George K

[James Vincent @jjvincent]

beautiful, mind-bending piece by T.J. Clark arguing that palaoelithic art like the chauvet caves were a reaction to the rising hegemony of language over thought — a counter-proposal that attempted to dissolve the linear, hierarchical structure of language https://t.co/Tq039iAiSH

[Matt Mullin @matthewwmullin]

You can now explore the entire US with 1-meter 3D LiDAR data!

I rebuilt my USGS elevation viewer so you can search any address, zoom all the way into the terrain, and slide between LiDAR and satellite imagery.

The detail is incredible. You can pick out riverbeds, ridgelines, old roads, drainage patterns, and terrain completely hidden by trees.

Try it here and export your own animated GIF: https://t.co/L0plghwfFY

Happy exploring!

[Andrew Yeung @andruyeung]

A few more observations from Copenhagen, Denmark, after spending two weeks there.

1. Couples here sleep with two separate blankets. Apparently, there's a thing called the Scandinavian sleep method - two blankets prevent couples from fighting over the covers and let each person control their own temperature.

2. Everything I had was cooked to technical perfection: eggs, steak, tomatoes, bacon. I literally didn't have a single dish in two weeks that was overcooked or undercooked.

3. There's an unwritten law called "Janteloven" preventing you from thinking you're special. Similar to how some places in the West have "tall poppy syndrome", it advises any person against thinking they are unique or promoting themselves. People here are pretty modest as a result.

4. I was told that consensus is a big part of Danish culture, even in the current political climate. In any disagreement, both sides keep talking until they find some type of common ground. This shows up in how many people join associations, communities, societies, etc.

5. The standard workweek is only 37 hours. Only 2% of Danish employees work long hours. Staying late is actively discouraged, and most offices empty out at 4 pm.

6. A local told me that the "measuring stick" here isn't money, career accomplishments, or ambition. But instead, it's your taste and your ability to create something you're proud of, which could be a well-designed apartment, a beautifully made object, or your sense of style.

7. Taxes here are really high. The average worker pays a net tax of ~36% If you make over $200,000, you'd be taxed ~44% vs ~36% in NYC.

8. People here love candles. The average Danish household burns 6 lbs. of candle wax per person per year, which is the highest rate in Europe.

[hunnogenomad @hunnogenomad]

A phylogenetic tree could be constructed for Indo-European cultures like this. The Yamnaya/CordedWare difference is important here; these were formed as a result of different migration waves that diverged from each other at an earlier period. Thanks to @Maptysk for their help.

[Kvali @kvali_app]

On the Evolution of the Southern Arc Hypothesis

There’s something rare and beautiful in science, and you don’t see it happen this publicly very often. In August 2022, a landmark trio of papers drops in Science under the banner of the Southern Arc. The lead author is Iosif Lazaridis, working with David Reich’s ancient DNA lab at Harvard. The big, splashy claim: the homeland of Indo-European languages - the entire family, including the Anatolian branch that gave us Hittite - lay south of the Caucasus. The evidence pointed to a genetic continuum running from eastern Anatolia through northern Mesopotamia into western Iran, a southern arc where the first Indo-Anatolian speech supposedly emerged. The steppe, in this model, was just a secondary staging ground for the later branches. News outlets run with it. Books get rewritten. The southern homeland looks settled.

Then, barely a year later, a preprint surfaces on bioRxiv in late 2023 titled "The Genetic Origins of the Indo-Europeans." Same lead author, Lazaridis. Same lab. But the conclusion has shifted dramatically. By the time the peer-reviewed version lands in Nature in early 2024, the southern arc as an ultimate origin is gone. The new root is placed further north, in something they call the Caucasus-Lower Volga cline, or CLV. This is a genetic and cultural gradient stretching from the northern foothills of the Caucasus into the lower Volga steppe, dated to roughly 4400-4000 BC. The deepest ancestor of all Indo-European languages, including Anatolian, was spoken right there, north of the mountains. The south didn't birth the family; it received an early offshoot.

The shift is not a small tweak. It’s a fundamental reorientation. In the 2022 model, Anatolian languages like Hittite were essentially stay-at-home southerners, remnants of a deep population that never left the original homeland. In the 2024 model, Anatolian is instead an early emigrant. The scenario now goes like this: the Proto-Indo-Anatolian community lived in the CLV cline. Sometime before 4000 BC, a group breaks south, crosses the Caucasus, and settles in eastern Anatolia. They carry the ancestor of Hittite, Luwian, and Palaic with them. Genetically, these people were mostly Caucasus-derived, without the full steppe ancestry package that would later define the Yamnaya. That explains why Bronze Age Hittites show zero steppe ancestry, a fact the 2022 papers had also established. The new model just flips the direction of movement and redefines the homeland.

What I love about this is how clearly it shows the self-correcting machinery of ancient DNA research. The data that forced the revision wasn't some ideological squabble; it was new samples and better models. The team kept finding that South Caucasus populations alone couldn't serve as the root for all the later Indo-European branches. The steppe groups didn't just spring from a fully southern source. They were an integral part of a cline that stretched north-south, and the linguistic root sat smack in the middle of that cline, not at its southern tip. The 2024 paper explicitly states that the South Caucasus was too genetically isolated and too diverged to be the fount of the entire language family. So the homeland walked north.

This has a pleasing symmetry. The Anatolian branch remains the first to split, the most archaic, the one that preserves sounds like the laryngeals that Saussure theorized decades before anyone dug up a Hittite tablet. But now its origin story mirrors its linguistic character: early, isolated, moving into Anatolia before the wagons and the kurgans and the massive steppe expansions, while the rest of the family stayed behind in the CLV cline and eventually gave rise to Yamnaya and all its thunderous consequences. The southern arc as a concept isn’t dead - it’s still the bridge - but it’s no longer the cradle.

For anyone following this stuff outside academia, the speed of the reversal is dizzying. Two years. That’s all it took for the same team to publish a blockbuster, then revisit their own evidence and publish a correction that reshuffles the map of the world’s largest language family. That’s not a sign of weakness. It’s what happens when the bones keep talking and you’re actually listening. The prehistory of language is being written and rewritten in real time, and the CLV cline is now the name you need to know, a corridor of grass and mountains north of the Caucasus where someone first said the words that would become "water," "night," and "daughter" in a hundred tongues.

An updated overview of aDNA relevant to Proto-Indo-European origins

[Khairallah AL-Awady @eng_khairallah1]

this is f*cking gold

Andrej Karpathy joined Anthropic five weeks ago.

A friend on his team just showed me the exact LOOPS.md file he actually uses.

I dropped it into my setup. The very first response was different.

Not slightly different. Completely different.

Claude stopped giving generic answers and started working exactly the way I think.

You don't talk to the model anymore. You build the system that talks to the model for you.

Bookmark it before it gets lost in your feed.

Read it now, then check the article below.

[Mr Panda @PandaTalk8]

不要再去读 X 上又臭又长的 harness 工程的文章了，
跟这篇文章相比， X 上的那些文章垃圾都不如。

Lilian-weng 的这篇新的博客文章，
是我目前读过的最写的清晰、最容易理解的 harness 工程， 也是递归自我改进的 harness 工程写的最好的文章。
文章是一篇框架式的详细描述了 self-Improment 的 harness engineering 全部知识框架。

链接是我做的中文版本

https://t.co/o3NpmBaQo5

[CyrilXBT @cyrilXBT]

Anthropic just dropped 5 workshops on building self-improving agentic systems from scratch:

00:00 - Ship your first Claude agent
36:44 - Build memory for Claude agents
1:05:06 - Make your agent autonomous
1:26:46 - Set up a proactive agent
2:03:35 - self-improving agents (tools,skills)

These 3-hours of free Claude workshops will replace 10 paid agentic courses.

Watch today, then read article below on how to build a self-improving agentic system with Fable 5.

[ClaudeDevs @ClaudeDevs]

https://t.co/v694m8Eaj6

[Ravid Shwartz Ziv @ziv_ravid]

1/ On Training in Imagination -
Dwarkesh's episode has a segment on dreaming as one of the next training paradigms. The idea is that a model learns mostly inside its own, by imagining what would happen, instead of trying out for real.

We have a recent paper on exactly this 🥳🥳🥳

[Rahul @sairahul1]

Anthropic engineers just dropped 40 minutes on how they actually build loops with Claude Code.

Not prompts. Not tools. Loops.

3 agents. 1 loop.

One plans. One builds. One judges.

Plan → Build → Judge → Repeat.

The agent doesn't stop until the app works.

Most people are building agents that run once and hope for the best.

The best teams are building agents that cycle, self-correct, and improve with every iteration.

That's the gap.

Watch it, then read this before you build your next loop.

[Tech with Mak @techNmak]

The person who built Claude Code mass-leaked the thinking behind it.

45 minutes of design decisions, mistakes, and where it's all going.

This is rare. Creators at this level don't usually talk this openly.

[Aarno @TheGlobalMinima]

In nearly 5 years of modern generative ai, this is the first book I’m seeing with a super high level of coverage and comprehension.
> language modelling
> inference optimisation
> RL and its methods
> system scaling
> applied concepts like agentic ai, rag, memory
> environments and benchmarking

These fields have a subtle boundary differentiating them, but ultimately overlap in modern applications. Agents require system scaling, memory needs inference optimisation, rl requires understanding of environments and benchmarks.

For the first time in my exp, all in one place. Found this on paperswithcode[.]co

[darkzodchi @zodchiii]

A Stanford team just published the 16-page PDF on “How to structure an AI agent”

Structure matters more than how you prompt it, and it's backed by hard numbers.

Build → Reflect → Curate → Reuse

• Build: the agent starts with a structured context, not a clever one-off prompt.

• Reflect: it watches what actually worked during execution, no labels needed.

• Curate: it folds those wins into an evolving playbook instead of a static prompt.

• Reuse: the next run starts from that refined structure, getting stronger each time.

This is exactly why senior engineers build the structure first in Claude Code, then let the agent run.

Read the paper, then grab the setup below 👇

[Nathan @arcane_bloom]

Amazing blog by Alisa on her job search. Here are the resources she used to study ML/LLM stuffs:

1. Stanford's "Language Modelling from Scratch" course: https://t.co/6SYi2CDHnN
(To understand the breadth of the field and keep a coherent picture in the mind)

2. After getting the breadth, she deep dived into concepts ONE at a time using blogs, papers, chatting with ChatGPT and Claude and implementing things from scratch.

3. Implementing / debugging a transformer comes up so often in interviews. Turn it into muscle memory: https://t.co/IBORUhmfU1

4. Ofc, do Leetcode🥲 https://t.co/VDawahkLOO

5. Other Learning resources she shared:
a. Self-Attention & Transformers: https://t.co/Vlowc8Ectr

b. The Illustrated GPT-2: https://t.co/f4Bui6UD1u

c. Backpropagation
https://t.co/FLoTwkMppw

d. Introduction to Policy Gradient for LMs
https://t.co/QlwnkKVTRd

e. Lightweight Guide to understanding GRPO and RL principles
https://t.co/FBBnnuZXZt

f. How to Scale Your Model
https://t.co/ryFY5Kpjmn

[Antonio Lupetti @antoniolupetti]

"An Introduction to Flow Matching and Diffusion Models" is a set of MIT lecture notes for the course "Generative AI With Stochastic Differential Equations" (2026) that provides a clear introduction to the mathematics behind modern generative AI.

The notes discuss flow matching and denoising diffusion models as core techniques behind many advanced generative systems, with references to models such as Stable Diffusion 3, FLUX, VEO-3, and AlphaFold3.

They develop the mathematical foundations of generative modelling, covering topics such as sampling from probability distributions, ordinary and stochastic differential equations, Brownian motion, diffusion processes, flow matching, score matching, classifier-free guidance, architectures for image and video generation, latent spaces, autoencoders, and discrete diffusion models for language generation.

What I particularly appreciated is the teaching style. The notes first build geometric and probabilistic intuition and only then derive the complete mathematical formulations. The result is a treatment that is rigorous, visual, and remarkably approachable.

This is probably one of the best freely available resources for understanding what is actually happening under the hood of diffusion models from a mathematical perspective.

https://t.co/J96rHCBPrb

[Antonio Lupetti @antoniolupetti]

"Transformers" by Daniel Jurafsky and James H. Martin is one of the clearest and most mathematically grounded introductions to the Transformer architecture I have ever read.

Chapter 8 introduces the Transformer as the standard architecture behind modern large language models. What makes this chapter particularly interesting is its step-by-step presentation of the underlying mechanisms: contextual embeddings, self-attention, query, key and value vectors, scaled dot-product attention, multi-head attention, residual streams, feedforward layers, layer normalization, masking, and the parallel matrix formulation of attention.

In particular, the treatment of attention as a weighted sum of contextual representations is especially valuable. The chapter first develops an intuitive, simplified view of attention and then gradually derives the full formulation using the Q, K, and V matrices. This approach makes it easier to understand what is actually happening inside the architecture from an algebraic and matrix-based perspective, rather than simply viewing the usual block diagrams.

I think it is an excellent resource for anyone interested in understanding how Transformers work from linguistic, mathematical, and computational perspectives.

https://t.co/3fitdPy6Fv

[George K @gkrivoro]

@DollarCars Watch out with these scam artists. At the SFO airport dollar, using some verbal gymnastics you will end up being upcharged to the moon without even realizing it. By my calculations, I was scammed about 500 dollars extra over a rental of 450. Illegal activities!

[༺ 𐂂 ༻ @touchmoonflower]

A bright and shining example of the suicidal absurdity of technocapitalist society occurs every year at this time, all across the country.
1/n

[U.S. Senator John Fetterman @SenFettermanPA]

I believe that there is a street safety crisis in America.

Losing hundreds of lives to unsafe streets is unacceptable. My bill, the Building Safer Streets Act, will help fix this chronic issue here in PA & across the country.