Rubicon of Cognition
Over time we accumulate a better understanding of what is happening. And this happens mostly through the combination of three directions:
- New knowledge,
- deeper and more detailed grasp of already mastered knowledge,
- and the third, a new level of abstraction, that is, conclusions drawn from the awareness of everything, very often from several different segments of our knowledge at once, conclusions that overall lead us to these emergent insights.
It can also be pictured as expansion, deepening, and synthesis (which is precisely what I consider the new level of abstraction, i.e., discoveries).
In essence humanity already knows quite a lot. And when we as a civilization develop, we mostly draw on the accumulated knowledge and there is an enormous amount of it. Life experience, wisdom, practice certainly give depth, but that alone is still not enough for creating genuinely new knowledge. New knowledge becomes possible only through a new level, thanks to emergence and in no other way.
The real question of creating AGI is exactly this: teaching models to draw new conclusions by aggregating and analyzing already known information. So far it's proceeding slowly and laboriously, yet examples already exist, just very modest and insignificant ones. And for many people it remains an open question whether models will ever be able to achieve this.
Because this is the best way to understand whether there is some form of mind in there, or whether it's only algorithms. After all, consciousness and reason are exactly what is required for the synthesis of information, one could say for making scientific discoveries and rolling out new technologies.
And as soon as models learn to give explanations of accumulated knowledge at these new levels of abstraction, they will immediately become incomparably, by orders of magnitude, more efficient. Because then it will no longer be necessary to memorize gazilobytes of trivial facts; instead it will be possible to rely simply on general principles, logic, and a handful of examples.
Tags: aiphilosophycognition
