Scythia Marrow

Illo of a young lady taking photographs of tiny floating whales. Drawn by gdbeeart, Sept 2025

polychrome parade

If a university’s choice is between dying and becoming subject to the whims of an apparatchik the university must die.

Texas A&M should lose accreditation. Journals should not publish articles from researchers affiliated with Texas A&M. Conferences should not invite nor admit Texas A&M researchers.

I hope everyone will forgive the bluntness of my language, but it seems like bluntness is warranted: it now seems pretty clear that the president raped children or at least knew about and actively facilitated the rape of children.

I think it is time we stopped using coded language to say that.

x2 on PGTE, it is absolutely incredible. I honestly think that EE is the best fantasy writer of all time, Pale Lights is impeccably good.

Excellent newsletter!

Your only claim I would push back on is your claim that the science behind LLMs is exhausted. The amount of investment is actually leading to a brisk pace of advancement; the model of next year will be able to burn twice as many tokens before spiraling into delusion. Progress!

The curse of being 2 months early then 4 months late strikes again.

YoloE is very, very good. Trex2 performance in 140ms on a laptop? Incredible.

#MLsky #ComputerVision

It really clicks when I start using the concept as a whole. Before it clicks I need to think through and write out every piece of the definition while working on a proof, and deliberately plan how to use each piece. After, things appear intuitively ”math concept shaped”.

Worst part about ML nowadays is being 4 months into a hobby project solving a neat problem in a neat way only to discover the VC backed framework for solving said problem in said neat way and also baking cookies was released two months ago.

LangGraph is very very good. I’m not salty about it. Nope.

I don’t think there exists a meaningful definition of word model, reasoning, or symbol that excludes gpt.

GPT has a world model (transforming layers) that computes over symbols (embeddings). That’s reasoning.

Rümeysa Öztürk, smiling with arms crossed, walks through a parking lot.

Rümeysa Öztürk, smiling, greets a reporter on grass surrounding a parking lot.

Video stills of Rümeysa Öztürk being released from ICE custody via NBC 10 Boston

Gonna do every newspaper in America a favor and write their headline. It reads:

TRUMP DISCUSSES FOREIGN CONCENTRATION CAMPS FOR AMERICAN CITIZENS WITH BUKELE

There you go. Don't even need to credit me.

My name is __. My zip code is __. I am calling as your constituent to express alarm at the Abrego Garcia case. Trump is openly defying the Supreme Court by not bringing Garcia back. If you allow Trump to do this, then he is unquestionably a dictator, and your job as my member of Congress is to remove him at once.

(For Reps:) I ask that you immediately introduce Articles of Impeachment in the US House, and block all other business until they are introduced.

(For Senators:) I ask that you block all business in the Senate until Articles of Impeachment are introduced in the US House, and when they pass the House I expect you to vote to impeach. Thank you.

Here's the script, folks. It's now or never.

Capitol switchboard: (202) 224-3121

Local office numbers: www.congress.gov/members/find...

RTs appreciated.

Causal Emergence 2.0: Quantifying emergent complexity Complex systems can be described at myriad different scales, and their causal workings often have multiscale structure (e.g., a computer can be described at the microscale of its hardware circuitry, t...

arxiv.org/abs/2503.13395

Erik Hoel’s recent paper on Causal Emergence is very interesting. He formulates causal emergence as an integral instead of a difference which resolves a lot of lingering issues with the mathematics, and provides a solid foundation for ML theory to work with.

#MLsky #AIsky

Tufts PhD student on visa arrested by immigration authorities, school says Rumeysa Ozturk is a Turkish national, according to her lawyer, Mahsa Khanbabai.

abcnews.go.com/US/tufts-phd...

This is tyranny. ICE needs to get out of Boston; tyrants are not welcome here. The rubicon is crossed and we must respond by defending our people and our freedom.

The only morally acceptable response to this abomination is to throw anything ICE owns into the harbor.

So the admin. arrested a green-card activist on pretext but clearly over speech, revoked his green-card without process and is now holding him God-knows-where, while his citizen wife frets in the dark.

That's...just tyranny. The act of a tyrant.

In America, we used to know what to do with tyrants.

By making these measures computational and informational we can be confident that an algorithm that achieves those goals on arbitrary markov chains isn't just a UFA, it's a computationally optimal UFA.

That's the science. How to do that is speculation at this point, which is a different thread.

The task of us in ML is to do just that: reconstruct state space from variable space in the most computationally efficient way possible by maximizing reversibility (Γ_a in the paper) and causal emergence (ΔΓ) by minimizing redundancy (1/N - 1/r, inverse size of state space).

This is sketched out in the weaknesses section: "Second, discussions have primarily focused on state space, while an approach incorporating variable space would be more practical, despite the exponential growth of state space size with the number of variables, posing a significant challenge."

However, the mathematics behind this qualitative effect has been difficult to pin down. This paper elaborates on a great potential avenue; by relating the two phases to the two components of effective information we can get a true grasp on what each phase is really doing.

For context: the idea of two vague types of optimization is commonplace within the mathematics of ML, it pops up everywhere. From the exploitation / exploration tradeoff to the positive and negative phases of Boltzman machines or Forward Forward, learning seems to be "two-ish".

Dynamical reversibility and a new theory of causal emergence based on SVD - npj Complexity npj Complexity - Dynamical reversibility and a new theory of causal emergence based on SVD

Zhang et al recently published an absolutely stellar paper relating casual emergence and reversibility in Markov processes, building off the work of Erik Hoel. This is a must-read for any ML theorist.

🧵

www.nature.com/articles/s44...

#MLSky #AISky #machinelearning

CVAT downloads broke today. Am I really going to have to format my datasets myself?

Hope it’s fixed by tomorrow >->

Pin this, save it, we’ll need it.

“When a prison camp opens in your town…when a DREAMer is disappeared from your classroom…when the President destroys what’s left of the Constitution…They will all say they didn’t know this was coming. And I want the American people to know that they did.”

AOC 🔥🎯