Ymir as Physical LEGO

Kris Thorisson

Is it possible to build physical models of computer models? Sure, it just takes an aweful lot of LEGO blocks.

When I was at LEGO my team and I developed an artificially intelligent character called Puff the Magic LEGO Dragon. I spent a lot of time explaining the workings of the Ymir architecture that powered Puff's mind. It was hard. After several minutes of giving the big picture I very often got a blank stare along with the question "but where is the intelligence?"

In the last few years I have heard this question more and more often. It doesn't only come from the general public but also from seasoned computer scientists, philosophers and psychologists. One of the most difficult things about explaining computer systems that simulate intelligence is that often these systems encompass a number of relatively complex subsystems that interact. Explaining each subsystem in turn does not explain the full system because it leaves the question open about how their interaction makes the whole system behave in some particular way, to produce its purported intelligent behavior. This is mostly likely also the way that human and animal intelligence operates: Intelligent systems don't really have any single part that we can point to and say "there, that's the 'intelligence' center". Rather than being centralized in any one place in the brain or encompassed in a single "intelligence principle" in some central "self" or "consciusness center", intelligence is more like the net product of an ecosystem of interacting elements, where the behavior of the whole system is considered intelligent but none of its small parts are really that smart.

So after spending countless hours explaining Puff to visitors and coworkers I realized that without some learning aid, this process would go on forever.

To remedy this I built a LEGO model of Ymir, the architecture that powered Puff (as well as Gandalf, another autonomous communicative agent that I built, at MIT). Because Ymir is such a modular system it turned out great as a physical LEGO model, especially once I figured I could use radio and colored antennas as a metaphor for communication between the different kinds of modules. As a result of this experiment the whole task of explaining how this thing worked got significantly reduced - with a physical model, people "just get it". The principles behind this effort are now being generalized into a foundation for modeling complex A.I. systems.

External Project Page

Project Items (downloadables)

Name Category Sector Description Size Date Download

Project created: 2004-11-20