Former Community – We Moved

🆕 Coached community here, for posts & feedback public community here.

ForumResources: Researches, Studies, TheoriesRandomAre we self-development nerds?

Please or Register to create posts and topics.

Are we self-development nerds?

#21 · July 4, 2022, 11:28 am

Quote from Lucio Buffalmano on July 4, 2022, 10:36 am
How Would It Look Like? Let's Do An Example
Do you have an example in mind of this would look like?
I'll go first and see if I understand:
Power Move: If someone pulls a power move
Response: And one goes meta on him...
Caveat: Provided that he does it well and "wins"...
Result 1 (on the attacker): The power mover will be shamed, as evidenced by either:
apologizing
self-soothing body language
submissive body language
Result 2 (on the bystander): the people around will side with the defender who effectively went meta, as evidenced by:
Clapping
Nodding
Laughter
Signs of approval such as "wow", "ouch", "that was brutal"
Result 3 (on the defender): the defender will gain status and power, as evidenced by:
proud body language
more speaking time
less interruptions going forward
decreased attacks or hostility from the now disempowered attacker
Something like that?
What do you think.

Quoting myself here with another idea:

Testing Techniques One At A Time

The approach would be to pick a simple technique and open a "testing thread".

People will test it out in real life, and report the results.

After we reach, say, 10 iterations, we can have a rough idea of how well it works (BTW, albeit not this structured, the forum also provided and provides plenty of these tests and data points).

Example:

The "that was a weird / not so friendly thing to say" response technique.

We open a thread for that, we try it out in real life for anything we feel is disempowering or a power move, and we report the results.

The YouTube approach and this one can also go hand in hand.

B and leaderoffun have reacted to this post.

Community, new content and Confidence University now available here.

#22 · July 5, 2022, 2:00 pm

Quote from leaderoffun on July 5, 2022, 2:00 pm
Addressing Kavalier's points in more detail here...

Argument 1: There are entire databases and catalogs of openings, and we don't have that for social behavior
This is a good time to think about the difference between theory and data.
In language, 'theory' is understanding the rules of the language (grammar), and the semantics (meaning of the words). This is very complex and very hard to measure, so we humans came up with a field, linguistics.
[The unreasonable effectiveness of data](https://research.google.com/pubs/archive/35179.pdf)
Peter Norvig says 'every time we fire a linguist, performance increases.'
That paper is interesting because in science we make progress by coming up with hypotheses and testing them. If you take what they say there to the ultimate conclusion: Hypotheses are not needed. Theory is not needed. Patterns will emerge from data, and this is a not-guided-by-theory activity.
I can see 3 levels.
1. Pure theory. [Unfalsifiable](https://www.merriam-webster.com/dictionary/unfalsifiable). For example, Psychoanalisis claims that children want to kill their father and 'own' (including have sex) access to mom. Philosophy tends to be unfalsifiable, this is why it's not science (not saying it's not useful; just that it's not falsifiable). Some aspects of science are barely falsifiable, for example parts of quantum theory.
1. Theory that is verified by empirical tests (data). This is what most science strives for; including social sciences. You formulate a hypothesis (theory), then you set up an experiment to collect empirical evidence. If it proves your theory, then you believe it more.
1. Pure data. This is what we have right now with machine learning models. There's no theory, the model is a black box. Or if you want to push the angle 'impossible! there must be some theory!'... then I'd say the theory is on what you use as input. If you believe body language is important in social interactions, then you must use photo/video as input, not only text. Because there's no theory (nothing in your black box is legible; it's just a bunch of weights in a neural network), you don't know why it works. But it does. For example, a neural net can detect dogs (or 10k categories of objects for that matter) with high accuracy, but there's nothing in the network that says: "my theory is that a dog is a living animal with two ears, 4 legs, a tail, and it barks". The equivalent for power dynamics would be a model that had no input from humans whatsoever (that is, it didn't read PU :) ) but could predict the results of social interactions really well. This sounds like sci-fi right now, but every year that elapses, less so.
If we look at the contents of PU as the theory of human relationships, then are at level 2, there's some empirical evidence in forum posts that report 'I tried this, it worked', in youtube videos with exercises in PU and commentary on the YT channel, etc. But it's still very early days. PU is very much falsifiable, and that's a good thing! And my gut feeling is that it will do really well when we try to falsify it. But as an empiricist I understand this is a gut feeling only.
For an individual, looking at those databases might not be the best investment of time. But dismissing it as 'old knowledge' that is not useful doesn't help: machines with no hypotheses, with no preconceptions about what constitutes a good move (what we would match to a rule in TPM) are doing well; exceedingly well indeed in games like Go and Chess for example.
Machines can do well in fields where behavior can be systematized and measured well. We hope that machines cannot do so well in social dynamics because 'Gosh, it's so hard to measure!' that takes me to your next point:
Argument 2: It's easy to quantify chess. Human behaviour, though... difficult and dangerous
Ah, this has been the struggle of psychology for the last century. What has changed? Now we have more data than ever before about mostly every human.
There's two components to your argument.
It's Difficult: yes but less so every year thanks to machine learning. this is what I'll try to argue for in this post.
It's Dangerous: this is the moral part of AI; I think we can continue the reasoning without having to deal with this as it will derrail the conversation.
In order to measure human interactions with nearly the same accuracy and effectiveness you can get from a chess game, we need huge amounts of data – data that is difficult to collect (it requires constant surveillance), to analyse, that people are often usually unwilling to give (there is a fringe, but growing number of people who resist smartphones, use alternative web browsers and search engines, legislative bodies are creating law barriers, technological race may go in a direction that preludes monopoly of data), and even if you are successful, this success come with dangerous consequences: that would be the wet dream of totalitarian dictatorships, after all.
This is happening already. Bigtech has ridiculous amounts of data, and govs too (at least some, like China). Privacy is a myth. Even the most privacy conscious people I know have somewhat given up on the hope of being off the databases of Bigtech companies.
In order to measure human interactions with nearly the same accuracy and effectiveness you can get from a chess game
Visual stimulus
See for example this video: https://youtu.be/Yqb7zpgWPPA?t=6626 The machine is looking at the pose and correcting mistakes via voice (like a yoga teacher would do)
Anything that has to do with body language is very much measurable and countable now (not in the future, but now, with existing pose detection tech).
Would you like to have a count of:
- Every shrug in youtube
- Every yawn in youtube
- Every time someone looks up and to the right with their eyes in youtube
- Every hand shake (of every kind) in youtube
- Everytime your interlocutor keeps eye contact and smiles in youtube
- Everytime your interlocutor nods in youtube
- Every time someone taps someone else in the shoulder in youtube
- etc
All of this is doable now.
Audio and text
The field of natural language processing (NLP) has made unprecedented progress in the last 5 years.
Example: this a [visual question answering system that can answer questions about anything in an image you upload](https://youtu.be/Yqb7zpgWPPA?t=1973):
Would you like to have a count of:
- Everytime someone said something pleasant to another person in youtube
- Everytime someone said something nasty to another person in youtube
- Everytime someone name dropped in youtube
- Everytime someone repeated a sentence from their interlocutor in youtube
- Everytime someone made reference to physical appearance of their interlocutor in youtube
- Everytime someone made **backhanded compliment** in youtube.
- Everytime someone congratulated someone else in youtube
- etc
Then you can combine modalities: visual and text
Would you like to have a count of:
- every sentence uttered in youtube that produced a frown in the interlocutor
- every physical action that made the interlocutor to say 'you are amazing' in youtube
- every time someone said 'I'm sorry and the interlocutor turned his back
- every time someone tried to shame someone else by pointing at her belly in youtube and saying anything referencing weight
- etc
I'm not putting too much effort into finding the exact situations to count that would make sense for us interested in social dynamics. I'm sure you can come up with more/better bullet points!
Because machines don't get tired of counting, they could go through ALL material in youtube and 'count' for you. You only have to define the input with a few examples. You don't even need a full theory, just a bunch of examples.
Of course there are limits to this and these machines _as they are today_ will fail at getting the most subtle expressions, like whether there's chemistry between a couple of actors, or whether someone was lying when they said something. At the pace they are improving? I wouldn't be so sure the limits stand in 5 years time!
Effective moralizing, social scalping, judge role... these may be too subtle to count today. But it all depends on whether we can find good examples.
If counting is possible, why don't we do it?
This is within reach only for a bigTech company like google because compute time will be extreme. And that's only one part of the problem: downloading all of youtube is probably illegal, on top of very very data intensive. I imagine this project would run in the 100s of millions at least.
So my argument is that we are quite close to having a database of human behavior similar to the body of recorded games of chess. That chess is far simpler and can be represented with a succinct text code of moves was an advantage but we re slowly getting there for any behavior to be captured and counted.
There was a time where searching the entire catalog of music from all times was sci-fi. Now we have that. It feels sci-fi that we can search for video of people shrugging and get exactly that. But I do think we will see that in the next few years.
How we do social science will change. And perhaps there will be less need of theory.
Argument 3: In chess you have this gigantic database, but it's not that useful to master chess. Hence if we had the same database for social interactions it might not be very useful
"Today chess is a game that has been measured and remeasured from every possible angle"
(...)
one would still be advised to ignore them and use one's valuable time to study the games of the masters".
^^ But one can write a program that consumes all that information and produces actions (or recommendations) that win more often than a human that spent his time not looking at that data.
So my gut feeling (that cannot be validated :) ) is that an algorithm database for social interactions as comprehensive as the one we have today for chess would kick ass. Any human being advised by such algorithm will do well.

Addressing Kavalier's points in more detail here...

Argument 1: There are entire databases and catalogs of openings, and we don't have that for social behavior

This is a good time to think about the difference between theory and data.

In language, 'theory' is understanding the rules of the language (grammar), and the semantics (meaning of the words). This is very complex and very hard to measure, so we humans came up with a field, linguistics.

[The unreasonable effectiveness of data](https://research.google.com/pubs/archive/35179.pdf)
Peter Norvig says 'every time we fire a linguist, performance increases.'

That paper is interesting because in science we make progress by coming up with hypotheses and testing them. If you take what they say there to the ultimate conclusion: Hypotheses are not needed. Theory is not needed. Patterns will emerge from data, and this is a not-guided-by-theory activity.

I can see 3 levels.

1. Pure theory. [Unfalsifiable](https://www.merriam-webster.com/dictionary/unfalsifiable). For example, Psychoanalisis claims that children want to kill their father and 'own' (including have sex) access to mom. Philosophy tends to be unfalsifiable, this is why it's not science (not saying it's not useful; just that it's not falsifiable). Some aspects of science are barely falsifiable, for example parts of quantum theory.

1. Theory that is verified by empirical tests (data). This is what most science strives for; including social sciences. You formulate a hypothesis (theory), then you set up an experiment to collect empirical evidence. If it proves your theory, then you believe it more.

1. Pure data. This is what we have right now with machine learning models. There's no theory, the model is a black box. Or if you want to push the angle 'impossible! there must be some theory!'... then I'd say the theory is on what you use as input. If you believe body language is important in social interactions, then you must use photo/video as input, not only text. Because there's no theory (nothing in your black box is legible; it's just a bunch of weights in a neural network), you don't know why it works. But it does. For example, a neural net can detect dogs (or 10k categories of objects for that matter) with high accuracy, but there's nothing in the network that says: "my theory is that a dog is a living animal with two ears, 4 legs, a tail, and it barks". The equivalent for power dynamics would be a model that had no input from humans whatsoever (that is, it didn't read PU :) ) but could predict the results of social interactions really well. This sounds like sci-fi right now, but every year that elapses, less so.

If we look at the contents of PU as the theory of human relationships, then are at level 2, there's some empirical evidence in forum posts that report 'I tried this, it worked', in youtube videos with exercises in PU and commentary on the YT channel, etc. But it's still very early days. PU is very much falsifiable, and that's a good thing! And my gut feeling is that it will do really well when we try to falsify it. But as an empiricist I understand this is a gut feeling only.

For an individual, looking at those databases might not be the best investment of time. But dismissing it as 'old knowledge' that is not useful doesn't help: machines with no hypotheses, with no preconceptions about what constitutes a good move (what we would match to a rule in TPM) are doing well; exceedingly well indeed in games like Go and Chess for example.

Machines can do well in fields where behavior can be systematized and measured well. We hope that machines cannot do so well in social dynamics because 'Gosh, it's so hard to measure!' that takes me to your next point:

Argument 2: It's easy to quantify chess. Human behaviour, though... difficult and dangerous

Ah, this has been the struggle of psychology for the last century. What has changed? Now we have more data than ever before about mostly every human.

There's two components to your argument.
It's Difficult: yes but less so every year thanks to machine learning. this is what I'll try to argue for in this post.
It's Dangerous: this is the moral part of AI; I think we can continue the reasoning without having to deal with this as it will derrail the conversation.

In order to measure human interactions with nearly the same accuracy and effectiveness you can get from a chess game, we need huge amounts of data – data that is difficult to collect (it requires constant surveillance), to analyse, that people are often usually unwilling to give (there is a fringe, but growing number of people who resist smartphones, use alternative web browsers and search engines, legislative bodies are creating law barriers, technological race may go in a direction that preludes monopoly of data), and even if you are successful, this success come with dangerous consequences: that would be the wet dream of totalitarian dictatorships, after all.

This is happening already. Bigtech has ridiculous amounts of data, and govs too (at least some, like China). Privacy is a myth. Even the most privacy conscious people I know have somewhat given up on the hope of being off the databases of Bigtech companies.

In order to measure human interactions with nearly the same accuracy and effectiveness you can get from a chess game

Visual stimulus

See for example this video: https://youtu.be/Yqb7zpgWPPA?t=6626 The machine is looking at the pose and correcting mistakes via voice (like a yoga teacher would do)

Anything that has to do with body language is very much measurable and countable now (not in the future, but now, with existing pose detection tech).

Would you like to have a count of:
- Every shrug in youtube
- Every yawn in youtube
- Every time someone looks up and to the right with their eyes in youtube
- Every hand shake (of every kind) in youtube
- Everytime your interlocutor keeps eye contact and smiles in youtube
- Everytime your interlocutor nods in youtube
- Every time someone taps someone else in the shoulder in youtube
- etc

All of this is doable now.

Audio and text

The field of natural language processing (NLP) has made unprecedented progress in the last 5 years.
Example: this a [visual question answering system that can answer questions about anything in an image you upload](https://youtu.be/Yqb7zpgWPPA?t=1973):

Would you like to have a count of:

- Everytime someone said something pleasant to another person in youtube
- Everytime someone said something nasty to another person in youtube
- Everytime someone name dropped in youtube
- Everytime someone repeated a sentence from their interlocutor in youtube
- Everytime someone made reference to physical appearance of their interlocutor in youtube
- Everytime someone made **backhanded compliment** in youtube.
- Everytime someone congratulated someone else in youtube
- etc

Then you can combine modalities: visual and text

Would you like to have a count of:

- every sentence uttered in youtube that produced a frown in the interlocutor
- every physical action that made the interlocutor to say 'you are amazing' in youtube
- every time someone said 'I'm sorry and the interlocutor turned his back
- every time someone tried to shame someone else by pointing at her belly in youtube and saying anything referencing weight
- etc

I'm not putting too much effort into finding the exact situations to count that would make sense for us interested in social dynamics. I'm sure you can come up with more/better bullet points!

Because machines don't get tired of counting, they could go through ALL material in youtube and 'count' for you. You only have to define the input with a few examples. You don't even need a full theory, just a bunch of examples.

Of course there are limits to this and these machines _as they are today_ will fail at getting the most subtle expressions, like whether there's chemistry between a couple of actors, or whether someone was lying when they said something. At the pace they are improving? I wouldn't be so sure the limits stand in 5 years time!

Effective moralizing, social scalping, judge role... these may be too subtle to count today. But it all depends on whether we can find good examples.

If counting is possible, why don't we do it?

This is within reach only for a bigTech company like google because compute time will be extreme. And that's only one part of the problem: downloading all of youtube is probably illegal, on top of very very data intensive. I imagine this project would run in the 100s of millions at least.

So my argument is that we are quite close to having a database of human behavior similar to the body of recorded games of chess. That chess is far simpler and can be represented with a succinct text code of moves was an advantage but we re slowly getting there for any behavior to be captured and counted.

There was a time where searching the entire catalog of music from all times was sci-fi. Now we have that. It feels sci-fi that we can search for video of people shrugging and get exactly that. But I do think we will see that in the next few years.

How we do social science will change. And perhaps there will be less need of theory.

Argument 3: In chess you have this gigantic database, but it's not that useful to master chess. Hence if we had the same database for social interactions it might not be very useful

"Today chess is a game that has been measured and remeasured from every possible angle"
(...)
one would still be advised to ignore them and use one's valuable time to study the games of the masters".

^^ But one can write a program that consumes all that information and produces actions (or recommendations) that win more often than a human that spent his time not looking at that data.

So my gut feeling (that cannot be validated :) ) is that an algorithm database for social interactions as comprehensive as the one we have today for chess would kick ass. Any human being advised by such algorithm will do well.