I think I have a slightly better understanding of what
Heidegger is trying to say. The following perspective is still
extremely wrong, because if you freely take this perspective, you
can apply it to so many different places where Heidegger in
Being and Time is not applying it to. But we
can try.
To a first approximation, a lot of “smart” algorithms (I will focus on algorithms for this one) at the end of the day, make a ton of assumptions about the world.
For example, many GOFAI algorithms (or even just a lot of normal, non-ML based algorithms) assume something like, “I will have some policies over some space, feed it data I understand, and spit out something nice”.
The magic of deep learning is that it can bypass all of this, and simply feed it the base data - and you get something that learns. But how does it learn, what is its internal representations, and can we extract out some kind of essence of it?
I think Heidegger (or at least people adjacent here) would just leave it at that as a universal negative result, but from an engineering perspective the issue is quite real. I mean we have terms for adjacent things like incommensurability and theory revision.
Heidegger doesn’t want you to say, “Well, things always rely on assumptions, and they bottom out in some unjustifiable too hard to compute thing”. That sort of implies that all you need to do is just feed something every possible reason ever and you will be A-OK. I mean, I guess AI models are trained sort of like that, but their intelligence clearly outstrips their training data.
A key analogy that I have now, that may or may not be incorrect (this is the same approach I had with Brandom/Sellars; anchored onto their “space of reasons” as a primitive meme even when I didn’t understand the whole argument) is something like this: The frame of what “matters” is not ever given a priori nor given as part of the agent itself or anything - it’s continuously given by the enviornment, and it isn’t some kind of matching relation where the agent receives some “objective” world (or even just a world in general) and tries to match it to some best course of action or representation.
That’s kind of bad, but I think it sits in the middle ground between the two axes of:
- “algorithms that rely on ontology fixing” vs “deep learning slop”
- a balance between agent & enviornment in who supplies the frame
and importantly, very importantly, no image/representationalism matching nonsense. Given that we know how much ontology can shape further discourse, don’t treat it as some kind of base, nor some kind of validation/verification layer, try to analyze how it actually changes.
As to why examine this through Heidegger rather than cogsci or something? Not sure, but this is my current understanding of him.
Engineering loops
Incidentally, this has got me to think about just engineering loops in general. Kind of a side tangent as it’s a “simpler” problem than the problem than Heidegger is trying to solve (as in, we’re more concrete and actionable here :D), but I think it’s important.
The issue with “base” things such as ontology and definitions is that they can and will constrain the space of things to talk about - no such thing as neutrality after all.
So how do you actually do stuff productively like:
- solidify noise in active conversation when possibly no shared model has been developed
- after said conversations / brainstorm in private, cooperatively upload to a knowledge database, while respecting that definitions and conceptions are contextual and will shift over time
Hard. Obviously it gets easier as the problem gets more and more “canoincal” over time but this is what ambiguity’s about.