Posts
Comments
Improved my intuitions, ty.
Keeps baffling me how much easier having a concept for something makes thinking about it.
What about this one:
"Hivemind" is best characterized as a state of zero adversarial behavior.
"Humanity becomes a hivemind" is the single least dystopic coherent image of the future.
Illustrative post. The downvotes confuse me.
Depression is a formidable cognitive specialization.
There may have been other, unmentioned optimization targets that also need eloquence
Predictions:
- (75%) Groups who successfully[1] adopt trust technology will economically and politically outcompete the rest of their respective societies rather quickly (less than 10 years).
- The efficiency gains feasibly up for grabs in the first 15 years compared to statusquo are over 100% (75%) or over 400% (50%).
- (66%) Society-wide adoption of trustbuilding tech is a practical path / perhaps the only practical path towards sane politics in general and sane AI politics in particular.
The whole gestalt of why this is a huge affordance seems self-evident to me, it's a cognitive weakness of mine to often not know which parts of my thinking need more words written out loud to be legible.
But one intuition is: Regular "natural" human cultures are accidental products sampled from environments where deception-heavy strategies are dominant, and this imposes large deadweight costs on all pursuits of value, including economic value, happiness, friendship, and morality. Explicitly: Most of our cognition goes into deceiving others, and the density of useful acts could be multiple times higher.
- ^
i.e. build mutual understandings at least to, but ideally surpassing, the point of family-like intimacy / feeling the others as extensions of oneself
I'm not eloquent enough to express how important I think this is.
I feel like such intuitions could be developed. - I'm more uncertain where I would use this skill.
Though given how OOD it is there could be significant alpha up for grabs
(Q: Where would X-Ray vision for cluster structures in 5-dimensional space be extraordinarily useful?)
Hmm. Yeah. It gets difficult to display points with the same XY coordinates and different RGB coordinates
With colors you can in principle display data in 5-dimensional space on a 2D medium without flattening.
Bottlenecks (cognitive):
- intuitively knowing the RGB values of colors you're seeing
- intuitively perceiving color differences as 3-dimensional distances
Feasible? Useful?
Latest in Shit Claude Says:
Credibility Enhancing Displays (CREDs)
Ideas spread not through their inherent quality but through costly displays of commitment by believers. Words are cheap; actions that would be irrational if the belief were false are persuasive.Predictive angle: The spread of beliefs correlates more strongly with observable sacrifices made by believers than with evidence or argument quality.
Novel implication: Rationalists often fail to spread ideas despite strong arguments because they don't engage in sufficient credibility enhancing displays. Effective belief transmission requires demonstration through personal cost[1].
The easiest way for rats to do this more may be "retain nonchalant confidence when talking about things you're certain are true, even in the face of audience skepticism"
- ^
I think the "personal cost" angle is mistaken. Costly Signaling only requires the act would be costly if you didn't posses the trait.
Aspies certainly seem to do this less!
You mean, like him as a blogger? Or as a person in real life?
The latter? Like, I subconsciously parse his blogging voice not unlike as if it were a person in my tribal surroundings, and I like/admire/relate to that virtual person, and I think this is what causes some aspect of persuasion
I mean yes it's embarrassing, but it's what I see in myself and what seems to be most consistent with what everyone else is doing, certainly more consistent than what they claim they're doing.
E.g. it seems rare for someone who actively dis-appreciates the sequences to not also dislike Eliezer for what seems like vibes-based reasons more than content-based reasons
But then again, all models are false!
If I peer into my own past, where arguably I was more autistic than today, I can see that my standards for admiration seem to have been much stricter. I basically wouldn't ever copy role models because there were no role models to copy. This may be the shape of an important caveat
They do, but the explanation proposed here matches everything I know most exactly and simply.
E.g. it became immediately clear that the sequences wouldn't work nearly as well for me if I didn't like Eliezer.
Or the way fashion models are of course not selected for attractiveness but for more mimetic-copying-inducing highstatus traits like height/confidence/presence/authenticity
and others
And yeah not all of the Claude examples are good, I hadn't cherrypicked
More thoughts that may or may not be directly relevant
- What's missing from my definition is that deception happens solely via "stepping in front of the camera", i.e. via the regular sensory channels of the deceived optimizer, ie brainwashing or directly modifying memory is not deception
- From this follows to deceive is to either cause a false pattern recognition or to prevent a correct one, and for this you indeed need familiarity with the victim's perceptual categories
I'd like to say more re: hostile telepaths or other deception frameworks but am unsure what your working models are
I'd say weirdness is about not being predictable
Perhaps along some generalized conformity axis - being perceived as a potential risk to the social order.
Deception: An optimizer falsifies another optimizer's models in order to steer its behavior
Had a minor braincoom discovering Mimetic Theory
Best model/compression I took away is a mental image evoked by "Desire is triangular, not linear" depicting how desires are created via copying
Claude 3.7 explains some basics:
Desire is triangular, not linear - We don't want things directly; we want what others want. Every desire has a hidden "model" we're unconsciously imitating.
Conversion happens through the model - We convert to a new worldview by imitating someone we admire, not through intellectual persuasion. Reason follows mimetic conversion.
The interdividual self - Girard rejects the autonomous individual entirely. The "self" is actually a collection of desires borrowed from others. What we call "personality" is just the unique pattern of our imitations.
Common Examples:
- Kids fighting over the same toy while ignoring identical ones
- Fashion trends spreading through social groups
- Career paths chosen because respected peers chose them
- Romantic triangles where someone becomes attractive once they're dating someone else
- Consumer frenzies (iPhones, limited editions) driven by visible queues and scarcity
- Gentrification patterns where neighborhoods become desirable because the "right people" moved there
- Academic research clusters forming around suddenly "hot" topics
Subtler Manifestations:
- The desire for "authenticity" itself (ironic since it's mimetically transmitted)
- Self-improvement goals based on what's celebrated in your social circle
- Political opinions adopted from respected figures in your group
- Food preferences that align with your aspirational identity group
- Hobbies pursued because they signal belonging to certain communities
- Creative outputs that unconsciously mirror admired creators
- Parenting styles that copy other parents you respect
Given the above, will antiandrogens make me more introverted? And if so, are there cognitive benefits to introversion? (I think so)
2 days ago started taking the supposed mild but statistically significant antiandrogens and OTC supplements Reishi + Chasteberry + Spearmint
I'll be amused if that before long ends my "frequent public posting" streak
(Vague musing)
There's a type of theory I'd call a "Highlevel Index" into an information body, for example, Predictive Processing is a highlevel index for Neurology, or Natural Selection is a highlevel index for Psychology, or Game Theory and Signaling Theory are highlevel indexes for all kinds of things.
They're tools for delving into information bodies. They give you good taste for lower level theories, a better feel for what pieces of knowledge are and aren't predictive. If you're like me, and you're trying to study Law or Material Science, but you got no highlevel indexes for these domains, you're left standing there, lost, without evaluability, in front of a vast sea of lower-level more detailed knowledge. You could probs make iterative bottom up progress by layer for layer absorbing detail-info and synthesizing or discovering higher-level theories from what you've seen, but that's an unknown and unknowable-feeling amount of work. Standing at the foot of the mountain, you're not feeling it. There's no affordance waiting to be grasped.
One correct framing here is that I'm whining because not all learning is easy.
But also: I do believe the solutionspace ceilng here is much higher than we notice, and that marginal exploration is worth some opportunity cost.
So!
Besides what's common knowledge in rat culture, what are your fave highlevel indexes?
What non-redundant authors besides Eliezer & co talk a lot in highlevel indexes?
Are there established or better verbal pointers to highlevel indexes?
Insightful: https://takingchildrenseriously.com/the-evolution-of-culture/
- Best intro on memetics I've seen
- Gave me an additional "evaluability dimension" for cultures and history
For this I could write an app that performs a gradual translation to chinese on the .epub file of a fiction I'm currently addicted to
Overly optimistic ballpark estimate is "800k words of text are enough to learn recognize 4k chinese characters"
Evidence in favour:
- Priors allow the possibility that languages are not created equal and lead to different cognitive speeds
- Common discourse is biased towards assuming neutrality in all tribal identity adjacent matters AND the two visible opinions are "English & Chinese are equally fast" and "Chinese is faster" -- "English is faster" is a missing opinion
- Chinese websites tend to look "full of text" to me, which is explained neatly if information absorption rate is higher
- (weak) The seeming competence of China
Evidence against:
- Priors allow the possibility that languages are created equal bc they do not present bottlenecks in cognition
Strongly considering learning to read in Chinese based on vague anecdotal reports pointing to higher reading / scanning speeds
(also as a status symbol / intelligence signal, tbc)
True!
Useless knowledge should neither be learned nor compressed, as both takes cognition.
The way I put that may have been overly obscure
But I've come to refer in my mind to the way the brain does chunking of information and noticing patterns and parallels in it for easier recall and use as just Compression.
Compression is what happens when you notice that 2 things share the same structure, and your brain just kinda fuses the shared aspects of the mental objcts together into a single thing. Compression = Abstraction = Analogy = Metaphor. Compression = Eureka moments. And the amazing thing is the brain performs cognition on compressed data just as fast as on original data, effectively increasing your cognitive speed.
For example, I think there's large value in merging as much of your everyday observational data of humans as feasible together into abstracted psychology concepts, and I wanna understand models like BigFive (as far as they're correct) much better on intuitive levels.
Yes. The product I bought identifies itself as "Sceletium tortuosum".
I've only tried 1 brand/product, and haven't seen any outstanding sources on it either, so I can't offer much guidance there.
I can anecdotally note that the effects seem quite strong for a legal substance at 0.5g, that it has short term effects + potentially also weaker long term effects (made me more relaxed? hard to say) (probs comparable to MDMA used in trauma therapy)
Compressing existing knowledge >> Acquiring new knowledge
- Kanna is a legal substance dubbed Nature's MDMA, likely working as an empathogen
- Kanna seems to quite precisely increase my BigFive::Extraversion::Warmth[1]
- Subjectively this feels like suddenly all the people you see are "your friends", and like you could just walk up to them, strike up a conversation, and start bantering
- DHEA is a metabolic precursor of testosterone, among other things, and an OTC supplement in the US
- Single dose DHEA seems to 1:1 have increased my BigFive::Extraversion::Assertiveness[1]
- As a subset of that, it seems to have boosted my executive functioning
- The execfunc boost is (as I read later) vaguely consistent with ADHD research
- It also changes my cognition and motivation in ways I'm not fully comfortable with
There's a subjective 15% chance the mindstate switch was instead placebo-induced
Downvoters: consider "Deception increases predictability"
"Honesty reduces predictability" seems implausible as a thesis.
OpenAI successfully waging the memetic war, as usual
Awesome!
My faves are #4 Intuition Flooding and #12 Incremental Reading. Will try them when I have slack and a topic of interest.
#2 Immersive Reading seems intriguing. I've noticed in myself a sense of my reading speed being capped by mental critical filtering processes. I feel like I could increase my comprehension speed at the cost of absorbing contents less discriminately.
#3 Recursive Sampling and #7 Spot the Core are strategies I've discovered myself, but no less useful for that.
#8. Triangulating Genius seems effortful but like a great fit for particular cases. #9 Expert Observation is great where the material exists. (Someone should do youtube videos liveblogging their math learning or social situation navigation, for me)
Amusing instructive and unfortunate this post's actual meaning got lost in politics. IMO it's one of the better ones.
Am left wondering if "local" here has a technical meaning or is used as a vague pointer.
What people need to get is that Lying is the weaker subset of Deception. It's the type you can easily call out and retaliate against.
Which is why we evolved to have strong instinctive reactions to it.
I take away:
- While doubt may involve encountering disconfirming evidence for a held belief - and it's proper to immediately update on the doubt-creating evidence and thereby factor the expected result of further inquiry into your belief-state, -
- Doubt itself is a pointer to a location of yet-unseen evidence. To a specific line of inquiry that may or may not disconfirm the held belief in question.
- The inverse, or perhaps a generalization to positive and negative cases, is then Suspicion.
- Suspicion points to locations of likely belief-creating or belief-modifying evidence.
Edit: meditating on what this post points to - finding in myself instances of the sensation of rational-doubt, and dwelling on them - proved useful.
I find it important for rationalists to think and talk more about deception.
While in honesty the post is a bit long for my taste, I like the way it approaches the overton window with this kind of dark-artsy, borderline-political topic and presents a plainly-insightful case study.
I'd say Accidentally Load Bearing structures are (statistically speaking) always the work of another optimizer: - someone saw the structure, and built another (architectural, behavioral,) structure on top of it.
So the key question is whether or not this structure may at some point have seemed useful to someone. (In a way that can be retrospectively broken.)
I think the post loses out on mental succinctness not explaining this.
Thanks for starting this rebellion, Eliezer.
Splitting the Great Idea into parts
Applied to "The Sequences", or Rationality:
- a collection of good predictive models
a foundation for a culture more productive and virtuous than mainstream culture
- Treating every additional detail as burdensome
It helps to apply scepticism to every post, and internally rank posts by usefulness and credence.
(I've since found https://www.lesswrong.com/rationality, which does the job.)
Same.
The “how to think” memes floating around, the cached thoughts of Deep Wisdom—some of it will be good advice devised by rationalists. But other notions were invented to protect a lie or self-deception: spawned from the Dark Side.
It's so unfortunate that "how to think" - the rules of proper belief - are not hardcoded in the system's firmware, and must instead be entered via user-supplied data the belief system is built to manage. I'd frame that this post is centrally about this user-caused systembehavior-variability, and the implicit security flaw.
Another aspect: Dominant memes - that is, memes that feel good, fair & highstatus - can be functionally dysvirtuous and unilaterally damaging.
Very cool. Less of a distinct mental handle, more of a subtle mental strategy one can find oneself executing across time.
This cognitive phenomenon is usually lumped in with “confirmation bias.” However, it seems to me that the phenomenon of trying to test positive rather than negative examples, ought to be distinguished from the phenomenon of trying to preserve the belief you started with. “Positive bias” is sometimes used as a synonym for “confirmation bias,” and fits this particular flaw much better.
Subtle distinction I almost missed here. Worth expanding.
I think this page would be more useful if it linked to the individual sequences it lists.
As far as I've seen, there is no page that links to all sequences in order, which would be useful for working through them systematically.
This works on a number of levels, although perhaps the most obvious is the divide between styles of thought on the order of "visual thinker", "verbal thinker", etc. People who differ here have to constantly reinterpret everything they say to one another, moving from non-native mode to native mode and back with every bit of data exchanged.
Have you written more about those different styles somewhere?
And this is how talking is anchrored in Costly Signaling.
(Note that "I dunno, probably around 9 pm." is still an assurance, though of a different kind: You're assuring that 9 pm is an honest estimate. If it turns out you make such statements up at random, it will cost you.)
And that's why talking can convey information at all.
TL;DR It often takes me a bit to grasp what you're pointing to.
Not because you're using concepts I don't know but because of some kind of translation friction cost. Writing/reading as an ontological handshake.
For example:
>How does task initiation happen at all, given the existence of multiple different possible acts you could take? What tips the mind in the direction of one over another?
The question maps obviously enough to my understandings, in one way or another*, but without contextual cues, decoding the words took me seconds and marginally-conscious searching.
* I basically took it as "How do decisions work?". Though, given the graphic, it looks like you're implying a kind of privileged passive state before a "decision"/initiation happens, but that part of the model is basically lost on me because its exact shape is within a meaning searchspace with too many remaining degrees of freedom.
>There are four things people confuse all the time, and use the same sort of language to express, despite them meaning very different things:
I think my brain felt a bit of "uncertainty what to do with the rest of the sentence", in a "is there useful info in there" sense, after the first 9 words. I think the first 9 words sufficed for me, they (with context below) contained 85% of the meaning I took away.
>Whether you're journaling, Internal Double Cruxing, doing Narrative Therapy, or exploring Internal Family Systems, there's something uniquely powerful about letting your thoughts finish.
Strikes me as perhaps a plain lack of Minto (present your conclusion/summary first, explanations/examples/defenses/nuances second, for that's how brains parse info). For the first half of the sentence my brain is made to store blank data, waiting for connections that will turn them into info.
Also reminded of parts of this, which imo generalizes way beyond documentations.
Dunno if this is even useful, but it'd be cool if you had some easy to fix bottlenecks.