I’ve been much obsessed by related questions lately. What are information, memory, and computation? What are learning and intelligence?
One can answer these questions in either direction, starting with information and going forward to intelligence, or starting with intelligence and going backwards to information. I typically find it easier to explain from the bottom up, so I’ll start with information.
Cesar Hidalgo, Director of MIT’s Media Lab and physicist, defines information as physical order. When energy flows through systems of matter, they have the ability to self-organize.
We can use the image of a whirlpool in bathtub to help us understand this. Water sitting in a tub is matter with little energy flowing through it. When we unplug a stopper, energy flows out of the tub and the water spontaneously organizes into a whirlpool. This physical order is information, so the whirlpool is information.
The problem with fluids is that they can’t store information. As soon as the energy stops flowing through the tub, the whirlpool dissipates and there is no memory of it. This brings us to the question of what is memory? It is information that is captured in a solid state. If we were able to freeze the whirlpool into ice, that would work as a memory device. So memory is just the crystallization of information in solid states.
Once we have information stored, we can begin using it to compute. Computation is the transformation of inputs into new output via a function. As Max Tegmark, another MIT physicist puts it, computation is the transformation of one memory state into another memory state.
Tegmark continues by explaining that if computation is all about the transformation of information from one memory state into another, then learning is simply the ability to improve the function used in computation. Learning is simply getting better at computation.
All of these allow us to understand what intelligence actually is. Intelligence is the ability to accomplish complex goals. It is the ability to use information, memory, computation, and learning in order to arrive at desired outcomes. This has two consequences; intelligence is value neutral and it makes little sense to quantify it in a single metric like IQ.
The first consequence is self-evident. We can select goals that are extremely benevolent or extraordinarily evil. World peace would be a benevolent goal. The “Final Solution” would be an evil goal. Both require massive intelligence in order to pull off as they are incredibly complex goals.
The second consequence is evident once we begin thinking of intelligence beyond the individual as the focal point. It makes little sense to think of intelligence as just an individual trait anymore. The ability to compute information to accomplish goals is better done in networks once the goal becomes complex enough. This is why two cognitive scientists, Steven Sloman and Philip Fernbach, believe intelligence is better measured at the collective level.
Instead of measuring g, or general intelligence in an individual, Sloman and Fernbach have advocated that c, or collective intelligence, is a better method. This measure can still be applied to an individual in a team, but it attempts to measure how difficult it is to replace a person within a team. If a person contributes highly to a group in accomplishing a goal, regardless of their g, then they would have a high c. Cognitive scientists have in fact generated a valid measure of c and Sloman and Fernbach believe it could become more useful in general.
The more general problem with intelligence, whether measured in humans with g or c, is that that there is nothing preventing machines from doing it better over time. Tegmark asks readers to think of a landscape of human competence, with lowlands, foothills, and mountains. Computers have been slowly advancing and their performance is like water flooding the landscape, forcing humans to retreat, first into the foothills, but eventually further and further up into the mountains.
As McAfee and Brynjolfsson point out in their newest book, Machine, Platform, Crowd, machines are already a better replacement for human minds on many things we value, such as judicial sentencing and mortgage and business loan processing. Machines do fail at some tasks, namely those that we haven’t bothered to model or figured out how to model with functions. However, machines mostly perform the same or better than humans across the board on tasks which they can model by computing inputs to desired outputs.
Those are my current understandings of information, memory, computation, learning, and intelligence. Information is physical order that arises from energy flowing through matter and can be stored in solid states called memory. This information can act as inputs to functions which compute them into outputs. Learning is the act of tweaking the function over time for improved outputs. Finally, whenever we are using our computational abilities to achieve specific outputs, we are acting intelligently, which often requires networks larger than the individual for complex goals.
What fascinates me about these questions is that they force us to very genuinely and openly ask ourselves what our collective aims and goals for society are. When machines can use intelligence to extract information from crowds using massive platforms to compute and learn, what do we hope and wish for those humans lost in the flood? What goals are worth using intelligence to accomplish? What information state should we be aiming at and will it be better for all or just a few?