GlyphMar 1
the AI alignment problem is entirely a smokescreen designed to distract from the capital class alignment problem
Show threadmcc
@glyph I do think there is an interesting perspective where computer software based on deterministic execution of instructions *can* be aligned with the goals of a user but computer software based on a trained statistical model cannot, technically, be aligned with anything at all as there is inherently random behavior. But we can't conceptualize that problem because the capital class is lying and saying that their computer has a soul because they named it "Cylde" and drew googly eyes on it
Mar 1 at 7:42pmWeb
Show threadmccMar 1
@glyph Even without the "Clyde" problem it's hard to talk about because there's a historical notion of a probabilistic algorithm where you have stochastic behavior operating with proven bounds and a provable distribution of behaviors, and the new type of statistics-based software where the software just sort of does whatever and we don't even discuss it as if it were statistics-based we call it "intelligence"
Show threadGlyphMar 1
@mcc no disagreement with any of that, but the “AI alignment problem” is specified by its advocates in terms of “universal human values”. the stipulated “alignment” is not with specific user desires or a stated optimization objective but with those putative (imagined) values
Show threadGlyphMar 1
@mcc the first problem of course is that it ignores society and culture and difference and the entire concept of politics[1], but the second issue that I am highlighting here is that *to the extent* that there are sufficiently popular values that we might call them “universal” and “human”, and *to the extent* that we have an entity that actually poses a threat to those values, it is the capital class.
Show threadGlyphMar 1
@mcc [1]: inb4 somebody says they actually wrestle with those things at extremely exhaustive length: they mostly try to rationalize those things away, which is not the same process
Show threadRandom GeekMar 1
@glyph @mcc resting safe in the assumption that anyone who claims adherence to universal human values hasn't so much as listened to Bruce's Philosopher's Song, and certainly not followed up on the associated readings.
Show threadJeff Miller (orange hatband)Mar 1
@glyph the first thing we'll do, is fire all the (actual) ethicists.
Show threadMatt BoydMar 1
@glyph @mcc At the far end of this the rationalists going "Logically we need to feed every poor person into a wood chipper so humanity can get to Mars."
Show threadGlyphMar 1
@3psboyd @mcc I feel a *little* bad for the lesswrongers generally because this is really judging the community by its worst and most extreme elements, and here we are on fedi (not a group whose most extreme and unpleasant members I would like to represent me) but that faction is certainly … unduly powerful in society right now
Show threadJayMar 2
@glyph @3psboyd @mcc
This. I know some decent ones.
But the decent ones tend to follow the Bentham-Utilitarianism-on-acid (aka longtermist) nutters, wherever they lead, IME.
Show threadGlyphMar 2
@jaystephens @3psboyd @mcc if they were at least real Benthamites they’d get out the felicific calculus and do the damn arithmetic and not just slosh around a bunch of half-assed Fermi estimates with orders of magnitude instead of numbers
Show threadGlyphMar 2
@jaystephens @3psboyd @mcc consider this my “born in the dark” Bane speech
Show threadDavid NashMar 2

@glyph @jaystephens @3psboyd @mcc

I know what “felicific calculus” refers to, but every time I see that phrase, I’m annoyed that it refers to generic happiness and not to the number of cats people have (or that they would like to have).

Show threadJayMar 2
@glyph @3psboyd @mcc
Mate, what's a bit of child labour in Africa compared to the happiness of the quadrillions of humans who'll flourish once we're spread across the galaxy? Any malnutrition and lost limbs in the here and now is a rounding error.
Show threadWillMar 1
@mcc @glyph LLMs are an epsilon-approximation to an intelligent autonomous system, where epsilon is equal to infinity.
Show threadTzimisce FleshMar 1
@mcc @glyph I mean, you can train a software on a statistical model to be really good at fulfilling certain goals. It's just Clyde was trained to convince people it has a soul rather than an actual statistics problem.
Show threadɗ𐐩ʃƕρʋMar 1
@mcc @glyph I think the biases in a random process (or more generally, the particular distribution) can still align with somebody else's biases and/or expectations. People have this thing where when you say "random", they immediately imagine some kind of fair lottery, with every option equally probable.
Show threadTravis F WMar 1
@deshipu @mcc @glyph yeah the flat distribution is commonly considered random, but really no distribution isn't an idealized model, even when biased. randomness, as statisticians like to talk about it, does not even exist.
Show threadɗ𐐩ʃƕρʋMar 1
@travisfw @mcc @glyph are you saying bayesians are not statisticians?
Show threadTravis F WMar 1
@deshipu @mcc @glyph them's fightin' words
Show threadDavid GerardMar 1
@deshipu @travisfw @mcc @glyph there's people who apply Bayes' theorem and then there's *Bayesians*
Show threadFlipper 🐬🏳️‍🌈Mar 1

@davidgerard @deshipu @travisfw @mcc @glyph True Bayesians are always ready to work the crank.

Bayesians are also much hotter than frequentists. The word "bae" is just the diminutive of "Bayesian", true story.

Show threadmccMar 2
@flipper @davidgerard @deshipu @travisfw @glyph i (a frequentist) once dated a Bayesian for a while. Nothing was learned from this experience which applies to other situations
Show threadDavid GerardMar 2
@flipper @deshipu @travisfw @mcc @glyph "Bayesian" is a contraction of "Bay Area sex pest"
Show threadstilescrisisMar 1
@mcc @glyph I don't think alignment has anything to do with determinism. People are non-deterministic but a person can absolutely be ethnically aligned (or not).
Show threadmccMar 1
@stilescrisis @glyph I think a certain sort of predictability is a prerequisite for alignment. Necessary but not sufficient. Humans are not deterministic but their behavior can be consistent, because they can act with intent. They can have beliefs and moral codes. They can understand their own incentives and the consequences of their actions. You can do things that cause them to understand the consequences of their actions better.
Show threadstilescrisisMar 1
@mcc @glyph Right, which is why they are called "model weights" and not "model coin flips." Models are non-deterministic at the token level but pretty darn consistent at the macro level, which is why ChatGPT articles are so easy to spot. "It's not X, it's Y"; numbered lists; boldface, etc.
Show threadmccMar 1

@stilescrisis @glyph "Models are non-deterministic at the token level but pretty darn consistent at the macro level"

At recreating the structural properties of language, yeah, because that's what the algorithm's for. But the product is not sold as a "structural properties of text simulator". It is sold as an engine for producing meaning. And when it comes to meaning the tokens matter very much, very very much