Short things, kept in the open. A commonplace book, formerly rabbitholes.garden.

From DeepSeek’s paper: Post-Training PipelineIn this release, we refrain from introducing novel post-training algorithms. The overall recipe follows the standard paradigm of supervised fine-tuning (SFT) followed by reinforcement learning (RL) and on-policy distillation (OPD) [Gu et al., 2024; Lu and Lall, 2025]…

I was just testing DeepSeek V4.1 Flash on a bunch of vision tasks with old book pages, and it’s not bad at all. It had issues with some Sanskrit type and diacritic-heavy text, but it still managed reasonably well. And considering the cost, it’s pretty impressive. During off-peak hours, the API is dirt cheap. That means…

So true:

“To poison a nation, poison its stories. A demoralised nation tells demoralised stories to itself. Beware of the storytellers who are not fully conscious of the importance of their gifts, and who are irresponsible in the application of their art: they could unwittingly help along the psychic destruction of their people.” [1, 2]

This quote originates from the book ++A Way of Being Free++ (1997), a collection of essays and aphorisms written by the Booker Prize-winning Nigerian author ++Sir Ben Okri++.

Hat tip to Jim O’Shaughnessy:

Post type: A very in-the-moment stream of consciousness, voice-typed and lightly cleaned up with ChatGPT. --- Language is probably one of the greatest inventions of mankind. I don’t think there would have been civilization, at least not in any recognizable form, without our ability to speak to one another, communicate…

This sounds like a description of Somalia or some other hellhole, but it’s Sweden. Sweden!!

Gunnar Strömmer, Sweden’s justice minister, calls gang crime “a systematic threat to society” on a par with or above that posed by terrorism. On a visit to Gothenburg, Sweden’s second city, he notes that almost a year has gone by since a deadly shooting there.

So bad was the country’s reputation that politicians in neighbouring Denmark and Norway talked about “Swedish conditions” as shorthand for everything they did not want — spiralling crime after a period of almost unchecked immigration. Swedish gangs prospered, using digital platforms to offer 12, 13 and 14-year-olds cash to commit murders or bombings, at home or abroad.

“Crime has been one of our main exports,” says Alexander Wallenius, lead investigator at the police’s national operations department, of a country more used to touting the success of its start-ups or music industry. He adds that the “No shooting” headline “should be the normal. It is a sign that the trend has changed.”

I often tell my friends that I’m white at heart because I know far more about American and European history than Indian history, and it’s a shameful fact that I’m working to fix. ​One of the books I bought a while ago was Ramachandra Guha’s India After Gandhi, which I hadn’t got around to reading. I finally picked it…

Not bad

Here’s what I got:

This is quite interesting:

Now, you might think that broader participation in the stock market would increase the strength of the wealth effect, with more households exposed to market volatility. But this is not quite right. ++New research++ from the Federal Reserve Bank of New York finds that the more widely Americans hold shares, the weaker the effect of a rate change on the real economy, which is the opposite of what you might expect: instinctively, the more people who own stocks, the harder a market swing should hit spending.

In fact, what matters, according to Juan Morelli, a New York Fed researcher, is not how many people own stocks but how concentrated the ownership is. When only a sliver of households own stocks, equity exposure is concentrated among relatively few households. So if a rate rise hits share prices, the resulting wealth shock is larger for each participating household, prompting a bigger reduction in spending.

I thought this was a thoughtful response:

At a higher level, I guess I’d say there’s a sort of refusal to think carefully about what models are or are not useful for in a rigorous way which I find personally quite annoying, and instead a reliance on some nebulous notion of being “AGI pilled” as a replacement for serious thought. I think people are very quick to anthropomorphize LLM intelligence because humans communicate through words and we infer the intelligence of human counterparties through comprehension of their language, but this leads them to wrong conclusions; for example if we observe that a new model proved some incredible mathematical theorem, some will say, “well, don’t we have AGI now, huh?” But to me, it’s actually more like, “well, given how hard it would have been for a human to do these mathematics, and given the limited economic effect of LLMs upon the world so far, isn’t it actually a negative datapoint vis-a-vis the generality of LLM intelligence?”

Andrew Ho @andrewho03 · September 3, 2026 ↗Not really observations that others before me haven’t made, but: - Despite the seemingly magical nature of LLMs, reflection over a >3 month timescale suggests my total productivity hasn’t increased by over 100%, or perhaps even by over 50%, and a lot of time is actually wasted…

As always, thoughtful response from Philippe Lemoine:

File this under mad genius:

The German chemist Felix Hoffmann (1868–1946) famously synthesized two drugs: aspirin, one of the most widely beneficial drugs ever, and heroin, one of the most harmful of illegal substances. These two drugs represent the efforts of late-19th-century chemists to create new substances that could be used as medicines, not just to isolate active ingredients from natural products or to imitate them.

One approach toward this end was to modify known physiologically active substances; another was to perform chemical operations on one or more of the myriad organic compounds created as products or by-products of the synthetic dye industry, which had developed in the 19th century and was particularly strong in Germany.

sciencehistory.org/education/scientific-biographies/fe… ↗

And people say LLMs are useless:

Been testing flash and it seems very promising:

The scale of China, no matter how much you read, continues to astound me:

Interesting:

India is ​preparing a framework that would let AI agents make small digital payments without approval for every ‌transaction, three sources familiar with the matter said, potentially making UPI one of the world’s largest networks with agentic payments.

While leading payment companies have launched such frameworks in the U.S., Europe, Singapore and Australia, a rollout on the popular Unified Payments Interface would ​also put India among the first countries to have a national infrastructure for agentic AI payments.

My thoughts on the topic:

/notes/2026-08-07-tell-your-agent-to-speak-to-my-agent/

The great Ben Evans posted this:

Posting for reference: the Lawrence Berkeley National Laboratory estimates 2023 US data centre water withdrawal at 66bn liters (17.4bn gallons). With all this investment, that might triple in the next few years. Big number!
But total US freshwater withdrawal is 102tr gallons a year. Data centres were about 0.017% of total US water withdrawal.
This can be a local planning issue, if you build in the wrong place. But the idea that AI overall is using lots of water is nonsense.
Sources: https://lnkd.in/gdnafw4z, https://lnkd.in/gaEmrn2T
Fun comparison: a kilo of almonds uses 10-15k liters of water, which is (roughly) a backyard swimming pool. A ChatGPT query uses a fraction of a teaspoon.
If you want, I can also transcribe the comments below the post or provide a direct summary of the argument being made. Let me know how you would like to proceed.

Heard DHH mention this on Lex Fridman’s podcast:

YouTube · youtube.com ↗DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux | Lex Fridman Podcast #501Lex Fridman on YouTube

Sometimes the feelings or emotions or meanings that we wish to convey run up against the limits of the everyday language that we use. This is probably one reason why we have poetry. As Tracy K. Smith put it evocatively in a conversation with Ezra Klein: “Poetry is the language that sits really close to feelings that…

Grok bot can now buy things for you:

/notes/2026-08-07-tell-your-agent-to-speak-to-my-agent/

This is a clever paper. I didn’t know this. I mean, it seems obvious once you hear it, but I didn’t know it. Reading this reminded me of something I heard someone say, maybe Liv Boeree or someone like that, on a podcast. She said something to the effect that individual rationality can lead to collective irrationality.…

Here’s an interesting study regarding Indian women’s preference for sons and its impact on mental health: The researchers surveyed a sample of married women of reproductive age and their co-residing mother-in-law in Jaunpur district, Uttar Pradesh. When there was more than one daughter-in-law in the household, they…

This is an interesting post on small language models, and I completely agree. For most tasks people have, whether in an organizational setting or in a personal capacity, frontier or large models are overkill. It’s like using a chainsaw to slice a piece of bread.

Most of our everyday use cases, inside a company or personally, are basic in nature: search, summarize, transform, set simple alerts, synthesize things, ask basic questions, and so on. For all those use cases, we don’t need GPT-4 or Claude, Opus 5 or Table. Simple small models like Gemini Flash, GPT Luna, DeepSeek Flash, and now GLM 5.3 Flash are more than good enough—and they’re dirt cheap.

This also opens up a lot of use cases for building small tools and apps by leveraging the power of cheap LLM models that can—if not be a million-dollar SaaS business—at the bare minimum have a few hundred or thousand paying users. I would bet on an efflorescence of such small apps and tools.

There is a lot of commentary online about Google falling behind in the AI race, especially considering that its long-promised 3.5 Pro never came. Since then, they’ve launched a couple of new models in the Flash series. While 3.5 and 3.6 were criticized heavily, Gemini 3.7 Flash seems to be getting some favorable…

What a beautiful thought: seek better questions.

The adoption of Chinese AI models continues unabated:

As of Tuesday, DeepSeek-V4-Flash was the most-used model in terms of token volume, while Chinese models dominated the top five, including Step 3.7 Flash from StepFun, GLM-5.2 from Zhipu, or Z.ai, and DeepSeek-V4-Flash’s updated 0731 version, which ranked second, fourth and fifth, respectively. The third spot was taken by OpenAI’s GPT-5.6 Luna, Vercel data showed.

The figures indicate that developers are increasingly choosing cheaper and open models for production workloads, particularly for autonomous agents that can consume large numbers of tokens for reasoning, writing code and calling software tools.

Interesting debate on AI, bloom and doom involving a bunch of the usual suspects:

Gold doesn’t hedge inflation. Learning this fact messed with my head when I first came across it. Ask ten people what the role of gold is in a portfolio, and the most common answer you’re likely to hear is: it’s an inflation hedge. But when you actually look at the data, gold is a terrible inflation hedge. Take pretty…

Pep’s favorite football players:

The AI vibes in America are almost definitively rotten. There was an interesting Pew survey on attitudes towards AI recently. They asked Americans whether the increased use of AI in daily life made them more concerned or more excited, and the responses are telling. Among 18–29-year-olds, 55% now say they are more…

“If the little grey cells are not exercised, they grow the rust.”

― Agatha Christie, The A.B.C. Murders

Looking at all the online discourse, it might feel like everybody is using AI. But that’s really not the case. Contrary to all the proclamations that AI is overhyped, if anything, I’d argue that AI is still underhyped. And I think the best available data is broadly on my side. If you look at surveys of AI usage…

Simon Willison on the new Qwen3.8-27B model:

The fact that a 17GB file can do all of this stuff on my home machines is a miracle. Once again, I’m delighted and amazed at how much progress local models have made this year. A year ago this would have been competitive with the best and most expensive of the proprietary models—today it can run on a capable laptop.

The only thing holding this back from being a daily driver is performance. It feels pretty slow on both the M5 Mac and the DGX Spark. That’s the catch with these dense (non-Mixture-of-Experts) models—they require a whole lot of memory bandwidth to perform well, and neither of the machines I have access to are top performers in that regard.

The most important thing about Qwen 3.8 27B is what it demonstrates. We can have an open weights general purpose model with a long context, effective tool calling, strong vision ability, and competent code generation, and we can fit the whole thing in just a 17GB file.

The geo-economics of rare earths:

Expensive to produce, expensive to store … so what’s the solution? Government intervention — and lots of it. All along the supply chain, the US has stepped in. It has bought stakes in several US rare earth producers, struck offtake agreements with some of those producers to ensure a minimum price for their products, and pledged $10bn in debt backing for Project Vault, a stockpiling initiative designed as a backstop against shortfalls in raw materials. So after all those moves, the west is FREE from Chinese hegemony, right? Right?