Bucky

Karma: 887

Bucky 11 Aug 2019 20:19 UTC
1 point

on: Why do humans not have built-in neural i/o channels?
Communication requires both input and output channels. All of the instances I can think of from the animal world involve a sense (hearing, sight, smell, touch) which has evolved with a different benefit. Then an output can evolve to take communicate using this sense as the input.
This seems orders of magnitude less complex than evolving input and output simultaneously which would be required for direct brain communication (a least I can’t think of another option).
Even if it could potentially happen, before it did there would be many instances of indirect communication evolving. Take-off happening first in a species with indirect communication is a fairly inevitable consequence of the relative complexity of the evolutions required.

Bucky 2 Aug 2019 21:28 UTC
1 point

on: Why Subagents?
Imagine a second agent which has the same preferences but an anti-status-quo preference between mushroom and pepperoni.
This would be exploitable by a third agent who is able to compare mushroom and pepperoni but assigns equal utilities to both. However the original agent described in the OP would not be able to exploit agent 2 (if agent 1′s status-quo bias is larger than agent 2′s anti-status-quo bias), so agent 3 dominates agent 1 in terms of performance.
Over multiple dimensions agent 3 becomes much more complex than agent 1. Having a status quo bias makes sense as a way to avoid being exploited whilst also being less computationally expensive than tracking or calculating every preference ordering.
Assuming agent 2 is rare, the loss incurred by not being able to exploit others is small.

Bucky 31 Jul 2019 19:41 UTC
7 points

in reply to: Dagon's comment on: Drive-By Low-Effort Criticism
Start with lower-effort posts, to get a sense of how people react to the headline and thesis statement.
Shortform seems like a great way to do this.

Bucky 24 Jul 2019 22:24 UTC
1 point

on: From Laplace to BIC
In removing the $O (1)$ terms I think we’re removing all of the widths of the peak in the various dimensions. So in the case where the widths are radically different between the models this would mean that N would need to be even larger for BIC to be a useful approximation.
The widths issue might come up, for example, when an additional parameter is added which splits the data into 2 populations with drastically different population sizes—the small population is likely to have a wider peak.
Is that right?

Bucky 21 Jul 2019 20:46 UTC
1 point

on: Laplace Approximation
Thanks for this sequence, I’ve read each post 3 or 4 times to try to properly get it.
Am I right in thinking that in order to replace $d P [θ] = d θ$ we not only require a uniform prior but also that $θ$ span unit volume?

Bucky 19 Jul 2019 9:24 UTC
4 points

in reply to: NikoTM's comment on: What do you think of cognitive types and MBTI? What type are you? What do you think is the percentage of the 16 different personality types on LessWrong?
The last one appears to be 2016 (this was a slightly wider survey which included other rationalist communities) which was before the lesswrong 2.0 relaunch. I haven’t heard of any plans for surveys—maybe a mod can fill us in.
Slatestarcodex does an annual survey of its readers. Scott pre-registers some investigations and then reports on results. This year, for example, he got a negative result on “Math preference vs Corn eating style” and more interesting results in the ongoing birth-order investigation.

Bucky 18 Jul 2019 22:51 UTC
23 points

on: What do you think of cognitive types and MBTI? What type are you? What do you think is the percentage of the 16 different personality types on LessWrong?
My own feelings on MBTI are similar to this SSC post—it’s unscientific but manages to kinda work as long as you don’t expect too much of it. I wouldn’t make any life decisions based on it!
For the third part of the question we don’t have to guess—the 2012 lesswrong survey included an MBTI question. Of the people who answered, 65% were INTP or INTJ, compared to 5-9% of Americans according to the MBTI website.

Bucky 14 Jul 2019 21:36 UTC
7 points

on: Let’s Read: Superhuman AI for multiplayer poker
Thanks for this.
Nitpick:
The description of a big blind:
Big blind: the minimal money/poker chips that every player must bet in order to play. For example, $0.1 would be a reasonable amount in casual play.
sounds more like an ante than a big blind. This is important for understanding the discussion of limping in Ars Technica.

Bucky 6 Jul 2019 22:00 UTC
1 point

in reply to: wait_here's comment on: Book Review: The Secret Of Our Success
Yes, that’s definitely upward selection pressure but I think that’s more evidence for “ability to solve problems” being the cause of our intelligence rather than “ability to transmit culture”.

Most cultural processes could be transmitted by being shown what to do and punished if you do it wrong. Language makes it easier but isn’t necessarily required. Chimps have some fairly complex tool kits knowledge of which appear to be transmitted culturally.

Bucky 5 Jul 2019 5:28 UTC
5 points

on: Everybody Knows
A version of this that I hear fairly often is “it’s common sense that...”

It works in the same way in that it makes it socially costly to argue against but is more insidious than “everybody knows” (at least in my circles “it’s common sense” has more of a veneer of respectability).

Both also have their proper uses which I think makes the improper uses more difficult to counter.

Bucky 30 Jun 2019 21:44 UTC
1 point

in reply to: johnswentworth's comment on: What are principled ways for penalising complexity in practice?
Thanks for this. I’m trying to get an intuition on how this works.

My mental picture is to imagine the likelihood function with respect to theta of the more complex model. The simpler model is the equivalent of a square function with height of its likelihood and width 1.

The relative areas under the graphs reflect the likelihood of the models. So if picturing the relative maximum likelihoods and how sharp the peak is on the more complex model gives an impression of the Bayes factor.

Does that work? Or is there a better mental model?

[Question] What are principled ways for penalising complexity in practice?

Bucky

27 Jun 2019 7:28 UTC

42 points

12 comments1 min readLW link

Bucky 25 Jun 2019 13:36 UTC
44 points

on: What’s up with self-esteem?
From the literature on self esteem
Previously, I thought that self-worth was like an estimate of how valuable you are to your peers
is sociometer theory and
Now I think there’s an extra dimension which has to do with simpler dominance-hierarchy behavior.
is hierometer theory.
Hierometer theory is relatively new (2016) and could be though of as a subset of sociometer theory if sociometer theory is interpreted more broadly. Accordingly it has less research backing it up and that which is there is mostly by the original proponents of the theory.
This paper gives an introduction to both and a summary of evidence (I found this diagram a useful explanation of the difference). The paper suggests that both are true to some extent and complement each other.
I’ve included some quotes below.
Sociometer theory
Sociometer theory starts from the premise that human beings have a fundamental need to belong (Baumeister and Leary, 1995). Satisfying this need is advantageous: group members, when cooperating, afford one another significant opportunities for mutual gain (von Mises, 1963; Nowak and Highfield, 2011; Wilson, 2012). Accordingly, if individuals are excluded from key social networks, their prospects for surviving and reproducing are impaired. It is therefore plausible to hypothesize that a dedicated psychological system evolved to encourage social acceptance (Leary et al., 1995).
…
The original version of sociometer theory (Leary and Downs, 1995; Leary et al., 1995) emphasizes how self-esteem tracks social acceptance, by which is implied some sort of community belongingness, or social inclusion.
...
In contrast, the revised version (Leary and Baumeister, 2000) emphasizes how self-esteem tracks relational value, defined as the degree to which other people regard their relationship with the individual as important or valuable overall, for whatever reason.
Hierometer theory
Like sociometer theory, hierometer theory proposes that self-regard serves an evolutionary function. Unlike sociometer theory, it proposes that this function is to navigate status hierarchies. Specifically, hierometer theory proposes that self-regard operates both indicatively—by tracking levels of social status—and imperatively—by regulating levels of status pursuit (Figure 1).
...
Note here some key differences between hierometer theory and dominance theory (Barkow, 1975, 1980), another alternative to sociometer theory (e.g., Leary et al., 2001). Dominance theory, plausibly interpreted, states that self-esteem tracks, not levels of social acceptance or relational value, but instead levels of “dominance” or “prestige,” by which some social or psychological, rather than behavioral, construct is meant.
...
Accordingly, hierometer theory proposes that higher (lower) prior social status promotes a behavioral strategy of augmented (diminished) assertiveness, with self-regard acting as the intrapsychic bridge—in particular, tracking social status in the first instance and then regulating behavioral strategy in terms of it. Note that the overall dynamic involved is consolidatory rather than compensatory: higher rather than lower status is proposed to lead to increased assertiveness. In this regard, hierometer theory differs from dominance theory, which arguably implies that it is losses in social status that prompt attempts to regain it (Barkow, 1980).
Findings
… our findings are arguably consistent with the revised version of sociometer theory, which is equivocal about the type of relational value that self-esteem tracks, and by extension, the type of social acceptance that goes hand in hand with it. Indeed, hierometer theory, and the original version of sociometer theory, might each be considered complementary subsets of the revised version of sociometer theory, if the latter is construed very broadly as a theory which states that types of social relations (status, inclusion), which constitute different types of relational value, regulate types of behavioral strategies (assertiveness, affiliativeness) via types of self-regard (self-esteem, narcissism). If so, then our confirmatory findings for hierometer theory, and mixed findings for the original version of sociometer theory, would still suggest that the revised version of sociometer theory holds truer for agentic variables than for communal ones.

Bucky 16 Jun 2019 11:30 UTC
1 point

in reply to: Zvi's comment on: No, it’s not The Incentives—it’s you
Take out the “10mph over” and I think this would be both fairer than the existing system and more effective.

(Maybe some modification to the calculation of the average to account for queues etc.)

Bucky 16 Jun 2019 10:57 UTC
1 point

in reply to: dxu's comment on: No, it’s not The Incentives—it’s you
On reflection I’m not sure “above average” is a helpful frame.

I think it would be more helpful to say someone being “net negative” should be a valid target for criticism. Someone who is “net positive” but imperfect may sometimes still be a valid target depending on other considerations (such as moving an equilibrium).

Bucky 15 Jun 2019 20:40 UTC
6 points

in reply to: Zvi's comment on: No, it’s not The Incentives—it’s you
Trying to steelman the quoted section:

If one were to be above average but imperfect (e.g. not falsifying data or p-hacking but still publishing in paid access journals) then being called out for the imperfect bit could be bad. That person’s presence in the field is a net positive but if they don’t consider themselves able to afford the penalty of being perfect then they leave and the field suffers.

I’m not sure I endorse the specific example there but in a personal example:

My incentive at work is to spend more time on meeting my targets (vs other less measurable but important tasks) than is strictly beneficial for the company.

I do spend more time on these targets than would be optimal but I think I do this considerably less than is typical. I still overfocus on targets as I’ve been told in appraisals to do so.

If someone were to call me out on this I think I would be justified in feeling miffed, even if the person calling me out was acting better than me on this axis.

Bucky 7 Jun 2019 23:23 UTC
6 points

on: Book Review: The Secret Of Our Success

Heinrich counters with his own Cultural Intelligence Hypothesis – humans evolved big brains in order to be able to maintain things like Inuit seal hunting techniques.

I can’t really see how this would work.

Partly this is because maintaining techniques like this doesn’t seem difficult enough to justify just how intelligent humans are—on a scale of chimp to human it seems like it’s more on the chimp end. The fact that inventing the technique is impressive doesn’t imply that learning the technique is impressive.

But mainly I can’t see the selection pressure for increasing intelligence. Not being able to remember the hunting technique is obviously bad but where is the upwards selection pressure?

I definitely agree that Cultural Intelligence is important and is one of the ways humans have used their intelligence but I think the Machiavellian Intelligence Hypothesis is a stronger candidate for the root cause.

Bucky 6 Jun 2019 9:38 UTC
20 points

on: Steelmanning Divination
In an innovation workshop we were taught the following technique:
Make a list of 6 things your company is good at
Make a list of 6 applications of your product(s)
Make a list of 6 random words (Disney characters? City names?)
Roll 3 dice and select the corresponding words from the lists. Think about those 3 words and see what ideas you can come up with based on them.
Everyone I spoke to agreed that this was the best technique which we were taught. I knew constrained creativity was a thing but I think using this technique really drove the point home. I don’t think this is quite the same thing as traditional divination (e.g. you can repeat this a few times and then choose your best idea) but I wonder if it is relying on similar principles.

Bucky 6 Jun 2019 7:30 UTC
2 points

in reply to: Raemon's comment on: FB/Discord Style Reacts
“I especially like/benefited from this bit:
Quote from post/comment”

Bucky 5 Jun 2019 21:13 UTC
3 points

in reply to: Richard_Kennaway's comment on: How is Solomonoff induction calculated in practice?
Well that explains why I was struggling to find anything online!
Thanks for the link, I’ve been going through some of the techniques.
Using AIC the penalty for each additional parameter is a factor of e. For BIC the equivalent is $\sqrt{n}$ so the more samples the more penalised a complex model is. For large n the models diverge—are there principled methods for choosing which regularisation to use?

Bucky

[Question] What are prin­ci­pled ways for pe­nal­is­ing com­plex­ity in prac­tice?

So­ciome­ter theory

Hierom­e­ter theory

Findings

[Question] What are principled ways for penalising complexity in practice?

Sociometer theory

Hierometer theory