even if you don't agree, there is now a wide gulf of difference between my view and how the world views AI today. The only way to reconcile this is to follow through. And as an entrepreneur it is a good bet.
How AI is seen today
The networking I participated in and the talks I had with potential clients taught me a lot about what people today think what AI is.
I had talks with Professor Clarissa Vogelbacher from ITM-predictive and the people responsible for the MMO Eve Online, a successful and vastly complex game. These people absolutely are not stupid, and yet I failed to convey what true AI actually is. They simply have preconceptions that I couldn't penetrate. In my talk with Eve Online it was possibly my fault as it was my first talk to present the topic, but in the later talks it was definitely that they had difficulty to imagine what a true AI actually is, how it works and what it can do.
What was interesting here is that they typically construct a conventional frame of reference, like the model of a car sharing company including their customers, different city locations, fluctuating demand depending on time, location and events. And then they use machine learning (NN) to augment the model with a prediction, for example to estimate the demand for cars at different times/locations, in order to come up with a flexible and adaptable pricing.
This is to a large extent how AI is seen today. Yes, this is useful but also very limited. You can imagine more complex systems with multiple machine learning modules to serve different functions and provide interfaces to different data sources. But these machine learning modules are technically black boxes that can only be trained to adapt to a few parameters. In many ways these machine learning modules produce heuristics. It's mostly an aspect of automated software engineering, i.e. you can save the cost on programmers to implement these heuristics. But these heuristics don't provide intelligence beyond that.
There is one thing that bothered me in a lot of discussions: that people behaved like the onus is on me to prove to them that my AI works and how it does it. You don't work twelve years to develop an advanced new technology and then tell everyone how they can make it work on their own. Even if this would be that simple and there wouldn't be ethical considerations. It's the company's capital after all. What I am responsible for as an entrepreneur is to provide clients with a product that is as good as I say, to be worth the asking price.
How I see AI
Whereas NN modules, or machine learning, are seen as black boxes that have to be trained, my work focuses on how these work. In other words training tries to shape these black boxes from the outside. What I am doing is understanding and developing methods to shape the AI, what supposedly is in the black boxes, directly.
In principle it is easy and straightforward to write heuristics for a behaviour that is dependent on a set of parameters. What is much more demanding is to implement a behaviour that adapts dynamically to a continuously changing context. A context that includes the entire environment, like the battlefield in an MMO. Such an environment contains many objects of which each has its own distinct behaviour. Which means all these objects impact each other, like an increasingly complex differential equation. The technical solutions required for this have to take into account the entire continuum, very much like Albert Einstein's general theory of relativity. It is this environment, its adaptability and solutions for elements that affect each other, where true intelligence resides.
My work on AI is focused on this continuum, mechanics and solutions that involve all the objects it contains and how these interact. With those it is possible to design intricate AI systems that can model everything and address all kinds of questions. Questions can be as complex as what strategies a general has to choose to win a battle, or how a town as a whole could become more sustainable.
With other words it is a universal toolbox that you shape as you see fit. Intelligence gets a new meaning. It isn't just a heuristic to predict a data flow.
The system I have researched is similar to how Nicola Tesla viewed electromagnetic fields to build advanced electric machines. You can view localized electric and magnetic fields as finite elements that affect each other through induction. Applied to AI you have finite elements that all have a state and interact dynamically with their environment: Tesla Minds.
In principle, whereas Google tries to shape a black box from the outside to train heuristics, Tesla Minds shape this black box from the inside and give it intelligence, they are its fabric, its intelligent building blocks. It gives you a wealth of new highly advanced tools and an efficiency that is unmatched.
If you can work with the AI directly, you don't need to shape it from the outside like a black box any longer. If you so will, I have all the know-how Google doesn't have, the part inside of the black box to build true AI.