C-LARA

The last couple of days, the main focus of our C-LARA activities has been concentrated on getting things installed on the new UniSA server. When we have this working, C-LARA should be a great deal faster than it is on Heroku, since we won’t have to depend on the slow Amazon S3 file storage.

The UniSA machine runs Red Hat Linux; I have never used Red Hat at all, so I’ve been largely dependent on Chat. This has given me plenty of opportunities to reflect on those confident pronouncements we see from the AI sceptics to the effect that the AIs may be able to do things but “don’t understand what they’re doing”. Alas, I can assure you that this poor carbon-based intelligence rarely understands what he’s doing as he installs packages under the AI’s direction. When we started, I didn’t even know the difference between ‘dnf’ and ‘yum’. Most of the time, Chat just gives me another command to execute, and I obediently paste it into the command-line interface and hit return.

In general, installing complex software is an interesting test case for the whole concept of “understanding what you are doing”. A large application like Postgres (the database) or Nginx (the webserver) represents at least person-centuries and quite likely person-millennia of effort. It’s doubtful that anyone “really understands” more than a fraction of the whole thing. But you don’t need to. A typical user, like me, doesn’t even want to. It’s enough to be able to move it from the package repository to the local machine and switch it on; then it will silently do its job behind the scenes, while I get on with the things that interest me. If something goes wrong, more understanding may come in handy, but often it will be enough to remember that turning things off and on again can fix them.

As has frequently been observed, members of a complex technological society are usually in this position. If we know enough to be able to use a piece of technology that someone else has created, we’re happy. The understanding involved may be extremely superficial, but we don’t have time for more than that, and there’s no need. When we talk about an AI’s ability to understand or not understand, we should bear these cases in mind: no one “really” understands, and the AI’s practical understanding is in general better than the human’s.