Teasers for the previous 100 posts
1/ The real AI challenge is figuring what will get a desired message across in advance, not generating enough entry-level drivel to sometimes get lucky, and that requires context we don’t know how to model.
2/ The “meaning” of communicative acts is the influence they have on action; more precisely, it is the delta they have on the actions agents would have otherwise taken.
3/ Language is the solution to a coordination problem between people who want to get stuff done, and a lot of that has to do with figuring out what futures people are mutually interested in and attempting to make them happen.
4/ The signals we use to interpret our world and, importantly, our effect on the world are markers we hope indicate actual underlying realities—our understanding of reality always bottlenecks at these markers.
5/ John Nerst’s (@everythingstudies) Cat Couplings (https://everythingstudies.com/2019/10/30/cat-couplings/) are a memetic technology that creates and reinforces types, e.g., Types of Guys.
6/ Generalized Reading is a concept we use a lot in TIS, which just means everything that can be manipulated by people will be used for communication—here are three worked-out examples.
7/ Cherry-picked examples make the dream of AI—and it’s current status—hard to assess through over-hyping, which actively holds back progress from being reported, since it won’t be perceived as progress at all.
8/ The pragmatic usage of language we see in people is grounded in the modeling of possible futures every agent does for every other agent, allowing us to be vague or wrong, because we expect the other person to snap onto useful interpretations.
9/ The problem of selecting options from possibilities is always pragmatically scoped to the circumstances that you expect selection to take place—that’s why rules of thumb are never universal—examples abound from drug mules to policing cinematic taste.
10/ Predictability is a form of communication—that’s why in adversarial games it’s so dangerous: being obvious is a “tell” and you don’t want to tell your opponent anything real, unless you can condition them and exploit their new found predictability.
11/ How much more can people know themselves than observers can know them given that they have privileged access to information, but extra motivation to distort its interpretation?
12/ Compartmentalizing your mind—whether out of fear or a sincere desire for self-manipulation—leaves your mind less efficient in its ability to coordinate itself or throw its full weight at something it might otherwise gobble-up delightedly.
13/ A system that’s trying to satisfy you will modify your preferences, but we tend to think of current content recommendation systems as fundamentally more pointed and agential, without much evidence to that effect.
14/ Thinking about things in terms of “either it happens or it doesn’t” only works if that thing has a reasonable probability of going either way, otherwise you’re liable to end-up with bad definitions that don’t tell you what you think they will.
15/ What are attempts at artificial ethical judgements actually trying to do?
16/ A quick overview of how some of us conceptually cut-up situations into elements of strategy.
17/ If you think about strategizing you’ll probably be bad at it, because you weren’t trained to do it consciously and most situations specifically try to winnow away your ability to.
18/ There are things you know that others can’t because they’re not you, and ways of thinking about things you don’t know because of lack of exposure: how do these two things interact in communication between different parties?
19/ Socializing one’s self into taste is often more effective than performing taste, because it makes use the reality of your internal preference engine, which you can’t consciously simulate with high fidelity.
20/ Information that’s preserved by a group is what coordinates it, and coordinating requires this kind of preservation—and often protection from other groups who would like to intrude on the current order.
21/ Helping people play signaling games or trying to enforce them away, tends to make them more nuanced and powerful.
22/ How hackable is the world we live in?
23/ The balance in human life, and indeed in most competition, is between doing what everyone else is doing to be part of the herd, and doing something different to stand apart from everyone else in the herd.
24/ A lot of spookiness comes from introducing frames we don’t know how to properly apply to a situation and then being stunned by what these frames insist is or can’t be true.
25/ “Will my uploaded brain whistle?”
26/ “free associative therapy can be thought of as a special sort of game, where the speaker conceives of the listener as having a certain special knowledge which can fix the speaker’s problems”
27/ The often overlooked meta-goal of “having a good time” underlies the dynamics of most games in ways often not expressed in formal rules or analysis—what happens when a player doesn’t abide by it?
28/ Cues (social or otherwise) shape how we view a situation, leading to metonyms (”suits” instead of business men or women) that compress how we see the world into what we see in the world.
29/ In the real world, people continuously develop their internal conceptual ontologies to map what they should do about recognized events, which often makes hypotheticals that play with these ontologies poor simulations of reality.
30/ The AI concept of “wireheading”—an agent changing its reward function to make it easy to achieve—isn’t a case of the AI cheating a test, it’s a case of bad definitions that make you think there’s a paradox.
31/ Doing something because it’s worked before is fundamentally different than doing something because you have a model of why it might work.
32/ What potlucks provide as a social technology: a series of intertwining games that people self-select themselves into, often benefiting the group total.
33/ The things we say are often more about carving grooves in social dynamics to make certain actions favored or disfavored, rather than saying anything specific about the world.
34/ A smattering of vocabulary we use to describe the games we see and the bottlenecks we use to explain them here at the Inexact Sciences.
35/ Selection games (like you choosing which shampoo to buy and brands competing to be chosen) are underwritten by the fact that complete honesty is not beneficial to the party subject to selection, allowing the Detective genre to arise.
36/ Searching for examples to describe new concepts is the search for “things that have this strong similarity to other lived experiences, but crucially have not yet been generalized”—prototypicality alone is not enough.
37/ Realness is often fetishized, and when it is considered consciously as such, it becomes a symbol in and of itself, that can be played around like the rest of our cultural symbols…is that bad or not?
38/ J. Robert Oppenheimer on the intimate vs. the surveying view of phenomena.
39/ Formless empiricism is our default working methodology in TIS: we aim to be doing pragmatically useful things, underwritten by our experience of reality, without requiring immediately paying the meaning debt through direct experimentation.
40/ Pragmatically scoped notes on an academic paper about game theoretic pragmatics: how should we mechanically model communication as a game?
41/ Narratives give us something that procedural description or declarative denotation of a system/phenomenon doesn’t: an opportunity to excavate categories through the patterns across narratives.
42/ Good examples are often worth more than large sample sizes.
43/ What if language is just finely-tuned associative babbling?
44/ Anti-inductivity is rarely a direct property of a system—it’s a limiting case that systems often asymptotically approach or avoid.
45/ Saying things that are hard to be reasonable about discussing or analyzing is selected-for in philosophical discourse.
46/ Two years into TIS: what are we doing and why are we here?
47/ Strong emergence is an incoherent proposition, as it can always be factored-out of any argument for its existence.
48/ “Thought sinks” that we keep coming back to are often interpreted as parts of ourselves that need attention or as neuroses that need resolution, but what if they’re requests to go meta?
49/ Winning a bored game in a way that ruins your friendships violates the meta-game of board games because “board games” are a social construct.
50/ If you want to be an X, you will often keep scraping at the bottom of the barrel of what can be done, because you’re incentivized to look like those who came before you.
51/ What is the relationship between new language and new concepts?
52/ We can learn about others by learning our own responses to different stimuli, and this is precisely why communication is so often conceived first in terms of its effects and only secondarily in terms of its literal content.
53/ Calling “all communication manipulation” makes it seem like that manipulation is reliable, as if one was pulling a lever in a machine, but that’s certainly not how we experience it.
54/ Victory through population appears in many cultures and times—and is often conflated with reaching paradise: why does this happen and at point will population mean something other than “living humans”?
55/ “Spooky” information—like what energy healers supposedly tap into—really isn’t all that spooky when we think of mundane examples like socially contagious ideas and feelings.
56/ A critique of C. Thu Nguyen’s concept of “value clarity” that asks: isn’t single mindedness something humans are constantly drawn to across contexts, cultures, and times?
57/ Why would moderators of a deception-based game enforce a rule about not signaling you’re telling the truth?
58/ Often, a player who is guaranteed to lose is the one who can decide who will win—we can call such players kingmakers, because they decide who has dominion.
59/ A couple mutually going overboard with affection and desire in romantic relationships usually reduces communicational bandwidth.
60/ Showmanship has been a part of science for hundreds of years, especially the kind involving scientists attempting to prove something by showing that they trust their own invention enough to use it on themselves.
61/ Being wrong is useful if you can get people to correct you with information you were seeking in the first place.
62/ “Heuristics are sticky just like prices.”
63/ Intuitive and embodied theory building for understanding social dynamics supports ambiguity that direct theory building simply can’t support due to its explicit, mechanical methodology.
64/ It is very difficult to see things that you’ve backgrounded, because you assume nothing is communicating with you through their details.
“We pushed the empty frame of reason out the cabin door
No we won’t be needing reason anymore.”
66/ How much of the technological and biological knowledge that got us here is stored only in lore and then lost, sometimes forever?
67/ Categories are heuristics for thinking about what we want from the categories we choose to name.
68/ Establishing a model of setup and payoff and exploiting its legibility after it’s established is the essence of performance.
69/ The way people use language to cut-up a complex conceptual space—like chess or biology—tells you a lot about what you can expect them to know.
70/ Perceivable clusters exist in the space of human attention, which has been pragmatically tuned to accent certain features.
71/ Testing science on yourself isn’t just performative—it’s good ethics.
72/ When you assume everyone is constantly self-interested, you tend to over-estimate how much they want because you’re not used to systematically reasoning out what they want expect, except in exceptional cases.
73/ Torque—when one party exaggerates its actual stance in the hopes of forcing the other party to meet them in the middle—acts on rules and policy all the time, because policy writers don’t trust enforcers or enforcees.
74/ Fastmusic (via machine learning) is going to change music like fastfood changed food.
75/ We always frame things inexactly for our own pragmatic purposes—these are called biases when they don’t work, but studying something only when it doesn’t work is called “not establishing a base rate” and it’s bad practice.
76/ Metrics express the institutions that choose them like writing expresses a writer.
77/ To represent one’s self is to attempt to choose one’s category in the mind of another, to do so in a way that slides attention off you is ontological camouflage.
- We understand a lot about consciousness.
- We fundamentally don’t understand consciousness.
- Consciousness exists on a spectrum.
- There is no test for consciousness.
- Consciousness is a process.
79/ Discovering the lack of novelty in your ideas is painful, but trying to be somewhat novel anyway or simply not worrying too much is often the only way forward.
80/ What AI Safety as a community is today emerged largely out of non-academic channels, making it an interesting case study of knowledge logistics in the era of the internet.
81/ A case study in how finite games evolve out of the infinite one, set at a Maui pool party.
82/ A dream reveals to Collin that the reason why people talk about utility functions is that they’re secretly hoping functions are simple and understandable.
83/ Misrepresentation is often weeded out through iterated games that would punish a misrepresenter, raising the question: when should you game a selection game?
84/ Sure, self-experimentation might be ethical, but there’s also a lot of machismo driving the engine of science throughout history.
85/ In a world of 1 hour time blocks, 40 minute meetings are a social technology that allows people to contract or expand the meeting time as needed, while saving face eitherway.
86/ Focusing on instrumental truths still leads one to models that correspond to reality, as the Good Regulator Theorem might suggest.
87/ Honesty is the best policy in communities that are ecologically huddled, otherwise surveillance becomes necessary to make it the best policy.
88/ Who is causing the torque in the statement “any alcohol during pregnancy is dangerous and reckless”?
89/ A general principle: you are probably “late” to any noticeable trend, good or bad.
90/ On the origins of the emotionally two-faced monotheistic God.
91/ A little bit knowledge logistics about asymmetrical justice (as a treat).
92/ Sorites paradox: actually you’re just bad at thinking about pragmatics.
93/ “Science is inextricably enfleshed in tinkering,” says Alex Boland, and TIS says “amen.”
94/ Showing how self-deception happens is the only reliable way to establish enough common ground to talk about how to make it not happen.
95/ Least Common Denominator messaging: ”idiot proof” warnings and communication meant to cover-up any actual advice that could have been given, due to the fear of litigation of actionable but only contextually applicable information.
96/ Flags are one of the oldest still used symbolic communication technologies—perhaps because of how easy they are to repurpose and contextualize.
97/ Zoom out and contextualize—a move in the game of analysis that stops you from being a pedant about definitions, and asks: does the indexical meaning of this thing present anything useful we can decontextualize?
98/ I explain why I want a theory of social dynamics: because I want to have been everybody.
99/ The scariest possible thing is other people finding things you allow to grow in your life illegible, unreasonable, and unmanageable—how will you convince them you’re a reliable coordinator if they don’t get it?
100/ A little retrospective (as a punishment).