Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Awaclus]

Does any of that imply that interpretability tools are impossible or not useful?

I don't think so, but it does imply that the usefulness of possible interpretability tools has limits.

[silverspawn]

=IF(VLOOKUP(D32;$A$32:$B$37;2;FALSE)>VLOOKUP(F32;$A$32:$B$37;2;FALSE);"mehr";IF(VLOOKUP(D32;$A$32:$B$37;2;FALSE)=VLOOKUP(F32;$A$32:$B$37;2;FALSE);"Gleich";"weniger"))

[silverspawn]

if only more than one person had been able to do this

[silverspawn]

One day in July, Amalia went out and caught a piece of summer in a pitcher,

and she placed it in the cellar on her shelf, where it collected dust.

Then one day in winter she poured it into the snow, and her garden rose.

The next day, on a cold spring night, she went to a cave to visit the oracle of Delphi

 and autumn wore heavily

and the treetops swayed

as Amalia stepped into the cave

the oracle led her below her garden

where the pitcher lay broken

and she became the seasons

and stayed in the cave forever

[silverspawn]

text for another exam, written by me & Kayo Dot & GPT-2.

[silverspawn]

well except most of it is in German

[silverspawn]

from waitbutwhy:



I'm very annoyed by this graph because it's so grossly inaccurate. The opposite is empirically true. If a person has one virtuous trait, it's evidence that they're optimizing for something, and hence they're more likely to be great in other areas. The reverse effect is much weaker (a bad trait is only weak evidence for other bad traits) because even good people can have blindspots, but it still goes into the other direction.

It's not a flipping coincidence that the founder of the Against Malaria Foundation is famous for replying to emails super quickly. Both "run a charity which does a single thing and has an ultra small bureaucracy (it's essentially still run by one person in a basement even though it now has considerable budget)" and "make it a major life's goal to reply to emails quickly" are weird things to do, and normal people don't do them. Obviously these are positively correlated, not negatively.

[silverspawn]

Maybe another reason I'm annoyed is that the graph in some sense claims you can't be a genuinely great person period

[silverspawn]

Probably a backroom not a basement though

[silverspawn]

and I get even more annoyed looking at it further because the several lines       imply different examples, and none of them go into the right direction. Really Tim Urban, you didn't meet a single person where you observed the more plausible outcome?

[silverspawn]

And here is a big meta analysis of peer reviewed scientific studies showing that homeopathy works

[silverspawn]

By the way my sister got covid. Unfortunately, the median outcome is probably that it doesn't affect her much and confirms her belief that the disease isn't a big deal.

[silverspawn]

Curious if it's delta or omicron

[faust]

And here is a big meta analysis of peer reviewed scientific studies showing that homeopathy works

My immediate reaction is that this is publication bias in action; i.e. a study with the positive finding that homeopathy works is much more likely to get accepted than a study with a negative finding.

My second reaction is wondering if my first reaction is all that different from an anti-vaxxer's reaction to being shown research that contradict his position.

[faust]

By the way my sister got covid. Unfortunately, the median outcome is probably that it doesn't affect her much and confirms her belief that the disease isn't a big deal.

Would you prefer your sister suffers and changes her view than that she doesn't suffer and sticks to it?

This it not meant to be rhetorical, I'm actually curious.

[silverspawn]

By the way my sister got covid. Unfortunately, the median outcome is probably that it doesn't affect her much and confirms her belief that the disease isn't a big deal.

Would you prefer your sister suffers and changes her view than that she doesn't suffer and sticks to it?

This it not meant to be rhetorical, I'm actually curious.

I would, as long as we're talking about something like "feel really shitty for three days and tired for a week but then be fine". I think the false belief is causing more suffering than that (primarily for my mother), even just considering the direct consequences. And I bet having a bogus world view will come with more, unforeseen downsides down the line.

[silverspawn]

My second reaction is wondering if my first reaction is all that different from an anti-vaxxer's reaction to being shown research that contradict his position.

This is why, when I made a reaction video to the pro vax video a few months ago, I showed almost no evidence for my position and only focused on debunking the pro-vax evidence

I mostly see it as a sign that medical research is in bad shape

[silverspawn]

*debunking the vax-is-harmful evidence

[silverspawn]

One day in July, Amalia went out and caught a piece of summer in a pitcher,

and she placed it in the cellar on her shelf, where it collected dust.

Then one day in winter she poured it into the snow, and her garden rose.

The next day, on a cold spring night, she went to a cave to visit the oracle of Delphi

 and autumn wore heavily

and the treetops swayed

as Amalia stepped into the cave

the oracle led her below her garden

where the pitcher lay broken

and she became the seasons

and stayed in the cave forever

So the following happened:

I do the exam. The story is split across 5 different HTML files the students ought to create. Four of them mentioned her name.

One student handed in a solution where he spelled the name like this

"Amelia" in 01
"Amilia" in 02
"Amalia" in 04
"Amailia" in 05

Files 02, 04, and 05 came with the text already written (don't want to penalize anyone for being a slow typer), and I just checked and I misspelled it in 05 but not in 02 and 04. Still unsure what happened there, but I find it really funny in any case

[silverspawn]

For the penultimate exam, I let GPT-2 (no longer have access to GPT-3) do almost all the writing

Once upon a time, there was a spaceship. On board was a princess, a knight, a rogue, and a dog. The princess had a single purple wing. The knight had a wooden leg. (Found this really funny idk why.) They were all together in a large, round room with a trapdoor. The trapdoor opened and an ugly green monster stepped outside. It had several claws like a troll. "Princess" it said, "I've been sent to take you". It took the princess and the knight and the dog to the great statue named Goliath. But goliath was asleep right now.



Supposedly there are people who legitimately use GPT-3 for serious fiction. I have not had success with that. Maybe I've done it wrong. It definitely doesn't work for dialogue scenes.

[silverspawn]

For this latest exam, the bonus question was to create a small file with a list. The difficulty here is that lists weren't done in class, so they had to google.

The student whose submission I'm scoring right now didn't google how to do lists. Instead, she must have googled how to insert the • character instead and copied that into the file. Doesn't look quite right, but still, not bad.

[silverspawn]

<p>flog </flog>

Lol

[silverspawn]

So many people in this class just make up weird stuff, e.g. in 04 you were given a html file and had to correct errors and some just added things and I'm like ???? why

My main takeaway is that having too much space between classes is bad. It's biweekly to begin with (not good!) and then  with this particular class, several times, it was cancelled. They seem to have no idea what they're doing to a much greater extent than others.

Instead of having classes biweekly, they should be weekly in 3 month blocks or something

[silverspawn]

The class I have weekly is the one I'm most invested in and where I got the 16.5 average. Granted, one of the biweekly ones has an even better average, but well there are just a lot of good students in there. Also it was never cancelled.

[silverspawn]

stuff like this is all over the place

margin: 60px0px;

aside from the obvious, there isn't even margin here. Left it out on purpose because I wanted to make it easier