The mystery of knowledge

Easy to see once understood, yet so hard to capture, integrate and communicate. Our symbolic understanding of reality, local or universal, keeps escaping even our most refined techniques.

In recent years we’ve seen unprecedented investment in digital computing (CPUs, GPUs, TPUs, QPUs) and in organic computing (people creating, thinking, classifying and operating) to collect all digitally accessible human knowledge and aggregate it into some kind of all-knowing, commoditized oracle.

Even if the esoteric probability of trillions of parameters resembles the behavior of a thinking entity, in the practice of things that matter, knowledge and wit keep slipping past transformers and neural networks.

Two lineages

Since the 1990s, many institutions and minds have worked to structure how we organize knowledge and classify information spread across public and private networks, so that machines and people can interpret it. That effort gave us the ideas and standards of the semantic web for modeling knowledge: RDF, Linked Open Data, ontologies and knowledge graphs.

In parallel, over those same decades, in the operational context of companies, we saw the tooling for data analysis and integration evolve, usually grouped under the Analytics umbrella, with goals closer to monitoring processes and producing indicators: data warehouses, dimensional modeling, distributed processing frameworks, OLAP cubes.

Even though the goal is similar, “we need to understand something better”, these two worlds barely talked, perhaps because of the large differences in tooling, subject matter and processes in each domain.

Entropy alone isn’t enough

Then, out of the generative AI revolutions, came the ideas of vector embeddings and attention, with the promise of “extracting” or “deriving” knowledge from raw data, with no need to model concepts first.

A few years into this wave, we’re realizing that even with all the money and compute in the world, “just analytics” still can’t produce knowledge out of entropy. Something is still missing to sustain understanding, and in the latest rounds of the game that something looks like the combination of knowledge graphs and vector embeddings.

Thinking better

I can’t know what the future holds, but I take one reflection from all this. While many seem to believe these technologies come to “help us think less”, to me they seem to demand that we think better about how we represent concepts and how we communicate.

It’s contradictory, to say the least, to want quantum processors and crawlers that scrape the entire internet so that our children can produce copy and illustrations without having to think or learn.