Over the last seven years, I've been heavily involved in a medical diagnostic project which unites human and machine judgement. This has always been cybernetic in my mind (and it was cybernetic insights which led to some pretty cool machine learning that sits behind it). It's about to be commercialised which is very exciting, not least because the technology is applicable to fields far beyond medical diagnostics - education, management, organisational risk and public health are all within scope of potential application.
Cybernetics relies on simple rules and metaphors, but these work in a wide range of contexts. The Law of Requisite Variety is the most important - the amount of variety (or complexity) that a controller has is the limit of the complexity of any system that it can control. Most simply, variety eats variety. Since most systems have to survive in environments of greater complexity than they possess, they must establish a controlled relationship with their environment through attenuation (selecting what information to pay attention to and what to ignore) and amplification (use their capabilities and understanding to create a niche in the environment - for example, a spider spinning a web). This can balance the variety equation.
A simple mechanical metaphor of cybernetics is the Watt Governor on a steam engine. The engine's speed, represented by the spinning of its flywheel, is controlled by a device (the governor) which uses centripetal force generated by the speed of the wheel to either slow down or speed up the flow of steam to the engine. This works because the wheel has exactly the same amount of variety as the governor: whatever state the engine is in is matched by a corresponding state of the governor.
This is fine as a metaphor, but in social life, there is no one-to-one mapping of environmental complexity to controllers, so we end up with very complex patterns of attenuation and amplification which can create dangerous positive feedback to the system. We are living through this in many ways at the moment - not just in the climate crisis, but in the political feedback from our online communication, the economic system producing runaway inequality, the Ukraine war, and so on.
Buckminster Fuller drew attention to a different kind of cybernetic feedback mechanism - the trimtab. Trimtabs are the small edges on the back of wings which wiggle as the plane is flying, and which serve to make the pilot's job of steering and stabilising the plane easier. In other words, the trimtab is part of a mechanism which connects the pilot to the machine. It is not self-enclosed like the Watt Governor, but translates the environmental conditions into a potentially controllable situation, which would otherwise be very difficult to control.
Buckminster Fuller thought so much of trimtabs that he had "Call me trimtab" written on his grave. He argued that the most important part of steering was not at the front, but at the back, and that each of us could be part of a "social trimtab" each feeding information about environmental conditions in a way which could facilitate effective steering.
The diagnostic AI which I and our team have created basically works like this. With our work, the "pilot" is the doctor, but the pilot's job is to steer through different environmental conditions in terms of differing degrees of prevalence, diagnostic certainty, organisational complexity, health economics, risk and potential positive feedback. To achieve this has entailed a very different approach to AI. Conventional AI is simply used to provide "answers", often with the intention of replacing the "pilot". That's not a good idea because it throws away huge amounts of information which can be critical to understanding the nature of the challenges we face. The trimtab (and our trimtab AI) by contrast preserves information, transforming complex data into the conditions wherein effective decisions can be made.
I've always felt that the most important thing education should do is to harness the uncertainty of individuals, because this information is information about the nature of our environment. What I've never been entirely clear about is how this "harnessing" looks - lots of forums, debate, etc, don't seem to work and in fact amplify social complexity. So we need a way of organising the many different signals coming from society as a means of facilitating effective steering for the planet (or Spaceship Earth as Fuller said). This may be the most powerful and effective use of AI.
No comments:
Post a Comment