Tomer Ullman

time for some escapist reading on the legal system in imperial China well damn

Jennifer Hu Personal website of Jennifer Hu, cognitive science and AI researcher.

major kudos to the co-authors, especially lead author Jennifer Hu, who just started as a professor in Johns Hopkins and would be an ideal person to work with: jennhu.github.io

Type theory in human-like learning and inference Humans can generate reasonable answers to novel queries (Schulz, 2012): if I asked you what kind of food you want to eat for lunch, you would respond with a food, not a time. The thought that one woul...

(some of the ideas build on the a suggestion that can be summarized as "the language of thought is typed", see more here:

arxiv.org/abs/2210.01634)

The General Discussion also opens with this, in case it convinces you to read it:

the General Discussion gets into several limitations and new directions, including going back to the idea of 'type coercion', in which a compiler that encounters a type error can attempt to fix it

we also go into questions regarding development, it's fun, you should read it

we created a novel set of stimuli with a hierarchy of types violations, and had people rate the likelihood of this nonsense.

we found:

(4) people rank 'far'-type violations as more nonsensical than 'near', though both are nonsense; in line with the 'severity of violation' idea

let's go back to the 'hitting an error in the world model' idea, just like how in computer programs you can also hit 'type errors'

perhaps this is about the 'severity' of the type violation?

So, we got ppl to rate how a good a metaphor a nonsense string is/

We found:

(3) People DO agree on whether something sounds like a good metaphor!

...But that's not correlated with likelihood

(it'd be an interesting thread to chase down, what makes a good metaphor, in the absence of mapping)

(aside, we had the following exchange when looking at the results:

A: "Filling your purse with luck does sound like a metaphor"

B: "Yeah...but a metaphor for what?"

A: ¯\_(ツ)_/¯

maybe this is about 'metaphorization'?

That is, "filling your purse with luck" is literally nonsense, but it sounds like a metaphor, more so than 'hanging your coat on a sneeze'. So maybe that?

so people have a graded sense for some inconceivable events.

what is that based on?

perhaps a simple notion of 'more linguistically likely'?

to make a long story short we find that

(2) No.

people's judgements of the inconceivable are not correlated with string probability

given all this, you might expect people to have categorical judgements about nonsense. Do they?

We find that

(1) People have *graded* judgements about the inconceivable. Some things are (consistently) more nonsensical than others.

(I do love the stimuli)

one way to think about *impossibilities* is that they still refer to some kind of world model, and you can reason about the probabilities of it.

But *inconceivabilities* are more like an 'error' in the instantiation of the world model.

More Background:

people have a sense of *graded impossibility*: Some things are more impossible than others (levitating a brush is easier than a boulder)

researchers find that interesting, because the way people structure the imagination can tell us something about how people structure reality

(aside: some of you are already thinking: “but maybe '5' refers to some strange tool, or a printed-out statue that looks like a 5, or..."

yeah yeah -- we call those ’coercions’, and note they require a transformation of the sentence to have a new meaning, we get to that later)

Background:

people distinguish:

- probable (painting a house with a brush)
- improbable (painting it with your hair)
- impossible (magically lifting the brush)
- inconceivable (painting it with the number 5)

you can read more about that here:

www.sciencedirect.com/science/arti...

fun pre-print for your start of week reading:

"People Make Graded Judgments About The Inconceivable"

(by Hu, Sosa, and me)

doi.org/10.31234/osf...

you have climbed the mountain to learn Kung Fu from the ancient Master, but you are not worthy

started watching the old Poirot show in the evenings and I think I'm in love

Congress is currently considering a piece of legislation that would effectively sever all scientific ties between the US and China. Unbelievable.

You can read the AAU response here:
www.aau.edu/key-issues/a...

that's survivor bias :) a separate issue, in which you incorrectly generalize from data D without correctly accounting for the data generation process.

to oversimplify, in extension neglect you made claims about a sample S without correctly normalizing by the sample size or extent |S|

"extension neglect" is just the agreed-upon term for a general phenomenon, which shows up in many more specific domains, but mostly taken to be unintentional

en.wikipedia.org/wiki/Extensi...

Your followers might hate you [This is a reposting of something I wrote just when I started Small Potatoes. I’ve substantially edited and revised it, but kept some of the old-timey examples.]

in other words: Paul Bloom is right here smallpotatoes.paulbloom.net/p/your-follo... that social media can be veridical...

BUT for that to be true you need to account for sample size.

to try and off-set a misunderstanding: i'm not saying that getting yelled at is a sign that you're NOT doing something wrong. if you did something in front of 10,000 people and 9,000 of them yelled at you that should give you pause.

to put it another way: the problem isn't social media per se, we show extension neglect everywhere in life. But what can be unique relative to the past is an increase in sheer volume and immediacy of interactions.

it is easy to generalize poorly from there:

* People hate me (a very small subset said something mean)

* People are confused morons! (no, most people get what you are saying, a small sub-set are confused + willing to say something it)

* Ugh, [generalized sub-set of the population] are CRAZY!

...but if you don't account for sample size, you're thinking "jeez, people today are worked up about ice-cream" or "wow, people hate me"

100,000 watching you is a crazy number in real life, but not on social media. if you're a Big Name this could be more like tens of millions of people.

independently, how many people would radically misunderstand what you're doing? 1 in 500? Seems a low-ball.

so, you're going to get at least 2 people screaming at you for going 'this is nice' while eating ice-cream.

If you kept in mind the sample size then you'd brush it off.

like, suppose you did *anything* in front of 100,000 people. Suppose you ate ice-cream and said 'this is nice'

of those 100,000, most would just not care. How many would feel a desire to react? Not many. Say, 1 in 100?

And most of those would be a thumbs up or 'mm, ice-cream'.

my grand theory of social media induced madness is that many of the ills of social media are about extension neglect.

...by which I just mean "people fail to take into account how many people *didn't* react to them"