A person could design a chess playing engine that compares the location of all the pieces to a data set of all chess positions and the best move played in that position to make a decision on what move to make.
A person could design a chess playing engine that uses an evaluation function to decide what move to make.
There may be several other approaches like brute force iteration, but I selected these two to see if I am even close to understanding model-driven vs data-driven.
Is the difference between data-driven (the first option) and model-driven (the second one)?
Edit: In the first example, instead of compares as if iterative, it maybe would “memorize” them such that if the input to the neural net is the position information the network having been trained on all positions would output a move based on this training.
The AI has developed its function based on “training” of a massive data set. The evaluation function is a hypothesis the best function without any data.
In these attempts to unify cross-domain consilience people are failing to recognize how strongly it disciplines the sociology of science. Although, in my more cynical moments, I strongly suspect that they do recognize it – at least subconsciously – and that’s why it never goes anywhere in academia. It isn’t just that they’ll be caught making mistakes. It’s that cross-domain consilience is the best tool we have for quantiatively reifying bias. Institutional corruption is now well beyond merely avoiding embarrassment over incompetence. Never attribute to mere stupidity that which can be attributed to unenlightened self-interest.
I’ve specifically asked the Algorithmic Information Theory community to address this at the most primitive level of raw data from measurement instruments of physical quantities. As a thought-example: Take a cross-disciplinary dataset of physical measurements that has temperatures from a wide variety of sources of various phenomena – and come up with Algorithmic Information Theoretic subtheory of bias that would reify latent identities of thermometer manufacturers and how their thermometers are biased so as to provide greater cross-disciplinary consilience (hence smaller executable archive of the data).
This is so incredibly important in the present context of large language models that I’m simply dumbfounded that the AIT folks are dumbfounded at my request. The ball is obviously sitting in their court and its a big ball… and its inflating… the pressure is building…
Maybe a better way of understanding Zenil’s video is the old George Boxtenet:
All models are wrong, but some are useful.
The implicit assumption in Zenil’s diagram is that on the left, the models are less wrong (but maybe not as useful), and on the right is models are more wrong (but still often useful):
This 2017 interview on artificial intelligence is aging well – particularly the part where I mention how those of us with limited data (unlike the social network monopolies) must wake up to the importance of data efficiency before they do.
