iiinova — where your digital self keeps evolving

Greg Langmead @greg.langmead.info · 03/10/2026

This is a worthy proposal, as you might expect from Grant. My PhD advisor, John Morgan, spent time writing books that packaged difficult material like Donaldson theory, Seiberg-Witten theory, and the proof of the Poincaré conjecture. These created lots of momentum and understanding.

010

Greg Langmead @greg.langmead.info · 03/10/2026

I specifically meant this article from the transformer-circuits interpretability site: transformer-circuits.pub/2021/framewo...

transformer-circuits.pub

A Mathematical Framework for Transformer Circuits

000

Greg Langmead @greg.langmead.info · 03/10/2026

Maybe it's no surprise that a bunch of computer programmers think it's God-like for a computer to be able to write programs. Yikes, people. Tell me you think you're a God without telling me you think you're a God.

000

Greg Langmead @greg.langmead.info · 03/10/2026

Here's another link which helps with some of the transformer math. We are behind on interpreting transformers. Into that gap has rushed God, as usual. But it's clearly a computer program that learns how to be various computer programs for manipulating vectors. Adding numbers and coding are doable.

transformer-circuits.pub

Transformer Circuits Thread

Can we reverse engineer transformer language models into human-understandable computer programs?

200

Greg Langmead @greg.langmead.info · 03/10/2026

They are made wide (multi-head) and then they are given depth by composing, by concatenating them (maybe dozens or hundreds deep at this point?). This paper invented a programming language called RASP that is equivalent to a transformer. arxiv.org/abs/2106.06981

arxiv.org

Thinking Like Transformers

What is the computational model behind a Transformer? Where recurrent neural networks have direct parallels in finite state machines, allowing clear discussion and thought around architecture variants...

100

Greg Langmead @greg.langmead.info · 03/10/2026

You don't need to stare dumbfounded at the *outputs* and wonder what it's *doing*. What it's doing is known beyond a shadow of a doubt (modulo trade secrets, which aren't what worshippers are pointing to). It's transformers and stuff. So what are those? They are little computer programs.

100

Greg Langmead @greg.langmead.info · 03/10/2026

If it was a stochastic parrot before, it still is, since reasoning mode just keeps sending ever more tokens back through the same network. The network was re-tuned with reinforcement learning to create this reasoning mode, but it's still tackling the problem with transformers. So there's no mystery.

100

Greg Langmead @greg.langmead.info · 03/10/2026

I see people getting excited by the idea that the LLMs in "reasoning mode" (putting marketing language in quotes) can add integers. But this is a low form of activity where only entailments free of human semantics are needed. Granted it used to not do this, and now it can, which is cool.

100

Greg Langmead @greg.langmead.info · 03/10/2026

Here’s an ad from MI senate candidate Abdul El-Sayed whose audio is the noise from a data center. youtu.be/AkkumU-72JU?...

youtu.be

Mike Rogers wants to put a data center in your backyard because he's getting rich off of them.

YouTube video by Abdul El-Sayed

000

Greg Langmead @greg.langmead.info · 03/10/2026

Amazon has some FUD and self-attested fact checks about narrow technical details that do not defuse the central animating issue that the country has been taken over (again) by oligarchs. People should keep fighting, Amazon obviously feels the pushback. www.aboutamazon.com/news/company...

aboutamazon.com

The race our nation can’t afford to lose, plus a new commitment from us and a fresh set of community investments

Amazon's commitment to community-focused data center development, plus introducing Built Together, a new investment program for data center communities.

010

Reposted by Greg Langmead

Timothy Snyder @timothysnyder.bsky.social · 02/10/2026

Of course Trump flatters AI. It flattered him first.

11402102

Greg Langmead @greg.langmead.info · 02/10/2026

A lovely and sympathetic rundown of the sentiments out there, and a position statement very different from the distressed bulk of mathematicians. One element missing from his picture is the violent colonization by weird tech zealots, which is the exact piece that most infuriates me.

000

Greg Langmead @greg.langmead.info · 02/10/2026

Cuckoo for Cocoa Puffs: "Anthropic appeared quite concerned with protecting Claude"

010

Reposted by Greg Langmead

Randall Munroe @xkcd.com · 01/10/2026

If I went to Switzerland and won a Swiss watch precision contest, I would literally never talk about anything else

558342841

Reposted by Greg Langmead

John C. Baez @johncarlosbaez.bsky.social · 30/09/2026

The Nine Wise Mathematicians have handed down their recommendations for AI companies who want to prove big mathematical theorems. Their first recommendation is "Don't." It kinda weakens their point that they don't stop there. agmai.org/general-sep29/

At present, some frontier AI labs are testing advanced mathematical problems on proprietary models that remain inaccessible to the broader scientific community. Our recommendations are formulated with this practical context in mind. However, ideally, they would not do so. We want to state clearly from the start: we do not endorse this practice, and we ask them to stop testing advanced mathematical problems on proprietary models.

From: https://agmai.org/general-sep29/

5426

Reposted by Greg Langmead

Randall Munroe @xkcd.com · 29/09/2026

Jupiter Icy Moons Explorer xkcd.com/3304/

4-panel comic. (1) April 2023: [JUICE space probe flying away from Earth] FROM EARTH: *JUICE* is en route to explore the icy moons of Jupiter! JUICE: *Wheee!* (2) August 2024: [JUICE flying past Earth] JUICE: ...So which planet was I visiting again? I got distracted. FROM EARTH: Jupiter! JUICE: Right, OK. (3) September 2026: [JUICE flies past Earth again] JUICE: Is Jupiter the one with all the rings? Or... FROM EARTH: No, the big one near it. JUICE: Got it! (4) January 2029: [JUICE returning to fly past Earth again] JUICE: And *which* moons am I looking at? FROM EARTH: Oh my *god*. FROM EARTH: Now I get why we budgeted for so many Earth flybys.

141939278

Reposted by Greg Langmead

John C. Baez @johncarlosbaez.bsky.social · 29/09/2026

Here the founder of Flock says: "And then unfortunately there’s terroristic organizations like Deflock whose primary motivation is chaos. They are closer to Antifa than they are anything else and that I think is unfortunate because we don’t want chaos, or I don’t want chaos. I like law and order."

youtube.com

The AI Startup Founder That Wants To End Crime In America

YouTube video by Forbes

363

Greg Langmead @greg.langmead.info · 28/09/2026

What a fantastic thread. If you are technical, it's worth remembering that almost nobody, and almost nothing that matters, is itself technical. That said, technical work is the killer app for AI. But it's tiny.

000

Greg Langmead @greg.langmead.info · 26/09/2026

"____ spreads, with 55 new cases reported since Wednesday." Measles? Hacking by AI-lab guilty bystanders? Both?

000

Greg Langmead @greg.langmead.info · 26/09/2026

From www.pewresearch.org/short-reads/...

000

Greg Langmead @greg.langmead.info · 26/09/2026

This issue's not going away any time soon. It's noteworthy that it's decoupled from other indicators like age, and politics. Those huge LLMs might need to find somewhere else to run, like on our laptops. Which we can probably manage without the big firms.

A chart with the caption "Since January, Americans have become more likely to say data centers are mostly bad for the environment (54%), home energy costs (50%) and people’s quality of life nearby (49%). This is true for people across age groups, partisan coalitions and community types."

100

Greg Langmead @greg.langmead.info · 25/09/2026

An excellent reminder: future math is not in the training data.

000

Reposted by Greg Langmead

John C. Baez @johncarlosbaez.bsky.social · 24/09/2026

Wow, more progress on the Riemann zeta function! We know in our bones that for all positive integers n, ζ(n) = 1⁻ⁿ + 2⁻ⁿ + 3⁻ⁿ + ⋯ is irrational. But it's only been proved for n even, n = 3, .... and now, it seems, n = 5. This may seem small, but it's huge. zenodo.org/records/2282...

zenodo.org

ζ(5) is irrational

We prove that ζ(5) is irrational. The proof constructs integer polynomials Qₙ of degree 37n with 0 < Qₙ(ζ(5)) < exp(−139n²/5) for all sufficiently large n. These polynomials are rationally normalized ...

25913

Reposted by Greg Langmead

John C. Baez @johncarlosbaez.bsky.social · 23/09/2026

The American Mathematical Society, London Mathematical Society, European Mathematical Society, etc. have democratic governance (some more than others), and are better placed to make recommendations than 9 elite mathematicians chosen by OpenAI. I hope the 9 will be wise enough to see this.

161

Reposted by Greg Langmead

Digital Tolkien Project @digitaltolkien.com · 21/09/2026

Today is the 89th anniversary of the publication of The Hobbit. Here's a thread on interesting glimpses the book gave us into the legendarium Tolkien had at that stage already been working on for over twenty years but which readers at the time would have had no sense of.

116073

Greg Langmead @greg.langmead.info · 21/09/2026

This was directly visible in real time, with only the tiniest, 100% justified, leap of logic. It doesn’t need any fancy research or reports. Please, world, do not make it complicated to hold them accountable for these acts.

010

Greg Langmead @greg.langmead.info · 19/09/2026

I haven't read it, I'll have to check it out.

100

Greg Langmead @greg.langmead.info · 19/09/2026

If, like me, you can't always tell to what degree Warhammer 40k is being serious vs silly, then the existence of this should rule out some possibilities. www.youtube.com/watch?v=59NO...

youtube.com

The Emperor Protects: A Warhammer 40,000 Musical

YouTube video by Warhammer 40,000

011

Greg Langmead @greg.langmead.info · 19/09/2026

I curated resources for someone wanting to learn how to code-with-the-help-of-AI. It links to an excellent free resource for learning python. But also, I think you should learn a little about how a computer works, and about what software is, so it has those too. greg.langmead.info/courses/code...

101

Greg Langmead @greg.langmead.info · 18/09/2026

I strongly support this concept! To build models and harnesses *as a community*, with the community's values and goals baked in. There's nothing a frontier lab can do that smart people can't do for their own use cases. Worth ensuring the funding is value-aligned as well. @teorth.bsky.social

terrytao.wordpress.com

SAIR’s Open Math Model initiative

Last year I co-founded the Foundation for Science and AI Research (SAIR) with some private donors to create a non-profit organization that could support responsible uses of AI in mathematics and th…

000

Greg Langmead @greg.langmead.info · 17/09/2026

Also! the new hotness in Sil variants, Sil-Morë, has its own thoughtful touch-based design for its Android build. I have gotten it running on iOS, fixed some rotation limitations coming from SDL3, and made it available here: github.com/glangmead/si...

Screenshot of an iPad classic roguelike game named Sil-Morë.

000

Reposted by Greg Langmead

Tolkien Society @tolkiensociety.org · 15/09/2026

On 15 September 1977: J.R.R. #Tolkien's The Silmarillion was published. Edited by Christopher Tolkien, the book details the entire history of the world before The Hobbit, and goes into the legends that underpin the entire mythology including the creation of the world.

19133

Greg Langmead @greg.langmead.info · 12/09/2026

OpenAI computer boys are as effective at controlling what their models are doing as RFK is at controlling measles. Which is to say, the word "negligence" is inadequate since it implies good faith that isn't there. The purpose of a system is what it does.

000

Greg Langmead @greg.langmead.info · 12/09/2026

Turns out they infiltrated the Pentagon in June. Turns out they launched the missiles 90 minutes ago.

010

Greg Langmead @greg.langmead.info · 11/09/2026

I killed the leper by accident when I was moving around doing some testing. Please forgive me. Unless you're one of the people who also kills the lepers because they get in the way, in which case, right?

000

Greg Langmead @greg.langmead.info · 11/09/2026

You'd download an .ipa file from those links and side-load to your iPad with a free developer account and a tool like sideloadly.io

100

Greg Langmead @greg.langmead.info · 11/09/2026

Sticking with the theme of bringing open-source games to iPad, here are three variants of Angband (vanilla, NarSil, FAangband), ported in a way that I hope could get merged upstream, by working within cmake and SDL. github.com/glangmead/an... github.com/glangmead/Na... github.com/glangmead/FA...

Screenshot of vanilla Angband 4.2.6 running on iPad with a few term windows showing inventory, monsters, and objects, and a new keyboard interface for playing by touch.

221

Reposted by Greg Langmead

Terence Tao @teorth.bsky.social · 11/09/2026

A group of 25 Fields Medalists, including myself, have made a joint declaration on Math and AI: mathandai.org . We welcome additional signatories. See also this article in the Economist announcing the declaration: www.economist.com/science-and-...

mathandai.org

Declaration — Math and AI

Read the declaration and add your name.

422055930

Greg Langmead @greg.langmead.info · 11/09/2026

Wonderful essay by Matilde Marcolli on what math really is, and how math accidentally reduced itself to narrow zero-sum goals, paving the way for the AI bros to bray about nothing of value and have it appear significant. Math is not computer science. www.its.caltech.edu/~matilde/AIm...

its.caltech.edu

121

Reposted by Greg Langmead

Jason Snell @snell.zone · 11/09/2026

The premise is, there are these funny puppets and they do a variety show with a famous human and there's a bunch of comedy bits. If you need to explain it to a Disney executive, I have some DVDs you can borrow.

0456

Reposted by Greg Langmead

Grant @syncingsinking.bsky.social · 10/09/2026

This is just to say I have solved the open questions that were in your paper and which you were probably saving for your phd students Forgive me they were delicious so incentive-laden and followed so readily once your postdoc's drafts mysteriously found their way into my training data

4356115

Greg Langmead @greg.langmead.info · 09/09/2026

Right on! There's a 50-50 chance that you are in the last half of people to think a thought.

111

Greg Langmead @greg.langmead.info · 09/09/2026

"Prestige hacking" is helpful language. I've struggled to find an analogy for what AI labs are doing when they intrude in this way. They behave as if they wish to be Real Mathematicians, but since they misunderstand what that even means, they just bring their everyday violence and it's ugly.

010

Reposted by Greg Langmead

John C. Baez @johncarlosbaez.bsky.social · 09/09/2026

I think the big AI companies have hacked the prestige reward mechanism of mathematics, just as the journal oligopoly did before (we work for free and pay for the product we make, just so we can get the "prestige" needed to get jobs and tenure). (1/2)

1204

Greg Langmead @greg.langmead.info · 08/09/2026

When you read something an AI lab puts out about math, hear it in RFK Jr's voice.

000