Steve Omohundro and the Future of Superintelligence

Steve Omohundro is a computer scientist who has spent decades designing and writing artificial intelligence software. He now heads a startup corporation, Omai Systems, which will license intellectual property related to AI. In an interview with Sander Olson, Omohundro discuss Apollo style AGI programs, limiting runaway growth in AI systems, and the ultimate limits of machine intelligence.

Steve Omohundro Question: How long have you been working in the AI field?
It’s been decades. As a student, I published research in machine vision and after my PhD in physics I went to Thinking Machines to develop parallel algorithms for machine vision and machine learning. Later, at the University of Illinois and other research centers, my students and I built systems to read lips, learn grammars, control robots, and do neural learning very efficiently. My current company, Omai Systems, and several other startups I’ve been involved with, develop intelligent technologies.

Question: Is it possible to build a computer which exhibits a high degree of general intelligence but which is not self-aware?
Omai Systems is developing intelligent technologies to license to other companies. We are especially focused on smart simulation, automated discovery, systems that design systems, and programs that write programs. I’ve been working with the issues around self-improving systems for many years and we are developing technology to keep these systems safe. We are working on exciting applications in a number of areas. I define intelligence as the ability to solve problems using limited resources. It’s certainly possible to build systems that can do that without having a model of themselves. But many goal-driven systems will quickly develop the subgoal of improving themselves. And to do that, they will be driven to understand themselves. There are precise mathematical notions of self-modeling, but deciding whether those capture our intuitive sense of “self-awareness” will only come with more experience with these systems, I think. Question: Is there a maximum limit to how intelligent an entity can become?
Analyses like Bekenstein’s bound and Bremermann’s limit place physical limits on how much computation physical systems can in principal perform. If the universe is finite, there is only a finite amount of computation that can be performed. If intelligence is based on computation, then that also limits intelligence. But the real interest in AI is in using that computation to solve problems in ever more efficient ways. As systems become smarter, they are likely to be able to use computational resources ever more efficiently. I think those improvements will continue until computational limits are reached. Practically, it appears that Moore’s law still has quite a way to go. And if big quantum computers turn out to be practical, then we will have vast new computational resources available.
Question: You have written extensively of self-improving systems. Wouldn’t such a system quickly get bogged down by resource limitations?

Many junior high students can program computers. And it doesn’t take a huge amount more study to be able to begin to optimize that code. As machines start becoming as smart as humans, they should be able to easily do simple forms of self-improvement. And as they begin to be able to prove more difficult theorems, they should be able to develop more sophisticated algorithms for themselves. Using straightforward physical modeling, they should also be able to improve their hardware. They probably will not be able to reach the absolutely optimal design for the physical resources they have available. But the effects of self-improvement that I’ve written about don’t depend on that in the least. They are very gross drives that should quickly emerge even in very sub-optimal designs. Question: How would you respond to AI critics who argue that digital computation is not suitable for any form of “thinking”? They may be right! Until we’ve actually built thinking machines, we cannot know for sure. But most neuroscientists believe that biological intelligence results from biochemical reactions occurring in the brain, and these processes should be able to be accurately simulated using digital computer hardware. But although brute-force approaches like this are likely to work, I believe that there are much better ways to emulate intelligence on digital machines. Question: The AI field is seen to be divided between the “neat” and “scruffy” approaches. Which side are you on? John McCarthy coined the term “Artificial Intelligence” in 1956. He started the Stanford Artificial Intelligence Lab with a focus on logical representations and mathematically “neat” theories. Marvin Minsky started the MIT lab and explored more “scruffy” systems based on neural models, self-organization, and learning. I had the privilege of taking classes on proving lisp programs correct with McCarthy and of working with Minsky at Thinking Machines. I have come to see the value of both approaches and my own current work is a synthesis. We need precise logical representations to capture the semantics of the physical world and we need learning, self-organization, and probabilistic reasoning to build rich enough systems to model the world’s complexity. Question: What is the single biggest impediment to AI development? Lack of funding? Insufficient hardware? An ignorance of how the brain works? I don’t see hardware as the primary limitation. Today’s hardware can go way beyond what we are doing with it, and it is still rapidly improving. Funding is an issue. People tend to work on tasks for which they can get funding. And most funding is focused on building near term systems based on narrow AI. Brain science is advancing rapidly, but there still isn’t agreement over such basic issues as how memories are encoded, how learning takes place, or how computation takes place. I think there are some fundamental issues we still need to understand. Question: An Apollo style AGI program would be quite difficult to implement, given the profusion of approaches. Is there any way to address this problem? The Apollo program was audacious but it involved solving a set of pretty clearly defined problems. The key sub-problems on the road to general AI aren’t nearly as clearly defined yet. I know that Ben Goertzel has published a roadmap claiming that human-level AGI can be created by 2023 for $25 million. He may be right, but I don’t feel comfortable making that kind of prediction. The best way to address the profusion of ideas is to fund a variety of approaches, and to clearly compare different approaches on the same important sub-problems. Question: Do you believe that a hard takeoff or a soft takeoff is more likely? What actually happens will depend on both technological and social forces. I believe either scenario is technologically possible. But I think slower development would be preferable. There will be many challenging moral and social choices we will need to make. I believe we will need time to make those choices wisely. We should do as much experimentation and use as much forethought as possible before making irreversible choices. Question: What is sandboxing technology? Sandboxing runs possibly dangerous systems in protected simulation environments to keep them from causing damage. It is used in studying the infection mechanisms of computer viruses, for example. People have suggested that it might be a good way to keep AI systems safe as we experiment with them. Question: So is it feasible to create a sandboxing system that effectively limits an intelligent machine’s ability to interface with the outside world? Eliezer Yudkowsky did a social experiment in which he played the AI and tried to convince human operators to let him out of the sandbox. In several of his experiments he was able to convince people to let him out of the box, even though they had to pay fairly large sums of real money for doing so. At Omai Systems we are taking a related, but different, approach which uses formal methods to create mathematically provable limitations on systems. The current computing and communications infrastructure is incredibly insecure. One of the first tasks for early safe AI systems will be to help design an improved infrastructure. Question: If you had a multibillion dollar budget, what steps would you take to rapidly bring about AGI?
I don’t think that rapidly bringing about AGI is the best initial goal. I would feel much better about it if we had a clear roadmap for how these systems will be safely integrated into society for the benefit of humanity. So I would be funding the creation of that kind of roadmap and deeply understanding the ramifications of these technologies. I believe the best approach will be to develop provably limited systems and to use those in designing more powerful ones that will have a beneficial impact. Question: What is your concept of the singularity? Do you consider yourself a singulitarian? Although I think the concept of a singularity is fascinating, I am not a proponent of the concept. The very term singularity presupposes the way that the future will unfold. And I don’t think that presupposition is healthy because I believe a slow and careful unfolding is preferable to a rapid and unpredictable one.