eeeey MCARena

Hero worship in science is rotten. There are no giants on whose shoulders we stand. We're all part of a community that spans time and geography. Getting fixated on origin stories is very Marvel. Science doesn't happen in heroland though.

On par for Microsoft products in my experience.

bsky.app/profile/kons...

It's 1637. The catholic peasants revolt.

Shogun: We hebben een serieus probleem.

This is the original definition of scooping.

Or rather, their metrics told them the solution with 10 clusters was best. They were like: nah, that's too many, and chose the second best option where a few of the smaller clusters had been merged into the larger ones.

I recently reviewed a paper about life course analysis. There were 10 possible occupational classes. The authors had clustered the life course time series into, you guessed it, 10 clusters. The clusters were basically based on the most common occupational class.

What's the take home message?

If you're submitting AI slop you're a loser. You're just making these great free services harder to run, and making it more difficult to separate signal (science) from noise (your crappy AI shit.)

I'm a busy academic, I don't have time to read superfluous words!

The only criticism I have is that sometimes it'd be nice to enjoy modern art that doesn't turn into a commentary about the art form itself. Fortunately that was a relatively minor part of the 90 minute whole.

Thanks, good to know. I guess I'll look at the original vglms first then.

(That game rules)

For diversity's sake, may I suggest you also play a Canadian game about climbing a mountain (Celeste)?

And a quick question for @tslumley.bsky.social : svy_vglm doesn't seem to have a predict method (although using predict(model_object$fit) seems to work). Does this mean predictions for these models aren't valid?

I see. vgam/vglm models do seem to have these methods, so is there some problem in the way they return the values, i.e. are they in some way nonstandard? And thanks for the link. Extending the package doesn't look too hard, so maybe I'll give it a go!

It's stated that vglm models are problematic _because_ they support categorical/multinomial responses. Does this mean that marginaleffects can't support models from svyVGAM? Is there something I could do to help make it happen? (Not that I have experience in package development, but I'd like to!)

That would rock because svy_vglm() can handle multinomial responses, and that response distribution happens to describe a lot of survey data well. However, the marginaleffects github repo issue for supporting new models has a cryptic line about vgam/vglm models.

github.com/vincentarelb...

Given the times we live in, that is.

Seriously? That sounds ludicrous, so on second thought it must be true.