Ben Recht

credit to alex peysakhovich (who's not on here) for the pith.

"It's called transfer learning if it works and overfitting if it doesn't."

Is it?

What is overfitting?

But why does the cash cow continue to produce?

Those two statements are equally depressing but describe slightly different things, no? The first is about early adoption, the latter about late majority.

There are probably sleaze theorems for all parts of the adoption curve.

Important ETA: Bullet 3 should read: "AI cap ex now growing more than all consumer spending combined."

bah, left out an important word "growing." should have been

"AI cap ex now growing more than all consumer spending combined"

(figure from here: bsky.app/profile/mims...)

what's the sleaze theorem?

"May you live in interesting times?"

These three things seem deeply interconnected:

1) app gambling culture.
2) the only industry adding jobs is health care.
3) AI cap ex now more than all consumer spending combined.

I need to write a ted talk about how.

The American economy now stands upon the abandonment of clarity and efficiency in software engineering.

Do you think it would be helpful for them to explain the uncertainty in the estimates?

At least we can be sure no one will drown.

If you have infinite data and compute, 1-nearest neighbors is also optimal.

Yes, and crazy that the original pac-man didn't even have randomness!

You already run into a problem when you consider all logical tautologies must be assigned probability 1 which requires knowing what most consider unknowable.

sorry, to clarify: if you read RL people, they come up with all sorts of neurobullshit explanations for the discount factor.

But the reason the discount factor is 100% a crutch to make the formulas nicer.

In RL it’s a crutch so you get a pretty Bellman equation with a time-invariant policy.

Agree. I should write something... but in lieu of that, read the fun history of arxiv I posted above.

Inside arXiv—the Most Transformative Platform in All of Science Modern science wouldn’t exist without the online research repository known as arXiv. Three decades in, its creator still can’t let it go.

This isn't exactly what you are asking for, but allow me to recommend an amazing article: www.wired.com/story/inside...

While this is true, I could not disagree more with @tedunderwood.me's credit assignment.

Physics was at least two decades ahead of CS with all of these things. And the same is true with math and EE.

CS's main achievement was making conference abstracts into CV beans to be counted.

Hear hear! I wish I could add 100 more likes.

I appreciate your commitment to this message.

Hey wait...

.... nevermind.

I know a guy who bought in on this cult and is one of the people chosen in the Zuckerberg lottery. Like in all cults, it has to appear to work out for some...

So much of the Bluesky vitriol against Graeber cites the time he was a dick in the Crooked Timber comments section...

I agree with this, but I also think most of the people talking shit about Graeber and Scott on this website suffer from the Bluesky willful misreading disease you often post about.

I need to see those frequencies quantified with error bars before I can agree with you...

I must know what the "two irrelevant traditional works" are.

What does bad data mean to you?

I tend to think that any two variables are correlated, and it's just a question of how much.