The AI Research Group of the Centre for Digital Culture has released a second publication in their series examining Artificial Intelligence and Catholic social teaching, ethics, and moral theology. The first volume - Encountering Artificial Intelligence was good and I’m very much looking forward to starting this one, which is available for free as a PDF: Reclaiming Human Agency in the Age of Artificial Intelligence.

So here is a rare post that refers obliquely to my professional life.

If AI seems like it’s basically everywhere now I can assure you that it’s even more pervasive for people (like me) who work in the technology field. I’ve noted here before some of my concerns regarding LLMs and their inability to weigh, categorize, or hedge their output. All output, regardless of provenance, is equally plausible to them and is presented to the user accordingly, though I’ve noticed more language-of-uncertainty creeping into Claude’s output recently.

I was introduced to the concept of mechanistic interpretability the other day. This is a research field that examines the massive neural nets that make LLMs tick and aims to figure out how and why they do the things they do. It’s bonkers to me, still, that we’ve created something that’s too complicated for us to easily understand and whose inner workings remain a substantial mystery to us. This is, I understand, a necessary consequence of building something that seeks to work more like a brain than a simple expert rules-system. The former is a thing that can be trained to do a task from scratch, where the latter is built to do one thing extremely well, such as play chess. The chess-playing bots can beat basically anyone now, but they can’t do anything else with this ability. We can examine the chess bot code and know why it made a particular move in a particular situation; we can’t easily know why an LLM produced that output in response to this input. This makes a sort of sense; similar insight eludes are own self-understanding, so maybe this is as good as it gets. As the models get bigger/better/faster, their inner workings become even more opaque. And then what, exactly?

To be clear - I believe these things are useful. I’ve seen what they’re capable of, even on a tiny scale, for the various things I use them for. I do worry a bit about the speed of things and the wholesale enthusiasm for it all. I’m trying very hard to not sound like a contrarian antiquarian here. Part of is the memory of the dot-com hype-cycles of the mid-90s and part of it is knowing how the technology sausage is made. In any case I’m grateful for this working group’s stuff. It’s important and worth the time to read, especially for those making technical decisions, policies, and so on.