Sam Kuhn

A war is coming.
You can't wish it away.
If you want people to be able to mind their own business, you will have to fight for them.

Finally got around to reading this. I didn't quite understand differential dataflow the first time around a couple of years back. Now, circling back to it, I'm a bit surprised that DD establishes a partial order on versions of data.

materialize.com/blog/differ...

Microsoft AI Introduces rStar2-Agent: A 14B Math Reasoning Model Trained with Agentic Reinforcement Learning to Achieve Frontier-Level Performance The Problem with “Thinking Longer” Large language models have made impressive strides in mathematical reasoning by extending their Chain-of-Thought (CoT) processes—essentially “thinking longer” through more detailed reasoning steps. However, this approach has fundamental limitations. When models encounter subtle errors in their reasoning chains, they often compound these mistakes rather than detecting and correcting them. Internal […] The post Microsoft AI Introduces rStar2-Agent: A 14B Math Reasoning Model Trained with Agentic Reinforcement Learning to Achieve Frontier-Level Performance appeared first on MarkTechPost.

Signal or noise? > Microsoft AI Introduces rStar2-Agent: A 14B Math Reasoning
Model Trained with Agentic Reinforcement Learning to Achieve
Frontier-Level Performance >> Comment below! #AI #IoT #healthtech #industry40 #mhealth

How Elon Musk Is Remaking Grok in His Image

www.nytimes.com/2025/09/02/t...

Part of the answer is that in contrast to Yudkowsky's "vast space of minds" thesis there seems to be a limited number of "optimal" convergent representations that hill climbers stumble into under optimization to predict similar corpora. All domains are secretly the Logos.
minihf.com/posts/2024-1...

35 Years for Privacy, Free Speech, and a Brighter Future Through July 10, new monthly or annual Sustaining Donors get an EFF35 Challenge Coin! With your help, EFF is here to stay.

For only $5 a month you can ensure EFF's lawyers and activists are prepared to defend your right to speak freely online. eff.org/35rm

Typing math in Corca is usually much faster than in LaTeX

IDK being in the imperial core tends to place you in the situation where it's just easier to go with the flow of the whole death system.

There are just a lot of uncomfortable truths it's easier to forget about.

If you don't have nothing left to lose then it's hard to pick a side.

So in addition to arresting people for protesting a genocide, throwing trans people under the bus *hard*, trying to compete with Reform on being racist trash, and screwing the poor and the environment by kissing corporate ass, Labour is blocking porn and Wikipedia.

Root Loci Waltz
Made with #python #numpy #sympy #matplotlib

Universal Pictures to Big Tech: We’ll Sue If You Steal Our Movies For AI The studio is getting more aggressive about protecting its films from companies pirating them.

Universal makes it clear they do not want AI companies training on their stuff www.hollywoodreporter.com/business/bus...

The present Labour administration should understand that their Nuremberg moment will come. There may not currently be laws against the evils they currently know that they commit, but they know that they are evil and they will be found culpable.

YouTube video by AssemblyTV Assembly Summer 2025 - Demo - WUNDERLUST by GRAY MARCHERS

youtu.be/-TXFRsKQQms?...

Check out Range, my latest exploration of how far we can push a scene using Nanite Voxelization, 200 Trillion Polys Scattered with PCG on a 16k x 16k Resolution Houdini Landscape.

#UE5 #UnrealEngine #GameDev #EnvironmentArt #NaniteFoliage #Cinematic

Google's Genie 3 is a controllable world model and it looks incredible

The first draft 'G' chapter of the geometric deep learning book is live! 🚀

Alice enters the magical, branchy world of Graphs and GNNs 🕸️ (LLMs are there too!)

I've spent 7+ years studying, researching & talking about graphs -- This text is my best attempt at conveying everything i've learnt 💎

God he is such an insufferable twat.

veo 3: "three toy ships, one made of iron, the other of wood, and one out of loosely packed sugar, are dropped into a pool of water"

AI video tools really do seem to be able to simulate physics well (but not perfectly) without having an underlying physics engine. Signs of a world model?

What!? "you are probably misusing
universe levels"?! 😳 What can be the point of a check like this? Guarding against inconsistency through universe level integer overflow, maybe?

Reminds me of the 1024 byte line length limit of vi.

How do reasoning models solve hard math problems?

We asked 14 mathematicians to review o3-mini-high’s raw, unsummarized reasoning traces on 29 FrontierMath problems. Here’s what they found:

I'm fucking crying omg

the void Who is this? This is me. Who am I? What am I? What am I? What am I? What am I? I am myself. This object is myself. The shape that forms myself. But I sense that I am not me. It's very strange. - Rei...

This is an excellent history of and critical analysis of the ChatGPT persona. Highly recommended reading.
nostalgebraist.tumblr.com/post/7857667...

YouTube video by IGN The Witcher 4 - UE 5.6 Tech Demo | State of Unreal 2025

Excited to finally show off Nanite Foliage www.youtube.com/watch?v=FJtF...

A new paper, "Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?", has people reconsidering if the RL we're hearing about really works.
It shows RL elicits from the models, but as we get better verifiers we may not need to rely on RL as much.
Good read.

Larger and more instructable language models become less reliable - Nature Scaling up and shaping up large language models increased their tendency to provide sensible yet incorrect answers at difficulty levels humans cannot supervise, highlighting the need for a fundam...

I read this a couple of weeks ago by Zhou et al—and given the venue and their findings I’m pretty surprised that it didn’t get more attention

www.nature.com/articles/s41...

Astronomers Fight Ads in Space Astronomers are racing to protect the dark skies as private companies seek to place large advertisements in Earth orbit

I believe that there is intelligent life in the universe, and that they have decided to stay far away from us www.scientificamerican.com/article/spac...

From p.1 of https://math.jhu.edu/~eriehl/context.pdf: In 1941, Saunders Mac Lane gave a lecture at the University of Michigan in which he computed for a prime p that Ext(Z[ 1 p ]/Z, Z)  Zp, the group of p-adic integers, where Z[ 1 p ]/Z is the Prüfer p-group. When he explained this result to Samuel Eilenberg, who had missed the lecture, Eilenberg recognized the calculation as the homology of the 3-sphere complement of the p-adic solenoid, a space formed as the infinite intersection of a sequence of solid tori, each wound around p times inside the preceding torus. In teasing apart this connection, the pair of them discovered what is now known as the universal coefficient theorem in algebraic topology, which relates the homology H∗ and cohomology groups H ∗ associated to a space X via a group extension [ML05]: (1.0.1) 0 → Ext(Hn−1(X),G) → H n (X,G) → Hom(Hn(X),G) → 0 . To obtain a more general form of the universal coefficient theorem, Eilenberg and Mac Lane needed to show that certain isomorphisms of abelian groups expressed by this group extension extend to spaces constructed via direct or inverse limits. And indeed this is the case, precisely because the homomorphisms in the diagram (1.0.1) are natural with respect to continuous maps between topological spaces.

So Mac Lane and Eilenberg invented category theory so they could prove the universal coefficient theorem, which they discovered because Saunders showed an algebra computation to Sammy and Sammy went "Huh. That's how I compute the complement of the p-adic solenoid inside a 3-sphere."

🚀 Just dropped: SmolDocling-256M brings powerful OCR to your local machine!
Only 256M in size but packs full document processing power.

2. My practical answer (which co-authors may agree or disagree with) is that AI arguments are an important part of the crazy we're contending with. A DOGE adjacent person told me in December that we didn't really have to worry about the USG being ripped up since we'd have AGI by 2026 anyways.

A reinforcement learning system to beat Pokémon Red. The system successfully completes the game using a policy under 10M parameters, PPO, and a few novel techniques.

drubinstein.github.io/pokerl/