AI and Ethics
In the past twelve months a lot of questions have been asked. The more straightforward questions have been where do I plan to take my company (Tesla Minds) and what does Tesla Minds' AI offer to customers.
However, as my work revolves around breaking new ground, pushing truly the envelop what we can do with physics and AI, a whole set of more fundamental questions have popped up. Questions like what can we do vs what should we do. If we accept that AI can transform our society, which is currently at a critical junction – will we face up to the challenges ahead and deal with these humanely – there is also the question, how does truly advanced AI need to be deployed as a generational tool: a technology that we develop over decades to shape the lives of a new generation of people. Who should own this technology? What should it enable and what not?
The discussions I had covered a wide spectrum, some business-minded, some scientifical and philosophical, and others controversial. But they all share a common element, people asking questions. Thus it is only logical to distill the essence of these discussions into the form of a Q & A session.
Q: Can you give us a brief overview what sets your AI apart?
My vision of AI is a system that recognizes the structure of an entity, derives its possible behaviour, and optimizes it for the entity to attain a given goal.
This is an overarching logical definition. Technology comes into play to implement its aspects, to make AI a reality. The main difference to existing approaches is that once you get your foundation right and have a universal solution, the world is wide open, you can tackle everything.
Q: This sounds deceptively simple. Every scholar can formulate a similar universal claim.
At Tesla Minds we develop the building blocks to make it happen. Ultimately we build the hardware and software to implement the AI for a client, a robot, the control unit for a vehicle or software that tells you what is what.
All applications have a common core, generic means of recognition, an interface to infer and control behaviour, and the capability to optimize an entity's behaviour to attain goals.
Our research reflects this. We have a systemic approach towards researching all the necessary universal building blocks. Once we have these, we are capable of implementing generic AI.
Q: What kind of research are you doing?
Nature has gifted us with a universe that is full of rules. Rules that underpin everything, order and chaos battling for supremacy. There is inherent intelligence on many levels that you can learn from, if you understand its fundamentals.
Many people think of biology if they hear this. It is interesting that Albert Einstein believed that for something so complex as organisms to work out on a high level, there must be a fundament to make it work.
Roughly fifteen years ago I started looking into this direction. I am as much a physicist as a computer scientist and besides reading about the cornerstones of Einstein's ideas I was always fascinated by Nicola Tesla's pursuit of practically advancing modern electrodynamics. There is this elusive link between electromagnetic fields and time and space. And once you have time and space you also have gravitation.
Coming from the direction of researching artificial intelligence, it one day became clear to me, that once you understand how electromagnetic fields, time, space and gravitation come into existence, you also hold the key to understanding how the universe orders itself. Kind of a long shot.
Q: How does this apply to the development of software?
Quantum mechanics is not as far removed from software development as many people might think. At the heart of quantum mechanics is the concept of interaction, and you can easily arrive at a model of software that is interaction-centric. Many quantum particles have well defined properties and the symmetries these have to fulfil have equivalents in integrity conditions that are important for designing reliable software systems.
Order plays a major part in designing advanced AI. The order in which you rank events, which is another term for interactions, and the order in which you probe different possibilities are of paramount value for a highly performant AI. So it applies not only to efficient data management, but also to the organization of threads and processes.
Going beyond time and space to understand the essence of order in the interaction of quantum particles can give you invaluable insight.
There is also the conservation of energy and momentum, which is in principle a property of our universe that stipulates to do everything as efficiently as possible. Understanding how this works, the mechanism that is behind it, enables you to design AI differently. It opens up a new world for you to build AI that finds maxima and minima with a minimum of computing. It is invaluable.
Q: What are you currently working on?
I wished there would be a more straightforward answer to this question. I have asked myself instead the question what makes most sense to work on.
In terms of software, that is the practical application of AI, we have decided to work on a project that offers the best ratio between gain and effort required. This mostly means profit, because a company needs fuel to go forward.
The type of AI we are developing is conceptionally so complex that only a few people can develop it. As it is these are the same people who also understand the underlying research in depth. So there is a competition for human resources.
In order to answer your original question, I am working mostly towards fulfilling our strategic mission. I believe we are in a similar situation as in Interstellar, as in Professor Brand sketching out a few possible solutions to be achieved with limited resources. We need AI but if we could advance our understanding of physics to master gravitation, it would be even more valuable.
In practice I am working on fundamental physics if the situation allows it. The transfer of results to AI core design happens mostly instantaneously, i.e. it doesn't consume much time, it is nearly as if these two fields are linked to each other by shared principles. Software development has a lower priority but has its place too. In practice we have set milestones but have less rigid time constraints. This is currently possible, but if we had a client who expects a delivery, we would have to change our mode of operation.
Q: Is it really feasible to go for a big scientific breakthrough that has eluded many brilliant physicists for a century?
I believe the angle I am coming from is unique and allows new insights that physicists didn't have. Actually, many of them didn't have computers.
While you never can tell how much work is left to complete a new scientific breakthrough, the work has yielded so many insights already, that it is definitely worthwhile even for a company developing AI.
Physically we have already a first model that incorporates electrodynamics and gravitation on a common foundation. It needs to be seen how much work is still required to address all physical aspects properly, but in terms of AI it is sufficient to propel us forward for a long time already.
Q: Would it not be better to work for one client and to produce one advanced AI that gives you recognition worldwide?
This has indeed been discussed and it makes sense, to an extent. The advantages are obvious, particularly in respect of earnings and establishing the company as a premier supplier of AI.
But if you look more closely at it, we are not loosing much by strengthening our know-how through research. We loose some time and earnings. On the plus side, we are avoiding the crunch of having to decide which part of the technology we transfer and what we keep in house.
Q: Who owns AI?
Currently the state of the art is still in its infancy and to a large extent public, research, white papers, books, workshops, best practice showcases and the like all play a part. To that extent the technology is still free.
Entrepreneurs have begun to take ownership by developing applications and maintaining control by owning the hardware and/or software. This is inevitable because if you make a significant investment you want to see a return.
For us the state of the art isn't useful. We leave it behind to open a new frontier. In principle we are taking ownership of it by developing theories, concepts and algorithms in house.
We understand that this creates issues. Is it right that a company takes ownership of a transformative technology? Ideally we would like to share the AI we develop to an extent that makes sense. But what that is isn't easy to answer.
Q: Do you see ethical limits?
If you see the potential of AI, what it could be if all technological challenges are solved, it is frightening how it can be abused.
If you understand this, it becomes clear that if you go all in to solve all technological challenges and then go about to decide what we will do with it and what not, it will be too late.
Which means you have to keep a lid on the know-how, either by limiting access to it or intentionally limiting the areas of research. Both approaches are extremely difficult to maintain.
We don't have an answer for this. For now we are keeping our research in house and focused on fundamental understanding of AI. Based on that we decide what applications we want to develop with the know-how and tech we have. We try to err on the side of caution, consciously stepping away from applications that might lead to areas that could become problematic. Coincidentally this turned out also to be a sound strategy to protect our know-how and we stand fully behind it.
There are definitely areas that you don't want to develop for. We have no plans to develop artificial consciousness. We also strive not to develop universal AI building blocks but rather build AI as application-specific algorithms.
Q: What is technologically possible?
What we do is based on the twin progress of two disciplines: AI and quantum mechanics. In many ways progress in one discipline helps progress in the other discipline. For this reason we see AI as a key that drives our understanding of the universe.
You could even say that we are on the brink of understanding quantum gravitation and then building quantum computers with it that are small enough to fit into nanomachines. Once you have a breakthrough like this anything becomes possible, even control of gravitation.
This is such a stark vision that we are at a loss whether such a breakthrough even makes sense, or how to make sure that it makes sense.
Q: Are we ready for AI?
This is a truly important question. During the last 2000 years we have proven again and again that we are not capable of handling power responsibly. Posing that we need more power now to deal with the mess we have created, just doesn't check out.
If our research is successful, it will be the main mission of the company to figure out how the technological breakthrough and resulting technology can be deployed responsibly.
My best idea here is to develop technology step by step as part of applications that fulfil clearly defined functions.
Q: Can you say more about the project you are working on?
Our current project is a playground that experiments with an AI that can discern the structure of complex mechanics and develop strategies how to use the means available to its benefit. It also asks questions about the destiny of humankind and how a truly scientifically advanced civilization deals with the challenges that come with advanced technology.
It is all software, developed entirely in house, and has a commercial scope. It will be an application that people can try out and toy around with.
Q: Is there anything else you wish to share?
During my time at Ubisoft I worked on a PC game called Heroes of Might & Magic V. Every gamer who is into turn-based strategy games knows it. It is in principle a wargame in a fantasy world in which you have to manage resources, build castles and deploy armies. It is quite complex and challenging.
Many years ago I began to develop an AI that is up to match its wits against human players. Now, many years later, I wish to finish this work. It is another good showcase of what properly designed AI can do. The idea here is that the AI does full strategic planning, by taking all aspects into account, and then goes step by step to fulfil its strategic goals.
Scientifically, there is more to it than meets the eye. Once you have an AI that can handle a large set of rules, discern situations methodically, and then go about to formulate a strategy to attain its strategic goals, you have created the blueprint for an application that can handle many different situations.
But just as importantly, I wish to fulfil the promise that I made to true fans of this strategy game. For an overview of what is coming and why it is an achievement in terms of AI, please check out the post on strategic planning.