Member-only story
Trivial Pursuit
Current AI models solve the knowledge bottleneck. They implicitly show that intelligence requires knowledge. While IQ can be measured as an abstract property on test questions, intelligence becomes applicable together with information or data. Together this creates knowledge. However, at the present time AI is indiscernible from a large set of trivia, random information taken from all sources a company could find. Why is this?
At the core of every AI model there is a large neural net, usually in the form of a deep learning net with many feature layers. There are mechanisms involved for learning from sequences — called attention — which allow to highlight tokens of information over large context windows. Finally there are many additional tools and much fine-tuning by human workers involved to create a more seamless experience, to implement guardrails and block certain topics, to improve on grammar and style, etc. When a system is so large that running it costs millions each day in electricity it is immediately apparent that human input for fine-tuning is incredibly cheap in comparison.
These databases are trained to memorize large sections of text and then to interpolate on new questions. The memorization must make mistakes for the interpolation to work. Otherwise, for complete memorization, the system has no capacity for interpolation. It is set up in such a way that a perfect retrieval…
