Moving an arm and thinking about moving an arm are two vastly different things. Even thinking about thinking about moving an arm is a natural thing to do even if reading it is very odd. Now the hard part, how to you design thinking? The deliberate process of simulating scenarios to either logical or illogical ends would seem like a great fit for computers that can do millions of calculations a second. The slow an deliberate winding down a thought path seems to be the missing link to truely intelligent machines.
We are very good at making machines. Even more so machines that actually produce things and have purpose. Since the earliest primitive forms of man, our tools are defined by their use. Or more aptly by the end result they achieve. So what is the end result of smart machines? To drive our cars, build our structures, do the heavy lifting and manage our lives? To do all that the ability to compute and apply action is needed but not actual thought. Most business tools and modern applications of machine learning do exactly this. When a social media platform serves you an ad it’s collating data and balancing metrics about engagement, retention, ROI, etc. the list goes on. But there is no actual thought taking place. A singular layer of analysis and action, spread across a miasma of data and computing to be sure, but still only an un-inspired un-inventive static piece of functional 1s & 0s. Machine learning currently is the process where a rudimentary neural network learns the weights between parameters to approximate an end result. I could easily program a function to add to numbers together. In order for a machine to learn to add numbers together I would need a bank of data and the underlying pattern would be that the first two numbers added together equal the last number. This seems like over kill for the example of addition but what about recognising faces? Or interpreting expression? Identifying a child that has carelessly stepped onto the road Infront of 2 tonnes of accelerating metal? How do you program for that?
Gpt3 is one of the great hallmarks of machine learning to date. With a whopping 175 billion parameters the language and text generator is a marvel. Compared to the 85 (or so) billion neurones in the brain it seems like an absolute powerhouse. But it only produces a facsimile of one specific tiny subset of basic brain functionality. It could be argued that communication & language are the very basis from which higher level thought can spring so we’re on the right path. To have truely intelligent machines the ability to think and recurse that process on itself enabling thinking about thoughts seems an essential foundation. Unlike the human mind that is bound by the conservation of energy the artificial mind would have no limit and that could very well lead to an infinite loop about thought on thoughts long after it is useful. The slow an deliberate contemplation of an idea must have a logical end point even when the subject of the idea is itself an illogical exploration. The real trick is how to design that process or even the conditions to generate that process.
Very nice write-up. I definitely appreciate this site. Thanks!