Hacker News Reader

Comment by gibsonf1 | Hacker News Reader

The problem with this is that relational databases are completely foreign to how humans store and retrieve information. For humans our data is all connected and we operate with it primarily with hierarchical inference and reasoning. The relational database separates the data into tables and is terrible at hierarchical queries, so a 20th century idea that should be deprecated when working toward intelligence.

View on HN

There's nothing wrong with relational databases. There's something deeply wrong with how developers build relational databases. You can build databases in a way that meshes very cleanly with how humans think; people just don't. When a database developer makes an Asset table and a Building table and a Vehicle table, he's explicitly saying that Assets are definitely not Buildings and Vehicles are definitely not Assets and Buildings are definitely not Vehicles. That's the part that's completely foreign to how humans think: in reality, things are kinda like other things. No, Table Inheritance does not solve this problem (does it solve the problem in an OO language? No!). Again: you could build relational databases that mesh with how humans think, you just don't.

Also underlying your comment is this idea:

> Anything foreign to how humans think cannot help achieve intelligence.

Or:

> Human intelligence is the only possible kind of intelligence.

Bullshit. You could just have easily written off fixed-wing aircraft by saying they're completely foreign to how birds fly. It's silly to think that our particular form of intelligence is the only possible form of intelligence.

If you can never assume that any 2 concepts are distinct from each other, how do you model your database? I'd be curious how you would model the example you stated as a relational database that mimics how humans think

Relational databases are an implementation of the relational model of data, which itself is an implementation of first order predicate logic, which of course was invented by humans. If AI can figure out human language, it should be able to handle human logic as well. We’re already seeing that with AI code generators and co-pilots anyway.

There are lots of things invented by humans that don't model how human intelligence works.

Interesting observation, @gibsonf1. What kind of hierarchical queries do you have in mind that would be challenging to write in SQL? Would you mind sharing them in natural language?

Documentation databases operate in tree structures

- MongoDB

- ZODB

- Others

- This is also how disk file systems operate

You access items by their path, instead of id. Also path is relevant how the data is physically stored, fetching nearby data being cheaper.

To work with this kind of data in SQL you need to construct recursive CTE queries. While it’s possible, it is a bit akward sometimes.

jsonb queries in postgres have been very performant for us over +tb data sets. however, we often materialize views of large json blobs, just cause we can.

Our conceptual awareness AI uses recursive conceptual inference where given some entity, looking at its states, the system determines which concepts can be inferred directly from those states, then precedes to then also infer that that entity inherits all the properties of the genera concepts of those starting conceptual types. If you follow the recursion, it can be quite a large number of jumps, but incredibly fast.

For our data, we use Apache Solr, but a schema modeled on how we think the human schema works - its quite simple actually.

Reply on news.ycombinator.com