Comment by avgcorrection | Hacker News Reader

Comment by avgcorrection | Hacker News Reader

Parent Full thread

A perfectly mediocre essay.[1]

> Computers wipe the floor with us anywhere we can keep score

Notice the trick? If you can keep score at something then you can probably make an algorithm for it. If you can make an algorithm for it then you can probably make a digital computer do it a billion times faster than a person, since digital computers are so good at single-“mindedly” doing one thing at a time.

> So what’s being stripped away here? And how?

> The what is easy. It’s personhood.

Why?

The Turing Test was invented because the question “do machines think?” was “too meaningless” to warrant discussion.[1] The question “can a machine pose as a human”? is, on the other hand, well-defined. But notice that this says nothing about humans. Only our ability (or lack thereof) to recognize other humans through some medium like text. So does the test say anything about how humans are “just X” if it is ever “solved”? Not really.

You put a text through a blender and you get a bunch of “mediocre opinions” back. Ok, so? That isn’t even remotely impressive, and I think that these LLMs are in general impressive. But recycling opinions is not impressive.

> (though in general I think the favored “alignment” frames of the LessWrong community are not even wrong).

The pot meets the kettle?

[1] That I didn’t read all the way through because who has time for that.

[1] https://plato.stanford.edu/entries/turing-test/

> A perfectly mediocre essay.

The author rightly draws attention to text.

LLMs showed they can do the classical NLP tasks and more: summarise, translate, answer questions, play a role, brainstorm ideas, write code, execute a step by step procedure, the list is unbounded. It's the new programming language.

All these abilities emerged from a random init + text. Guess what was the important bit here? Text. It's not the architecture, we know many different architectures and they all learn, some better than others, but they all do. Text is the magic dust that turns a random init into a bingChat with overactive emotional activity.

Here I think the author made us a big service in emphasising the text corpus. We were lost into a-priori thinking like "it's just matrix multiplication", "it's just a probability distribution predictor over the next token". But we forgot the real hero.

The interesting thing about words is that they are perceptions, they represent a way to perceive the world. But they are also actions. Being both at the same time, perception and action, that makes for an interesting reinforcement learning setup, and one with huge training data. Maybe text is all you need, it is a special kind of data, it's our mind-data.

> Notice the trick? If you can keep score at something then you can probably make an algorithm for it

You are basically arguing P = NP, but it isn't known to be the case. As far as we can tell, keeping score is much easier in general than finding states that yield a high score.

I seriously doubt that “anything/[everything] we can score” has been conquered by AI,[1] but I was assuming that the author meant those typical AI milestones.

[1] What about some kind of competition where you have to react and act based on visual stimulus? And you have to do it perfectly?

Little too broad. If the act is tell you what color it is, then computers will win every time. Again, if you can score it. A computer will win.

Nice counter-example. Why would the test be that simple?

Then bring forth a complex but indivisible test?

Can’t be bothered.

Yeah, like sorting apples into good and bad piles really fast?

Sounds like a leading question which is supposed to debunk that whole category by suggesting one counter-example. So no.

Oh, was the point to only pick examples that made your point seem stronger?

The point was to allude to simple tasks for our sense-system which AI has issues with at present. Without getting lost in the weeds with y’all well-actuallying every little concrete example because you love to be contrarians.

Given that we don’t (seem to) have widespread beta-ready autonumous cars yet, I’m guessin that this is still the case…

I think you should try reinterpreting this, if there are a bunch of holes in your examples that are meant to be clear and defining, then maybe it isnt so open and closed?

Its also a little strange to be called a contrarian in this instance (yes generally speaking true) because it does not seem to be contrarian to point out obvious counter examples. It feels like a way for you to easily dismiss any other opinion because you already binned it as some one denying reality in order to pointlessly argue.

>A perfectly mediocre essay

One might even say "premium mediocre" [1]

[1] https://www.ribbonfarm.com/2017/08/17/the-premium-mediocre-l...

Reply on news.ycombinator.com