trading posts

[totient]

Scattered across the American southwest are trading posts. Not the Hudson's Bay company fur trading posts that mostly closed in the 1840s; these are trading posts founded to trade with the Navajo and other tribes in the 70 years after the Long Walk. At their peak, just before World War II, there were nearly 150 of them, of which dozens remain in operation -- I don't know exactly how many as there is no comprehensive list, nor a category in Wikipedia. I've been to several: Hubbell, Cameron, Gouldings, Ismay, Twin Rocks, McGee. They're all different, sometimes drastically so: Cameron's Trading Post is absolutely enormous with tour buses pulling in to it every few minutes, whereas there wasn't much sign that anyone other than myself went into Ismay's Trading Post at all on the day I was there, towards the very end of its 90+ year existence. Hubbell is now a museum; McGee is mostly an art gallery. Each has taken its own path.

But for all their differences, they have some things in common, so much so that when researching this post, I read yelp reviews stating "that's not a trading post" and nodded, because the establishment in question didn't fit the pattern. I'm not sure I can exactly put my finger on what the rule really is. In my experience generally they're right on a political boundary, right next to at least one stunning landmark, and buy things as well as selling them. But I don't think this is what those yelp reviewers were talking about and there are exceptions to these and any other rule you might come up with.

So it is, I think, with the sorts of science fiction conventions I like to work on. DragonCon and WesterCon are as different as Cameron and Ismay. But there is a kernel of sameness within them anyway. And there's no need to define that essence, nor any one rule that exactly describes what that kernel is.

Comments

[foms]

I've thought about what it means to prefer a given kernel and to want or try to keep a gate. I've also thought about punching up and punching down and about the history of the ghetto and about cultural imperialism.

I'm amusedly reminded of the first time that I experienced someone referring to Ebay as The Bay (company).

[rmd]

> But there is a kernel of sameness within them anyway.

Very much. And very much a "I know it when I see it" sort of thing.

[jducoeur]

Back in college, I took an AI / Cognitive Science class that was enormously formative in my thinking. And the most formative session was the one where the professor sat us all down and said that today's challenge was simple: define "chair". The entire class spent the better part of an hour at it, and could not come up with a rule set that he couldn't provide counter-examples for.

Human cognition is all about pattern-matching, and those patterns are vague. We tend to agree on them, and they often aren't even controversial, but as soon as you try to describe them in terms of rules, you're probably in trouble.

The outcome of that class was me deciding that most of AI as it was defined back then -- very rule-and-category-centric -- was clearly nonsense, and that multi-layered neural networks with vast amounts of data were the only way to go if you wanted "AI" worthy of the name. The thirty years since have been pretty gratifying, although we still have a fair ways to go.

[totient]

OMG the state of AI in the late 80s and early 90s. I remember taking a natural language processing class that was all about symbolic reasoning in Lisp. Ahaha. Ten years later I recycled the textbook because everything in it had been completely discredited.

[totient]

I've just come back to this, because current generative AI is using the definition "the set of things for which the locus of objects previously identified as chair are collectively a closer match than any other locus".

Which is still wrong, because there are also negative rules. We have objects that have previously been identified as not chairs, and those don't form a locus so their relevance has to be considered in some other way.

[rmd]

Generative AI saying "Idunno, but here's what everyone else seems to think is a chair?"

[totient]

Ultimately I think humans are also saying "I dunno, but here's what everyone else seems to think is a chair". But doing so more accurately, because they're more correctly modeling what people don't think are chairs.

I suspect that there's a fundamental difference in how humans learn from positive vs negative feedback, and in particular that there's a more likely to be a symbolic step in negative feedback. When we see an exception to a rule we look at the rest of the data set to find the basis vector on which the exception is operating in a way that I suspect we don't when the data point expands but doesn't conflict with the current model.

AI could be doing this, but I don't think it is.

[jducoeur]

Right -- LLMs aren't even really asking the question "what is a chair?", but instead, "what are the words most likely to follow chair?" (more or less). That produces sensible results a surprisingly large fraction of the time, but is a somewhat separate question from "how do humans think about the idea of chair?"