Software engineer, blogging and editing at /nickai/. PM me your fluid-g-increasing ideas.
NicholasKross
Karma: 1,345
I was texting multiple Discords for a certain type of mental “heuristic”, or “mental motion”, or “badass thing that Planecrash!Keltham or HPMOR!Harry would think to themselves, in order to guide their own thoughts into fruitful and creative and smart directions”. Someone commented that this could also be reframed as “language-model prompts to your own mind” or “language-model simulations of other people in your own mind”.
I’ve decided to clarify what I meant, and why even smart people could benefit from seemingly hokey tricks like this.
Heuristics/LMs-of-other-people are something like, a mental hack to trick our godshatter human brains into behaving smarter than we reasonably would have if left unprompted, due to our computational (and in particular our “recall” memory) limitations.
Like, yes, optimal Bayesian reasoners (plus a few Competent People who might exist, like Yudkowsky or Wentworth or mumblemumble) can do this unprompted, presumably because they have years of practice at the “mental motions” and/or better recall ability (combined, of course, with the rarity of having read any of the Sequences at all, to make a lot of this stuff explicit). Thus, the heuristics help those of us who don’t consciously remember every Bayes-approximating tip we’ve ever heard about a given mental situation.
Yep, we can easily have multiple hypotheses of the form “something we don’t (yet) understand has caused this”. My odds are more on “weather/camera/light/experimental aircraft we don’t understand” than “aliens we don’t understand”.
One problem with “using a simpler example”, is that there’s a lower bound. Prime numbers are not-too-hard to explain, at some levels of thoroughness.
Like, some part of my subconscious basically thinks (despite evidence to the contrary): “There is Easy Math and Hard Math. All intuitive explanations have been done only about Easy Math. Hard Math is literally impossible to explain if you don’t already understand it.”
Part of the point of Mathopedia, is to explicitly go after hard, advanced, graduate-level and research-level mathematics. To make them intelligible enough that someone can learn them just from browsing the site and maybe doing a few exercises.
Even if they need to go down a TVTropes-style rabbit-hole (still within the site) to find all the background knowledge they’re missing.
Even if we add increasingly-unrealistic constraints like “any non-mentally-disabled teen should be able to do this”.
Even if it requires laborious features like “there should be a toggle switch / separate page-subsection that replaces all the jargon in a page with [parentheses of (increasingly recursive (definitions)]], so the whole page is full of run-on sentences while also in-principle being explainable to an elementary schooler”.
Even if we have to use some incredibly hokey diagrams.
Pretty tangential, feel free to remove:
The YouTube “BarelySociable” did a 2-part video a while back, trying to figure out who Satoshi Nakamoto was. He gave pretty decent evidence it was a British guy who’s not any of the 3 candidates mentioned.
Yep! My main hope is that it works in a niche of people who needed specifically-it (or who find it more “intrinsically fun” to read and/or contribute to than the other options).
choose something that is difficult for others but simple for you.
Yep, a broader life lesson I’m still learning haha.
IIRC Paul Graham recommended such a tactic, framing it as “easier gains from moving around in problem-space than solution-space”.
And your other recommendations definitely make sense here. In my giant bookmarks folder about the “mathopedia” idea, this post and the comments are bookmarked.
This is “merely” one safety idea… but it’s a slam-dunk one that there’s not (as far as I can tell) good reason to not do.
Very good points, yeah!
I actually attempted making an example-page in a Wikipedia sandbox, but did not have the energy/deeper-requisite-knowledge for the topic I chose (Godels Incompleteness Theorems ;-;), so I didn’t finish it. But I do agree that, if I launched this, I’d need at least one good example-page.
Another part of the problem, which Arbital especially failed at, was getting others to contribute. Reddit and StackOverflow solve this by basically giving people literal “status points” for writing
helpfuleffort-signaling posts. So I’d want some kind of MediaWiki/other plugin that says “Hey, new contributor! Here’s a list of subsections of different articles, where we want X type of examples for concept Y. Mind adding more stuff there?” I even had a harebrained idea for a multi-authorship-explicit-credit plugin of some sort.
These ideas seem promising!
How do you distinguish feeling of epiphany and grokking?
Good point, I haven’t really done that here. We could differentiate by e.g. having practice-problems, and people can login to track their progress. Similar to the multi-explanations/teaching-methods setup, there could be a broad variety of example problems --> less likely someone gets lots of them right without actually understanding the concept.
Forgot to mention in-post, but the site BetterExplained is also quite good.
Dreams of “Mathopedia”
For this incentives-reason, I wish hardcore-technical-AI-alignment had a greater support-infrastructure for independent researchers and students. Otherwise, we’re often gonna be torn between “learning/working for something to get a job” and “learning AI alignment background knowledge with our spare time/energy”.
Technical AI alignment is one of the few important fields that you can’t quite major in, and whose closest-related jobs/majors make the problem worse.
As much as agency is nice, plenty of (useful!) academics out there don’t have the kind of agency/risk-taking-ability that technical alignment research currently demands as the price-of-entry. This will keep choking us off from talent. Many of the best ideas will come from sheltered absentminded types, and only the LTFF and a tiny number of other groups give (temporary) support to such people.
Final link seems broken.
Yeah, I keep finding myself wishing that every other message/communication platform I use, would add Discord-style custom emotes for hyperspecific situations.
I acknowledge this is phrased kinda weirdly. I would say this fits the spirit of the question (albeit as a noncentral example), plus “opting-into reacts as a whole” is required on a by-post basis.
Finally, a multi-karma (well, multi-vote) system!
Abstraction is Bigger than Natural Abstraction
Also possibly relevant (though less detailed): this table I made.
Also possibly relevant (though less detailed): this table I made.
(sources: discord chats on public servers) Why do I believe X?
What information do I already have, that could be relevant here?
What would have to be true such that X would be a good idea?
If I woke up tomorrow and found a textbook explaining how this problem was solved, what’s paragraph 1?
What is the process by which X was selected?