down among the unclefts

Claude: Human name

GPT: Robot name

Gemini: Mythological name

So rather than an amateurish marketing decision, it's a decision to just keep using the name that was originally meant for an audience of fellow ML researchers.

The most likely reason for that non-decision is that OpenAI did not intentionally decide to pivot to mainstream. They'd been doing experiments for months seeing if there was any significant commercial use case for GPT-3 variants with lukewarm success, and suddenly their experimental interface "ChatGPT" became the fastest-growing product in history. After that, GPT was the name the public knew and it was hard to change.

You're on a large rotating disk. You have a ruler, and you use it to measure the diameter and circumference. Your ruler shortens from Lorentz contraction when you measure the circumference, but not when you measure the diameter. So the relation between your measurements is not the one given by Euclidean geometry, that the circumference is π times the diameter.

(By the way, it seems that Einstein first published this thought experiment in 1912, according to Pais's Subtle is the Lord, chapter "The Prague Papers", so it's not just exposition but may have played some role in the development of general relativity)

This seemed to me like it shows a particular use of thought experiments, to show the incompatibility of the conclusions drawn from two different principles. It's particularly interesting because until partway through the 19th century, people did not seem to clearly make the distinction between Euclidean geometry being a description of the world or of geometric intuition.

Nozick's paper on Newcomb's problem seems to use it in the same way. The title is "Newcomb's Problem and Two Principles of Choice". The "two principles" are expected value and dominance. He tells us that one-boxing maximizes expected value, but two-boxing is dominant.

It seems very... philosophical, though. Like, you're in this situation, what would you do.

Or maybe... behaviorist? I've been reading Savage and it seems connected to the prominence of behaviorism at the time. His style of subjective Bayesianism with its tradition of "eliciting" probabilities and utilities, trying to fit them to someone's expressed preferences. I guess the philosophical tradition of testing principles against thought experiments is similar in a way.

But... I don't know, if I'm thinking through problems like this, sitting at my desk, maybe drawing some diagrams or doing some calculations, I guess I want to be doing it based on some model of how the situation I'm in works? Maybe here I'm just arguing against some Tumblr Newcomb's problem discourse from years ago:

You don’t know the mechanism, you have to guess; whatever he’s using to predict people, he can use it now, on you, but you don’t know what it is. Deduce and decide.

Is just "deciding" what I want here? What I want is to reason from assumptions about the situation to their consequences, in this case for hypothetical actions I could take. Which I guess brings me back around to whether this thought experiment is playing a role analogous to Einstein's thought experiment. I guess I feel like it should be "mechanistic" like a physics thought experiment if I'm going to have an answer at all. Not because every situation I'll face in life is mechanistic, but... more like, that's what I think this kind of pencil and paper reasoning is good for, and what I want from the exercise is to advance the art of pencil and paper.