Everyone knows actions are fundamentally different than predictions, but it's hard to write this distinction in math.

No matter what the models say, one person's statistical fact is another's statistical outlier. (Part 7 in the going-for-2-down-8 microcosm)

A machine learning view of the randomized controlled trial, as a bridge from patterns to actions.

alright... so... joe flacco throws a pick 6 it gets called back on BS grounds bengals score TD to go down 8 go for 2 and make it get the onside kick on the most sketchy foot contact get another TD to go up 1 forget to tackle the bears tight end who scores flacco throws a pick bengals lose

oh my

OK, I have an analytics post drafted, but there is still plenty of time left on the clock for Joe Flacco.

Ah, it's becoming even clearer! Probabilities change if you know the future.

How two mathematicians resolved a 50-year-old open problem by finding the solution in an 80-year-old paper. Human knowledge is unsearchable.

The Jets went for 2 down 8 and won! This changes everything.

This paper is a delight along multiple axes. It describes an Erdős problem that was solved 30 years before Erdős declared it open. arxiv.org/abs/2510.19804

It is maddeningly frustrating to try to explain the incoherent assemblage of heuristics that is reinforcement learning.

A great rant by hater/legend @drewmagary.bsky.social about why sports is the best when it's improbable. Also, appreciate the shoutout to my anti-analytics analytics.

Revisiting last week’s open problems scandal, I wrote about LLMs as Lore Laundering Machines and why some are blind to the novelty whitewashing.

Conformal prediction in a tweet: predict an empirical quantile instead of an empirical mean. Trustworthy AI, here we come.

The list of signatories on the latest "Please Skynet, don't kill us" letter is BONKERS.

I am unironically writing a book about this.

Randomization is a powerful algorithmic tool, but its value strongly depends on context. Just compare quicksort to confidence intervals.

PS The title of today’s post is inspired by this must-read classic by Stark and Freedman.

Win probability calculators are one of the more absurd artifacts of our obsession with data. I look into how they are made and who they are for.

Now that's what I call probability!

Speaking of football analytics, Sean Payton just had the fun sequence Go for 2 down 13 Go for 2 down 12 Go for 1 down 4. He's trolling the spreadsheet nerds.

This article in the WSJ claims that in college football last season, teams went for it on 4th-and-1 from their own 34-yard line 57% of the time. I find that unbelievable and would love to see the actual list of plays. For comparison, in the NFL, teams only did this 2 out of 9 times in 2024.

Do AI chatbots excel at text extraction because these companies depend on document scanning for training data?

Sam raises an important and uncomfortable question about the sociology of open problems in mathematics.

I'll never understand why people willingly attach their names to quack pseudoscience like this. We should run some psychometrics on the coauthor list.

If anyone wants a counterargument against AGI, look at this three-year chatGPT regression and its persistent denial of facts plain as day.

Riffing on Ed Leamer’s “sinning in the basement,” I ponder a theoretical reformation in machine learning. (There are ImageNet scatterplots, IYKYK.)

The replies here have brought out the best and the worst of bluesky. Fascinating.

Going for 2 down 8 has led to exactly 2 wins in the history of the NFL. Hear me out: what if simple probability models lead to bad decision-making in professional football?

Every class on machine learning should study the datasets upon which the field was built. They provide more insights than any probability bound.

Update: the Bengals never got the ball back and lost by 9.

The Bengals just went for 2 down 8. Now I need to tune in to this garbage game to see how it wraps up.

Amazingly, multiple NFL games have ended with the beautiful score 11-10.

The new Jay Som record is so good. jaysom.bandcamp.com/album/belong

Almost a decade ago, I coauthored a paper asking us to rethink our theory of generalization in machine learning. Today, I’m fine putting the theory back on the shelf.

Kevin successfully shamed me into closing all of my tabs. And for that I am thankful.

OpenAI’s new recursive nostalgia ads corroborate @himself.bsky.social's theory of Philip K Dick and AI.

Optimization theory has a few important lessons for machine learning, but the list of actionable lessons is surprisingly short.

Wrote about football again, amused by cases where even the data nerds concede you might have to consider the messiness of human psychology.

A pox on early meetings messing up my ritualized posting schedule.

In non-LLM AI news, Waymo has quietly been involved in two fatal car crashes in 2025. Phil Koopman details that this fatality rate is now comparable to humans. People: 7/billion miles in all conditions Waymo: 5/billion miles in geofenced conditions

Tying together many previous threads on the blog, I respond to Ben Chugg’s critique of “A Bureaucratic Theory of Statistics.”

In machine learning, do you need to know any optimization algorithm other than stochastic gradient descent? A reluctant but best-faith argument for no.

Really? A fine-tuner for Llama 3? You don't have to be an AI doomer to think Murati has pulled off one of the most impressive cash grabs in recent Silicon Valley history.

I wrote what are likely some very outdated thoughts on approximation theory. All of this was known in 1990. Which new perspectives am I missing?

This looks like another football post, but it's actually about the vulgar arrogance of the vulgar positivism of decision theory.

Not Mina Kimes, too!

Greg Olsen is so committed to the decision-theory bit: "Just because the outcome is good, doesn't mean the decision was correct." Anyway, Carson Wentz.

Without any theory, machine learning has evolved a universal theory of data representation. Can we post-hoc rationalize it with math?