Meditations by Marcus Aurelius has been translated into English more than dozens of times in the modern age. There is great debate about which translator did the best job, with different answers depending on what aspects, such as accessibility and faithfulness, to consider. Now suppose an algorithmic translator (like Google Translate) based on state-of-the-art machine learning techniques, that can perform near-perfect Ancient Greek to modern English translations. We would get one, de-facto translation of the text. The variations, or “imperfections”, of the human translated works would be lost.

More generally, an argument can be made that as we increasingly rely on machine learning and similar technologies to operate, our world will become more homogenous. The family of techniques loosely termed as “machine learning”, or ML, is all about making new things (e.g. predictions) based on past information. The data in the past is unchanging, so as the algorithms for extracting information out of the past improve and proliferate (i.e. the commoditization of ML techniques), and as more people get access to the same data set to train on (i.e. the democratization of data), the output that everyone gets should converge on some fixed point.

In a way, people function the same way as an ML algorithm (after all we based the field in large part on how ourselves behave). Each of us has a set of algorithms running (that would be our personalities, the way we think, etc…), and our future behaviors are determined by how the parameters of these algorithms change based on our past interactions (which would be how we change based on life experiences). Yet there is much less convergence effect here. This is because, first, our minds are all different, and they are not easy to change, and second, each of us only has access to an extremely small and different subset of all experiences. We can be thought of as different ML algorithms, all running on some tiny random sliver of the total available training data. This will result in everyone being different, with high heterogeneity.

This thought experiment about an ML-driven world then, is in a way, asking the question of what happens if we took the best parts of everyone’s minds, and combined everyone’s experiences, and made a meta-human out of it. This reminds me of the joke that in the future, the whole world will be full of beige people, because with true globalization and racial mixing, that’s the average skin color of our descendants based on current population distributions. The difference of course is that the technology world moves much faster, and has much fewer barriers to progress. So are we, through the increasing usage of machine learning, building the “beige” mind? If the world, or even just how most of the world operates, truly converges on some fixed optimum, I wonder how we will feel. We might try to deliberately sabotage the “perfect” algorithm out of boredom, just to watch something different happen.