Mir(Mir)

Karma: 26

Flowers are selective about what kind of pollinator they attract. Diurnal flowers use diverse colours to stand out in a competition against their neighbours for visual salience. But flowers with nocturnal anthesis are generally white, as they aim only to outshine the night.

O mysterious temple of chance, for what did the scattered parts of your whole link arms to shape you? And the mysterious ceasefire they declared to heed you? O quiescent eye of a hurricane, what holds you together? O molecular weave of water and sunlight, why do you even bother?

Mir 21 Sep 2023 22:45 UTC
1 point
0
on: Memory bandwidth constraints imply economies of scale in AI inference
Wow, this is a good argument. Especially if assumptions hold.
1. The ALU computes the input much faster than the results can be moved to the next layer.
2. So if the AI only receives a single user’s prompt, the ALUs waste a lot of time waiting for input.
3. But if many users are sending prompts all the time, the ALUs can be sent many more operations at once (assuming the wires are bottlenecked by speed rather than amount of information they can carry).
4. So if your AI is extremely popular (e.g., OpenAI), your ALUs have to spend less time idling, so the GPUs you use are much more cost-effective.
5. Compute is much more expensive for less popular AIs (plausibly >1000x).

Mir 19 Sep 2023 18:56 UTC
1 point
0
in reply to: TsviBT’s comment on: Sum-threshold attacks
cool third point! i may hv oversold the point in my first comment. i too try to name things according to their thingness, but not exclusively.
to make a caricature of my research loop, i could describe it as
1. trying to find patterns that puzzle me (foraging),
2. distilling the pattern to its core structure and storing it in RemNote (catabolic pathway),
3. mentally trying to find new ways to apply the pattern
  1. ie, propagating it, installing hooks (which I call isthmuses) into plausibly-related contexts such that new cryptically-related observations are more likely to trigger an insight (metaphor), allowing me to generalise further or discover smth i need to refactor
4. going abt business as usual, repeating 1-3 until unfolding branch meets unfolding branch from the other side, indicating i might hv found a profitable generalisation
  1. an important consideration re keeping isthmuses alive enough to trigger connections: i don’t want to hv memorised this specific instantiation of the pattern so it’s crystal clear. if it fits neatly into a slot and it’s comfortable w its assigned niche, it’s unlikely to trigger in novel situations. imprecision/fuzziness is good when the concept is still in exploratory phase (and not primarily tool-stage).
5. fix everything
loop is often bottlenecked by the high cost of refactoring anything. i rly wish i could find a general algorithm/strategy for refactoring complex systems like this, or a clever approach to building that minimises/eliminates the need.
the optimal conlang isn’t a new set of words. it’s a new set of practices for naming things, unnaming things, generalising & specialising, communal decision-processes for resolving conflicts, neat meta-structures that minimise cost of refactoring (somehow), enabling eager contributors w minimal overhead & risk of degeneration, etc.

Mir 19 Sep 2023 14:44 UTC
1 point
0
on: Write the Worst Post on LessWrong!
The thing people need to realize is that when somebody writes a bad post, it doesn’t harm the readers (except insofar as you falsely advertised quality). If something is badly argued readers are immune. If something is persuasively argued, but wrong, readers that fall for it now have an opportunity to discover a hole in their epistemic defenses.
Mostly, people read arguments as a proxies for making judgments about the quality of the author, or whether they agree/disagree with them. Notice that these purposes are orthogonal to actually learning something.
Unrelatedly, what would you say are your top ideas/insights/writings as judged by the usefwlness you think they’d confer to you-a-year-or-two-ago? (Asking you this as a proxy for “give me your best insights” because I feel like the latter causes people to overweight legibility or novelty-relative-to-average-community-member or something.)

Mir 19 Sep 2023 14:05 UTC
1 point
0
on: Computer Input Sucks—A Brain Dump
This is awesome. I found it via searching LW for variations of “voice typing”, Which I was motivated to search because I had just discovered that average conversational speed is around 3x average typing speed (~150 vs ~50 wpm, cf ChatGPT). (And reading speeds are potentially in the thousands.)
At the moment, I’m using Windows Voice Access. It’s accurate, has some nice voice commands and gives you visual feedback while speaking. The inadequacy, for me, is the lack of immediate feedback (compared to typing), and customisability. I’ll attempt to test your repo tomorrow.
Also, I’m surprised that stenotypers can type faster than regular keyboards. I had previously speculated about the benefits of making keyboards more or less modular. I thought making them more modular would face similar problems as Ithkuil: too many serial operations to build the vector you mean. English relies extremely on caching specific words for specific situations with very shallow generalisations, trading semantic reach for faster lookup-times, or something. But I guess what matters most is modularity in the right places, and mainstream keyboards aren’t optimised for that at all.
I speculate that the ideal keyboard should have keys that chunk word-pieces somewhat according to Zipf’s law: the $N$ th most common key should be such that the frequency by which you have to use it is ~ $\frac{1}{N}$ . I guess there should also be a term somewhere for the distance the finger has to travel to reach a key or something, to be able to weight the frequency of keys by some measure for ergonomicity.

Mir 19 Sep 2023 12:34 UTC
6 points
in reply to: Johannes C. Mayer’s comment on: Johannes C. Mayer’s Shortform
I highly recommend trying to get a prescription for something like adderall (or dextroamphetamine, lisdexamphetamine) if your doctor is willing to diagnose you with ADHD. Just go off it if it doesn’t work, but it seems likely that it will given your response to bupropion. Some vague reasons for recommending it are:
1. Amphetamines affects dopamine (& norepinephrine) via more than just reuptake-inhibition. I’m not sure yet which mechanisms drive desensitization, it just seems tentatively better to spread the attack vectors out.
2. I’ve mostly forgot the reasons I suggest amphetamines over methylphenidate, but at least it has a stronger effect on egosyntonic behaviour (/executive function or whatever you want to call it) than bupropion.
Furthermore, if you’re on amphetamines/methylphenidate, I would recommend sort of using it strategically. I only take them on days I know I plan to be productive, and have hope that they enable me to be productive. If I ever take them on a day I fail to motivate myself, it weakens the semiotic/narrative signal for my brain to switch into a mode where it expects productivity & silences competing motivations. Plus, if I know I mostly won’t be productive for a while (e.g. vacation or something), I go off them to reset desensitisation (usually resets a long way within a week it seems).
I’ve been on lisdexamphetamine for 1.5 years, and they still produce a very clear effect. I’ve had depression and, turns out, ADHD, for most of my life. People who have no dysregulation of dopamine will probably have negative effect from too much upregulation (e.g. reduced mental noise/creativity, inability to zoom out from activities and reprioritise).

Mir 19 Sep 2023 11:37 UTC
1 point
on: Finding Useful Things
This is among the top questions you ought to accumulate insights on if you’re trying to do something difficult.
I would advise primarily focusing on how to learn more from yourself as opposed to learning from others, but still, here’s what I think:
I. Strict confusion
Seek to find people who seem to be doing something dumb or crazy, and for whom the feeling you get when you try to understand them is not “I’m familiar with how someone could end up believing this” but instead “I’ve got no idea how they ended up there, but that’s just absurd”. If someone believes something wild, and your response is strict confusion, that’s high value of information. You can only safely say they’re low-epistemic-value if you have evidence for some alternative story that explains why they believe what they believe.
II. Surprisingly popular
Alternatively, find something that is surprisingly popular—because if you don’t understand why someone believes something, you cannot exclude that they believe it for good reasons.
The meta-trick to extracting wisdom from society’s noisy chatter is learn to understand what drives people’s beliefs in general; then, if your model fails to predict why someone believes something, you can either learn something about human behaviour, or about whatever evidence you don’t have yet.
III. Sensitivity >> specificity
It’s easy to relinquish old beliefs if you are ever-optimistic that you’ll find better ideas than whatever you have now. If you look back at what you wrote a year ago, and think “huh, that guy really had it all figured out,” you should be suspicious that you’ve stagnated. Strive to be embarrassed of your past world-model—it implies progress.
So trust your mind that it’ll adapt to new evidence, and tune your sensitivity up as high as the capacity of your discriminator allows. False-positives are usually harmless and quick to relinquish—and if they aren’t, then believing something false for as long as it takes for you to find the counter-argument is a really good way to discover general weaknesses in your epistemic filters.^[1] You can’t upgrade your immune-system without exposing yourself to infection every now and then. Another frame on this:
I was being silly! If the hotel was ahead of me, I’d get there fastest if I kept going 60mph. And if the hotel was behind me, I’d get there fastest by heading at 60 miles per hour in the other direction. And if I wasn’t going to turn around yet … my best bet given the uncertainty was to check N more miles of highway first, before I turned around.
— The correct response to uncertainty is *not* half-speed — LessWrong
IV. Vingean deference limits
The problem is that if you select people cautiously, you miss out on hiring people significantly more competent than you. The people who are much higher competence will behave in ways you don’t recognise as more competent. If you were able to tell what right things to do are, you would just do those things and be at their level. Innovation on the frontier is anti-inductive.
If good research is heavy-tailed & in a positive selection-regime, then cautiousness actively selects against features with the highest expected value.
— helplessly wasting time on the forum
finding ppl who are truly on the right side of this graph is hard bc it’s easy to mis-see large divergence as craziness. lesson: only infer ppl’s competence by the process they use, ~never by their object-level opinions. u can ~only learn from ppl who diverge from u.
— some bird
V. Confusion implies VoI, not stupidity
look for epistemic caves wherefrom survivors return confused or “obviously misguided”.
— ravens can in fact talk btw
1. ^
  Here assuming that investing credence in the mistaken belief increased your sensitivity to finding its counterargument. For people who are still at a level where credence begets credence, this could be bad advice.
  VI. Epistemic surface area / epistemic net / wind-wane models / some better metaphor
  Every model you have internalised as truly part of you—however true or false—increases your ability to notice when evidence supports or conflicts with it. As long as you place your flag somewhere to begin with, the winds of evidence will start pushing it in the right direction. If your wariness re believing something verifiably false prevents you from making an epistemic income, consider what you’re really optimising for. Beliefs pay rent in anticipated experiences, regardless of whether they are correct in the end.

Mir 17 Sep 2023 21:22 UTC
1 point
0
on: Telopheme, telophore, and telotect
is that spreadsheet of {Greek name, Germanic name, …} public perchance?

Mir 16 Sep 2023 23:02 UTC
1 point
0
in reply to: Alex Vermillion’s comment on: Sum-threshold attacks
Thank you! : )
Please continue complimenting people (or express gratitude) for things you honestly appreciate.

Mir 16 Sep 2023 18:43 UTC
2 points
0
on: What’s up with psychonetics?
i googled it just now bc i wanted to find a wikipedia article i read ~9 years ago mentioning “deconcentration of attention”, and this LW post came up. odd.
anyway, i first found mention of it via a blue-link on the page for Ithkuil. they’ve since changed smth, but this snippet remains:
After a mention of Ithkuil in the Russian magazine Computerra, several speakers of Russian contacted Quijada and expressed enthusiasm to learn Ithkuil for its application to psychonetics—
deconcentration of attention
i wanted to look it up bc it relates to smth i tweeted abt yesterday:
unique how the pattern is only visible when you don’t look at it. i wonder what other kind of stuff is like that. like, maybe a life-problem that’s only visible to intuition, and if you try to zoom in to rationally understand it, you find there’s no problem after all?
oh.
i notice that relaxing my attention sometimes works when eg i’m trying to recall smth at the limit of my memory (or when it’s stuck on my tongue). sorta like broadening my attentional field to connect widely distributed patterns. another frame on it is that it enables anabranching trains of thought. (ht TsviBT for the word & concept)
An anabranch is a section of a river or stream that diverts from the main channel or stem of the watercourse and rejoins the main stem downstream.
here’s my model for why it works:
1. Working Memory is a loop of information (parts of the chewbacca-loop is tentatively my prime suspect for this). it’s likely not a fully synchronised clock-cycle, but my guess is that whenever you combine two concepts in WM, their corresponding neural ensembles undergo harmonic locking to remain there.^[1]
2. every iteration, information in the loop is a weighted combination of:
  1. stuff that’s already in working memory
  2. new stuff (eg memories) that reaches salience due to sufficient association with stuff from the previous iteration of WM
  3. new stuff from sensory networks (eg sights, sounds) that wasn’t automatically filtered out by top-down predictions
3. for new information (B or C) to get into the loop, it has to exceed a bottom-up threshold for salience.
4. the salience network (pictured below) determines the weighting between the channels (A, B, C), and/or the height of their respective salience thresholds. (both are ways to achieve the same thing, and i’m unsure which frame is more better.)
5. “concentrating hard” on trying to recall smth has the effect of silencing the flow of information from B & C, such that the remaining salience is normalised exclusively over stuff in A. iow, it narrows the flow of new information into WM.
  1. (bonus point: this is what “top-down attention” is. it’s not “reach-out-and-grab” as it may intuitively feel like. instead, it’s a process where the present weighted combination of items in WM determines (allocates/partitions) salience between items in WM.)
6. this is a tradeoff, however. if you narrow all salience towards eg a specific top-down query $↓_{Q}$ , this has smth like the following two effects:
  1. you make it easier to detect potential answers $↑_{Q}$ by reducing the weight of unrelated competing noise
  2. but you also heighten the salience threshold $↑_{Q}$ must exceed to reach you
in light of this, here some tentative takeaways:
1. if your WM already contains sufficient information to triangulate towards the item you’re looking for, and the recollection/insight is bottlenecked by competing noise, concentrate harder.
2. but if WM doesn’t have sufficient information, concentrating could prematurely block essential cues that don’t yet strongly associate from $↓_{Q}$ directly.
3. and in cases where features in $↓_{Q}$ itself are temporarily interfering w the recollection, globally narrowing or broadening concentration may not unblock it. instead, consider pausing for a bit and try to find alternative ways to ask $Q$ .
Ithkuil
Natural languages are adequate, but that doesn’t mean they’re optimal.
— John Quijada
i’m a fan of Quijada (eg this lecture) and his intensely modular & cognitive-linguistics-inspired conlang, Ithkuil.
that said, i don’t think it sufficiently captures the essences of what enables language to be an efficient tool for thought. LW has a wealth of knowledge about that in particular, so i’m sad conlanging (and linguistics in general) hasn’t received more attention here. it may not be that hard, EMH doesn’t apply when ~nobody’s tried.
We can think of a bunch of ideas that we like, and then check whether [our language can adequately] express each idea. We will almost always find that [it is]. To conclude from this that we have an adequate [language] in general, would [be silly].
— The possible shared Craft of Deliberate Lexicogenesis (freely interpreted)
1. ^
  Furthermore, a relationship with task performance was evident, indicating that an increased occurrence of harmonic locking (i.e., transient 2:1 ratios) was associated with improved arithmetic performance. These results are in line with previous evidence pointing to the importance of alpha–theta interactions in tasks requiring working memory and executive control. (Julio & Kaat, 2019)

Mir 13 Sep 2023 18:28 UTC
1 point
0
in reply to: TsviBT’s comment on: Sum-threshold attacks
i think that goes into optimising for b in my taxonomy above. how easy is it to recall the structure of the thing once you’ve recalled the word for the thing? these are just considerations, and the optimal naming strat varies by situs ig. 🍵

Mir 12 Sep 2023 21:42 UTC
1 point
0
on: The Evidence for Question Decomposition is Weak
This seems like a question one shouldn’t be using statistical evidence to make an opinion about. It seems tractable to just grok (and intuify) the theoretical considerations and thus gain a much better understanding of when vs when not to decompose (and with how much granularity and by which method). Deferring to statistics on it seems liable to distort the model—such that I don’t think a temporary increase in the accuracy of final-stage judgments would be worth it.

Mir 12 Sep 2023 20:13 UTC
0 points
−2
in reply to: niplav’s comment on: Subscripts for Probabilities
did you know that, if you’re a hermit, you get infinite weirdness points?

✧*｡ヾ( >﹏< )ﾉﾞ✧*。

Mir 11 Sep 2023 12:38 UTC
8 points
1
on: Sum-threshold attacks
when making new words, i try to follow this principle:
label concepts such that the label has high association w situations in which you want the concept to trigger.^[1]
the usefwlness of a label can be measured on multiple fronts:
1. how easy is it to recall (or regenerate):
  1. the label just fm thinking abt the concept?
    low-priority, since you already have the concept.
  2. the concept just fm seeing the label?
    mid-priority, since this is easy to practice.^[2]
  3. the label fm situations where recalling the concept has utility?
    high-priority, since this is the only reason to bother making the label in the first place.
if you’re optimising for b, you might label your concept “distributed boiling-frog attack” (DBFA). someone cud prob generate the whole idea fm those words alone, so it scores on highly on the criterion.
it scores poorly on c, however. if i’m in a situation in which it is helpfwl for me to notice that someone or something is DBFAing me, there are few semiotic/associative paths fm what i notice now to the label itself.
if i reflect on what kinds of situations i want this thought to reappear in, i think of something like “something is consistently going wrong w a complex system and i’m not sure why but it smells like a targeted hostile force”.
maybe i’d call that the “invisible hand of malice” or “inimicus ex machina”.
i rly liked the post btw! thanks!
1. ^
  i happen to call this “symptomatic nymation” in my notes, bc it’s about deriving new word from the effects/symptoms of the referent concept/phenomenon. a good label shud be a solution looking for a problem.
2. ^
  deriving concept fm label is high-priority if you want the concept to gain popularity, however. i usually jst make words for myself and use them in my notes, so i don’t hv to worry abt this.

Mir 9 Sep 2023 19:45 UTC
7 points
2
on: Subscripts for Probabilities
here’s the non-quantified meaning in terms of wh-movement from right to left:
$S : I think what - ---- -      image: you thinking it'll what - ---- - tomorrow      (circumfix time operator) rain    image: rain      image: rain inside your prospective thoughts$
for conlanging, i like $_{95 %}$ this set of principles:
1. minimise total visual distance between operators and their arguments
2. minimise total novelty/complexity/size of all items the reader is forced to store in memory while parsing
  1. every argument in memory shud find its operator asap, and vice versa
  2. some items are fairly easy to store in memory (aka active context)
    like the identity of the person writing this comment (me)
    or the topic of the post i’m currently commenting on (clever ways to weave credences into language)
  3. other items are fairly hard
    often the case in mathy language, bc several complex-and-specific-and-novel items are defined at the outset, and are those items are not given intuitive anaphora.
    another way to use sentence-structure to offload memory-work is by writing hierarchical lists like this, so you can quickly switch gaze btn ii., c, and 2—allowing me to leverage the hierarchy anaphorically.
so to quantify sentence $S$ , i prefer ur suggestion “I think $_{55 %}$ it’ll rain tomorrow”. the percentage is supposed to modify “I think” anyway, so it makes more sense to make them adjacent. it’s just more work bc it’s novel syntax, but that’s temporary.
$S^{'} : {what - ---- -}_{55 %}      (quantifier) I think what - ---- -      image: you thinking it'll what - ---- - tomorrow      (circumfix time operator) rain    image: rain      image: rain inside your tentetative prospective thoughts$
otoh, if we’re specifying that subscripts are only used for credences anyway, there’s no reason for us to invoke the redundant “I think” image. instead, write
it’ll rain tomorrow $_{55 %}$
in fact, the whole circumfix operator is gratuitously verbose!^[1] just write:
rain tomorrow $_{55 %}$
$S^{''} : {what - ---- -}_{55 %}      image: you tentatively thinking tomorrow what - ---- -      (time operator) rain    image: rain      image: rain inside your tentetative prospective thoughts$
1. ^
  natlangs smh…^[2]
2. ^
  tbh i wish we had an editor w optimised LaTeX keyboard shortcuts so we cud effortlessly use $underbraces or undersets!$ wherever we fancy.^[3]
3. ^
  additionally, we should just make much more use the dimensionality afforded to us by the editors we have.
  it’s ~~free~~ cheap $semiotic semantic syntactic$ real-estate.

Mir 1 Sep 2023 16:09 UTC
2 points
0
on: The Economics of the Asteroid Deflection Problem (Dominant Assurance Contracts)
Surely, there is a clever mechanism that can fix this issue? So I went to the Wikipedia page of the Free-rider Problem and scrolled to the bottom, and lo and behold it was just sitting there: Dominant Assurance Contracts.
I feel like every assurance-contract proponent went through a moment like this. It’s like “Don’t Look Up” except wrt a civilization-scale solution. I’m tempted to think some people ‘just don’t get it’ because they’re missing gears-level models re coordination problems in the first place, but that would be hasty.

Mir 13 Jan 2023 9:37 UTC
8 points
2
on: Iron deficiencies are very bad and you should treat them
If you’re vegan and drink lots of tea (guilty), consider avoiding tea at least one hour before and after the time you eat iron-rich food. The worry is that polyphenols in tea bind with non-heme iron, forming a stable-ish complex that prevents absorption. There’s weak evidence that this makes a practical difference, but it was especially relevant for me since I drink tea almost exclusively.
From the available evidence there is no need to advise any restriction on tea drinking in healthy people with no risk of iron deficiency. In groups at risk of iron deficiency the advice should be to drink tea between meals and to wait at least 1 h after eating before drinking tea. (source)
Most studies have found no association between tea intake and iron absorption. However, drinking excessive amounts of black tea may inhibit the absorption of iron, and may harm people with anaemia. (wikipedia)

Mir 13 Jan 2023 9:10 UTC
1 point
0
in reply to: Slider’s comment on: Dangers of deference
I was about to request clarification on this too. I don’t get
“science is the way of making beliefs come apart from their pre-theoretical pragmatic implications.”
And I would like to get it.

Mir 13 Jan 2023 9:07 UTC
1 point
0
on: Dangers of deference
Endorsed. I think what you should do about deferral depends on what role you wish to play in the research community. Knowledge workers intending to make frontier progress should be especially skeptical of deferring to others on the topics they intend to specialise in. That may mean holding off on deferring on a wide range of topics, because curious scientists should keep a broad horizon early on. Deferring early on could lead to habits-of-thought that can be hard to override later on (sorta like curse of knowledge), and you might miss out on opportunities to productively diverge or even discover a flaw in the paradigm.
Explorers should mostly defer on value of information, not object-level beliefs. When someone I trust says they’re confident in some view I’m surprised by, I’m very reluctant to try to tweak my models to output what I believe they believe; instead I make a note to investigate what they’ve investigated, using my own judgment of things all the way through.

Mir(Mir)

I. Strict confusion

II. Surprisingly popular

III. Sensitivity >> specificity

IV. Vingean deference limits

V. Confusion implies VoI, not stupidity

VI. Epistemic surface area /​ epistemic net /​ wind-wane models /​ some better metaphor

deconcentration of attention

Ithkuil

VI. Epistemic surface area / epistemic net / wind-wane models / some better metaphor