Bhuvan: notes

Thank you, Perry

Wed, 14 Oct 2026 00:00:00 GMT

I woke up to the news that the great Perry Mehrling has passed away. He was 67. I was shocked. It feels unfair that all the cool and awesome people keep getting taken away while assholes still roam free on Earth.

Perry was brilliant. A lot of what I understand about global macro, the dollar and the global financial system comes from listening to him. The very first time I heard him, I fell in love with his heterodox view of the global financial system. It was refreshing. I was sick and tired of the same old takes from people in macro, mostly academics.

Perry’s view was different, and you could see the insane amount of work that had gone into it. His “money view” was sophisticated and built on the work of Hyman Minsky, Charles Kindleberger and others. His last book, Money and Empire, was about Kindleberger and the dollar system.

Predictably, everyone is sharing his famous INET videos, and rightly so. If Perry is in a video, it’s worth watching. But for me, most of what I know about the dollar-based financial system and the centrality of the dollar comes from his many interviews and podcasts over the years.

What struck me most was how intellectually honest and open he was. He was generous in sharing what he knew and, more importantly, willing to admit what he didn’t. That was the first thing I felt grateful for when I heard the news. Plenty of academics love sharing their research, but you’d be surprised how many have an “oh my precious baby” syndrome about their work. Perry just wanted to teach people.

At the very least, I owe him a deep debt of gratitude. A large part of how I understand money has been shaped by listening to him.

Perry, I wish you were still around, doing all those podcasts and interviews so I could keep learning from you. Rest in peace, and thank you for so graciously sharing everything you knew.

Some of my favourites:

Perry Mehrling on Trump’s echoes of the Nixon shock (Odd Lots, Bloomberg)
“The US dollar is still the least dirty t-shirt in the drawer” (Prague Finance Institute)
The rise of the global dollar system (Hidden Forces)
Money markets and capital markets: the shadow banking system (CIGI, False Dichotomies)
Payments vs funding: the law of reflux for today (Young Scholars Initiative)
A money view on the emergent new international economic order (Young Scholars Initiative)
Financialization and its discontents

If you want to go deeper:

Economics of Money and Banking, his course on Coursera, filmed with INET
His lecture notes for the course
His articles and papers at INET, including “Payment vs. Funding: The Law of Reflux for Today”
His blog, papers and videos
His books, from The New Lombard Street to Money and Empire

The real change India needs

Wed, 14 Oct 2026 00:00:00 GMT

Source: x.com

The real change India needs is for people to clean up after themselves and to maintain things. We’ll go a long way if we just maintain what’s already built. A cult of preservation and maintenance is underrated. Building things is sexy; preserving and maintaining them is not.

When facts change I change my mind. What do you do sir?

Tue, 13 Oct 2026 00:00:00 GMT

Source: x.com

They seem to be living in an era where if you don’t have the ability to change your opinion on something on a dime, you’re toast.

We live in exponential times and we need exponential mental models to navigate these disturbing, treacherous, exciting, wonderful, shocking, and scary times.

Twitter is annoying, frustrating, but still awesome

Mon, 12 Oct 2026 00:00:00 GMT

Hard agree. For all it’s flaws, Twitter is still awesome

Daring Fireball · daringfireball.net ↗X Once Again Marks the SpotIf you returned from exile on a desert island and took a look at X today for first time in a year, you might reasonably guess that Elon Musk got bored and sold it while you were away.

William Hazlitt on Jeremy Bentham

Fri, 09 Oct 2026 00:00:00 GMT

Source: x.com

This is funny

The times, they have changed.

Wed, 07 Oct 2026 00:00:00 GMT

Source: x.com

The rage economy

Tue, 06 Oct 2026 00:00:00 GMT

Source: x.com

“Rage economy” is a really nice framing:

AI videos are crazy

Mon, 05 Oct 2026 00:00:00 GMT

These AI videos are crazy:

Hawkeye1812Z @Hawkeye1745 · October 4, 2026 ↗🇰🇵🇷🇺🇺🇸American AI generated this: MOM @mom_agency_ · October 5, 2026 ↗This is why GPUs cost $6,000 The AI Investor @The_AI_Investor · October 5, 2026 ↗We need more GPUs 😂 Ankit @BytesWithAnkit · October 5, 2026 ↗AI's gotten so good that nobody's gonna believe the top video is the AI one 😌

On brain rot

Sun, 04 Oct 2026 00:00:00 GMT

Source: publicbooks.org

A thoughtful essay by Mahika Dhar in Public Books on brain rot as political resistance. Please read it.

I want to preface this by saying I could be wrong. This might well be a misreading of her piece. But I came away with a small confusion, and I think it’s worth sitting with.

A little mental exercise: replace “brain rot” with “satire” throughout the essay. A lot of the meaning still holds. That makes me wonder if there’s a subtle reframing going on. I don’t want to make the mistake of suggesting the author doesn’t know what she’s doing. Right at the outset she points to how fluid the term is, how it can hold a whole lot of meanings. But if you take brain rot in its popular sense, on its own terms, it’s vacuous content. Plain nonsense. Not the nonsense of Dylan, or of the poets who used it to poke power in the eye. Just drivel. (And I say this not as one of those hand-wavy critics lamenting the decline of literary culture. That’s not my point at all.)

That raises the question of who exactly Mahika means when she says brain rot. It seems to me there are three parties: the people who produce it, the people who consume it, and the people who reuse it. I think the reusers are what she’s really getting at, and it’s a brilliant point. Even as I write this, the Cockroach Janta Party is back at it, demanding the Chief Election Commissioner resign over the SIR controversy. You can see the same playbook: taking what would otherwise be throwaway brain rot and using it to drive home a political point.

But treating all brain rot as the same doesn’t seem quite apt to me. Is all brain rot nonsense? Is all nonsense brain rot? And if there’s a sharper point embedded in a piece of brain rot, is it still brain rot? Or has it become satire, sarcasm, absurdist commentary?

Take the ur-example of satire, Swift’s A Modest Proposal, where he suggests in a surgical deadpan that the Irish poor sell their children as food for the rich. Run it through the brain rot test. Is it brain rot, or is it satire?

Maybe this is all semantic yogasana on my part, and it obscures Mahika’s actual point: that brain rot works as a tool for saying serious things about power. Fair enough. But elevating all brain rot to the same level risks, paradoxically, flattening the very many meanings of the word she points to at the start. You could just as easily swap “brain rot” for “slop” in a lot of cases. What’s brain rot to one man is slop to another.

Having said all of this, I think the important point Mahika makes still stands, and it’s worth appreciating in full: that brain rot, done well, can be appropriated to speak back to power. It’s an underrated point, and not many people have made it. Kudos to Mahika for seeing past the absolutist readings and moving (delving, wink wink) into the nuance.

In a way, it’s a new linguistic development, if you think about it. I think we’re past the age of manifestos and polemics, that highly intellectualized, Europeanized mode of protest and speaking back. Memes, brain rot, slop have become tools in the linguistic toolkit of a new generation, used to raise grievances, to ask for what’s due, to poke power in the eye. In 2026, a good brain rot meme, used well, can do more than a ten-page manifesto.

Is this for the good or the bad? I honestly don’t know. And here I must confess I might sound like just another person with a generational gripe, but that’s not where I’m coming from. I just don’t have a good sense of how this plays out.

Gandhi vs Ambedkar

Sat, 03 Oct 2026 00:00:00 GMT

Source: thehindu.com

A snippet from this Hindu article with an excerpt from Rama Chandra Guha’s Arguments with Gandhi:

Ambedkar saw Gandhi as a timid and hesitant reformer. Orthodox Hindus, on the other hand, saw Gandhi as a dangerous revolutionary attacking the very basis of scripture and tradition. This disjuncture in viewpoints brings to mind some comments of the great black abolitionist Frederick Douglass on his great white contemporary Abraham Lincoln. In a speech of 1876, Douglass remarked that ‘viewed from the genuine abolition ground, Mr. Lincoln seemed tardy, cold, dull, and indifferent; but measuring him by the sentiment of his country, a sentiment he was bound as a statesman to consult, he was swift, zealous, radical, and determined.’

The two greatest opponents of untouchability in the first half of the twentieth century were Gandhi and Ambedkar. Their differing social locations and political affiliations produced sharp disagreements between the two men. In recent decades, as Dalits have grown more self-aware and better organised, they have turned more, much more, to Ambedkar than to Gandhi, to the activist intellectual who was himself a Dalit rather than to the guilt-ridden upper-caste reformer. This elevation of Ambedkar over Gandhi was inevitable, and appropriate. However, the politics of the present have also led to a wholesale repudiation of Gandhi on the part of those who seek to fight against the injustices of caste.

amazon.in · amzn.in ↗Buy Arguments with Gandhi Book Online at Low Prices in India | Arguments with Gandhi Reviews & Ratings - Amazon.inAmazon.in - Buy Arguments with Gandhi book online at best prices in India on Amazon.in. Read Arguments with Gandhi book reviews & author details and more at Amazon.in. Free delivery on qualified orders.

Sukumar Sen and India’s First Election

Fri, 02 Oct 2026 00:00:00 GMT

Source: amzn.in

I’m reading Ramachandra Guha’s magisterial India After Gandhi, and one section stopped me in my tracks: his account of how India conducted its first general election. The scale of it is staggering. Guha walks through the numbers, the infrastructure, the people and the resources it took, and he credits much of it to one man, Sukumar Sen, about whom, as Guha notes, we know very little.

With the country in the middle of important debates over the SIR, the Special Intensive Revision of electoral rolls, it seems a good moment to step back and see how far we’ve come. India was one of the first countries to begin its democratic life with universal adult suffrage, rather than expanding the vote bit by bit. Through all the ups and downs since, the right to vote has remained an indispensable part of what it means to be Indian. That history should frame the arguments we’re having today.

So I asked Claude to write a short note on Sukumar Sen and the sheer scope of India’s first general election. Here it is.

An act of faith

Between October 1951 and February 1952, a country that had been free for barely four years asked roughly 176 million adults to choose their government. Some 85 percent of them could not read or write. Most had never voted for anything, and the world largely doubted it could be done.

It was the largest election humanity had ever attempted, spread over a million square miles of mountains, deserts, jungles and floodplains. The historian Ramachandra Guha has called it an act of faith. The man charged with turning that faith into ballot boxes was a quiet civil servant named Sukumar Sen.

The mathematician in charge

Sen was a Bengali, educated at Presidency College in Calcutta and University College London, where he won a gold medal in mathematics. He joined the Indian Civil Service in 1921 and rose to chief secretary of West Bengal. In March 1950 he was appointed India’s first Chief Election Commissioner.

His first and perhaps most important act was to say no. Prime Minister Nehru wanted elections as early as 1951. Sen, whom Guha suspects was thinking like a mathematician, asked him to wait. A census had to finish, election laws had to pass, constituencies had to be drawn, and every adult in the country had to be found and written down. A rushed vote that failed would have discredited democracy itself.

Building the machine

The numbers Guha gathers convey the scale. Nearly 4,500 seats were at stake, about 500 for Parliament and the rest for state assemblies, all elected at once. Some 224,000 polling booths were set up and about two million steel ballot boxes were made. More than 56,000 presiding officers ran the booths, backed by some 280,000 support staff and roughly as many police as there were booths.

The rolls were the first hurdle. Thousands of clerks were hired to compile them by hand. In parts of northern India, women would not give their own names, only “wife of” or “mother of” some man. Sen called this a senseless relic and ordered officials to record real names. About 2.8 million women who could not be properly identified were struck off, a painful price he accepted to keep the rolls honest.

He also had to design a vote for people who could not read. Each candidate got a pictorial symbol, such as a pair of bullocks, a hut or a lamp. Instead of marking a ballot, voters dropped a paper into the box carrying their candidate’s symbol. Mock polls were staged in 1951 to test the system, films and radio talks explained it, and phials of ink, nearly 390,000 by Guha’s count, were used to mark voters’ fingers against fraud.

The vote

Polling was staggered over about four months so that officials, boxes and police could move across the country. Snowbound hill districts voted first, in late October 1951, before winter closed the passes. A schoolteacher from Kalpa in Himachal Pradesh, Shyam Saran Negi, is remembered as the first person to cast a vote in independent India; he kept voting into his hundreds.

The rest of the country followed through early 1952. Turnout was roughly 45 percent, remarkable for a first-time electorate. The Congress Party won a large majority, and Nehru became India’s first democratically elected prime minister. Just as important, the losers accepted the result.

A forgotten founder

The success made Sen briefly famous abroad. Sudan invited him to oversee its first election in 1953, and India gave him the Padma Bhushan in 1954. He ran the second general election in 1957 too, using ballot boxes salvaged from the first, and left office in 1958. He died in 1963.

Yet Guha laments how little we know about him. He left no memoir, and his name rarely appears beside Nehru, Patel or Ambedkar. That may be the truest measure of what he achieved. Sen built a process so trusted that Indians came to take it for granted, and a democracy that the world expected to fail has held elections ever since.

Further reading

Ramachandra Guha, India After Gandhi (2007), chapter “The Biggest Gamble in History” — the best-known account. An extract is online at The Caravan.

Ornit Shani, How India Became Democratic (2017) — how the electoral rolls were built from 1947 onward, even before Sen took charge.

Election Commission of India, Report on the First General Elections in India 1951-52 — Sen’s own two-volume official account.

Election Commission of India, Leap of Faith — an illustrated history dedicated to Sen.

Chambers Book of Indian Election Facts (Hachette India) — figures and anecdotes; see the Mainstream review.

Note: sources differ on the ink. Guha records ink used in 1951-52, while others date the modern indelible ink to the 1962 election.

A tale of two Chinas

Thu, 01 Oct 2026 00:00:00 GMT

Source: blackrock.com

Nice note from BlackRock:

Outside China, the key question is whether companies use Chinese technology or compete against it. Cheaper Chinese industrial robots and machinery can boost productivity and ease labor shortages for their customers. For rival producers, the economics are less appealing. European machinery makers illustrate the squeeze. China has moved from a major buyer of advanced machinery to a formidable competitor at home and abroad. In 2025, it overtook Germany as the world’s largest machine-tool exporter for the first time.

iAI creates a similar divide. Cheaper Chinese models could accelerate AI adoption while commoditizing the model market. That could shift value toward the physical infrastructure needed to support greater AI use - computing power, data centers and electricity - reinforcing the scarcity investment theme from our Midyear Global Outlook. Physical AI could amplify demand for power electronics and batteries while broadening demand for components like motors and actuators – areas where China already has a manufacturing and supply-chain advantage. But technology sovereignty cuts both ways: export controls and efforts to reduce reliance on Chinese technology could limit market access. That makes geopolitics a key risk to where the opportunity ultimately lies.

Inside China, industrial policy has helped build scale and competitiveness, but growth alone does not guarantee shareholder returns. Long-standing challenges – including weak domestic consumption, a soft property market and an aging population – remain, while intense competition can squeeze margins even in industries where China has gained global market shares. The divergence is stark: profits at Chinese electronic-device makers rose 110% year over year in the first seven months of 2026, supported by global AI demand, while auto profits fell 20.4%, according to data from the National Bureau of Statistics of China. Likewise, AI hardware suppliers have expanded margins, while solar manufacturers have seen theirs collapse. Markets echo that divide: the MSCI China index has lost 11% in U.S. dollar terms this year through Sept. 21, while the MSCI China Information Technology index has gained 7%.

Indian government sucks at industrial policy

Wed, 30 Sep 2026 00:00:00 GMT

Source: frontline.thehindu.com

Raghuram Rajan on industrial policy:

I’d say that I am not against the government’s actions; I am very much for the government’s actions on dealing with the fundamentals. So let’s focus on public goods such as education and healthcare, let’s focus on building good roads, let’s focus on creating a good business environment, let’s focus on utilities, let’s focus on water, electricity, logistics. These are areas where you have a much better chance of success improving things rather than [picking winners and sectors].

Whenever our government showers a sector with love, it has not ended particularly well except for the areas I just mentioned. Kudos to the current government for building out Indian infrastructure. That’s great. But I think if you test them on manufacturing, have they really created more manufacturing jobs? The line is flat [on the chart]. Look at the line. I’d say that just the evidence suggests that we have not been great. Maybe some other countries have seen success.

Sales of weight loss drugs in India

Wed, 30 Sep 2026 00:00:00 GMT

Source: bloomberg.com

The weight loss revolution in India seems to be doing ok?

And right on cue, the fakes have arrived on the scene:

Spurious drugs are a recurring problem in India, said Mahesh Zagade, a former commissioner of the Maharashtra Food and Drug Administration. He said that while the country’s laws are robust, implementation and enforcement have been lacking. Short-staffed regulators, the possibility of corruption, and easy access to raw materials have contributed to the problem, in his view. “This is a failure of the official machinery,” he said, adding that until action is taken to fix the broader system, “this situation is just not going to improve.”

Begum Aizaz Rasul

Tue, 29 Sep 2026 00:00:00 GMT

Source: en.wikipedia.org

What a legend. Came across the name for the first time in India after Gandhi:

With the enactment of the Government of India Act 1935, the couple joined the Muslim League and entered electoral politics. In the 1937 elections, she was one of the few women who successfully contested from a non-reserved seat and was elected to the U.P legislative assembly. Begum Aijaz Rasul remained a member until 1952. She held the office of the Deputy President of the council from 1937 to 1940 and acted as the Leader of Opposition in the council from 1950 to 1952–54. She was the first woman in India and the first Muslim woman in the world to reach this position. Despite her family background, she was known for her strong support for zamindari abolition. She also strongly opposed the demand for separate electorates based on religion.

In 1946, she was elected to the Constituent Assembly of India and was one of 28 Muslim League members who finally joined. She was the only Muslim woman in the Assembly. In 1950, the Muslim League in India dissolved and Begum Aijaz Rasul joined the Congress. She was elected to the Rajya Sabha in 1952-54 and was a member of the Uttar Pradesh Legislative Assembly from 1969 to 1989.

Between 1969 and 1971, she was the Minister for Social Welfare and Minorities. In 2000, she was awarded a Padma Bhushan for her contribution to social work.[2]

How much foreign investment did India attract?

Sun, 27 Sep 2026 00:00:00 GMT

Source: thisindianlife.today

New entry in This Indian Life:

NVIDIA Nemotron 3 Diarization

Fri, 25 Sep 2026 00:00:00 GMT

Interesting model. Yet to test it out:

NVIDIA Nemotron 3 Diarization is an open-weight, 100M-parameter model that ranks #1 on VoiceArena’s Diarization-Bench leaderboard with a 14.72% Diarization Error Rate (DER). Supporting up to eight speakers across live and recorded conversations, it handles overlapping speech, chunked processing for flexible recording lengths, and customizable streaming latency

huggingface.co ↗**Know Who Spoke When: Build Real-Time, Multi-Speaker AI with NVIDIA Nemotron 3 Diarization**A Blog post by NVIDIA on Hugging Face

Galbraith’s Antidote to Economic Dogma

Thu, 24 Sep 2026 00:00:00 GMT

Source: en.wikiquote.org

Years ago, I came across a quote by John Kenneth Galbraith in Tom Butler-Bowdon’s 50 Economics Classics—a brilliant book, by the way, and an accessible introduction to fifty foundational works that most of us are unlikely to ever read cover to cover. Galbraith’s point was simple: “Where the market works, I’m for that. Where the government is necessary, I’m for that… I’m in favor of whatever works in the particular case.”

I find myself thinking about this often, especially during the shrill debates that keep erupting between self-identified capitalists and socialists, libertarians and liberals, or the big-market versus small-market camps. These labels get endlessly thrown around in arguments that are completely devoid of nuance and context—and, more importantly, devoid of any real comprehension of the historical conditions under which these economic logics, ideologies, and systems actually arose.

Here’s the full quote:

I react to what is necessary. I would like to eschew any formula. There are some things where the government is absolutely inevitable, which we cannot get along without comprehensive state action. But there are many things — producing consumer goods, producing a wide range of entertainment, producing a wide level of cultural activity — where the market system, which independent activity is also important, so I react pragmatically. Where the market works, I’m for that. Where the government is necessary, I’m for that. I’m deeply suspicious of somebody who says, “I’m in favor of privatization,” or, “I’m deeply in favor of public ownership.” I’m in favor of whatever works in the particular case

amazon.in · amzn.in ↗50 ECONOMICS CLASSICS : Butler-Bowdon, Tom: Amazon.in: Books50 ECONOMICS CLASSICS : Butler-Bowdon, Tom: Amazon.in: Books

Coding from my phone is absurd

Wed, 23 Sep 2026 00:00:00 GMT

Source: chatgpt.com

I’ve been using Claude Code, and to a lesser extent Codex, from my phone, and it is genuinely game-changing.

I’m very late to this, obviously. But this is really awesome.

I just switch on my laptop, open a bunch of sessions, and then no matter where I am — at a coffee shop, out for a walk, whatever — I can quickly tell Claude or Codex to do something, check the progress, start something new, and move on.

It is kind of crazy that this is even possible in 2026.

Just think about the sheer absurdity of what is happening here. People will hate me for using this term, but there are intelligent agents at my disposal which I can prompt from the comfort of my phone, and they will go away and write code that is better than what a lot of programmers can write.

And all of this is being done by a person who doesn’t know C++ or JavaScript from idli vada and vangi bath.

Do you see the irony here?

And yet people are still arguing over whether LLMs are useful, whether they are good or bad, whether they “really” do this or that, and so on and so forth.

The gulf between people who have figured out how to get actual utility out of these models and people who are still making these abstract arguments about whether LLMs are good or bad is just nuts.

Claude Code Docs · code.claude.com ↗Claude Code on mobile - Claude Code DocsStart, monitor, and steer Claude Code tasks from your phone with the Claude app for iOS and Android.

Mancur Olson on collective action

Tue, 22 Sep 2026 00:00:00 GMT

Source: en.wikipedia.org

Today in learning:

In his first book, The Logic of Collective Action: Public Goods and the Theory of Groups (1965), he theorized that what stimulates people to act in groups is incentive; members of large groups do not act in accordance with a common interest unless motivated by personal gain (economic, social, etc.). While small groups can act on shared objectives, large groups will not work towards shared objectives unless their individual members are sufficiently motivated.[7]

In 1982, he expanded the scope of his earlier work in an attempt to explain The Rise and Decline of Nations (1982). He argues that groups such as trade associations, business cartels, and labor unions have an incentive to form lobby groups and influence policies in their favor. These policies will tend to be protectionist, which will hurt economic growth; but because the benefits of such policies are concentrated, and their costs are diffused throughout the whole population, there will be little public resistance to them. As distributional coalitions accumulate, nations burdened by them will fall into economic decline. His work influenced the formulation of the Calmfors–Driffill hypothesis of collective bargaining.

How to value companies when you’re a cokehead

Mon, 21 Sep 2026 00:00:00 GMT

Source: ft.com

Valuations for cokeheads:

Now, TAM is de rigueur in any venture capitalist’s slide deck or IPO prospectus. Its usefulness is another matter. Uber touted a $12.3tn TAM in its 2019 listing; its annual revenue today is less than $60bn. WeWork’s $3tn addressable market of office properties didn’t save it from bankruptcy. But for a company like Anthropic, TAM is as good a place as anywhere to start.

Investors aren’t groping totally in the dark, thanks to SpaceX. Elon Musk’s company, which makes the Grok chatbot as well as owning social network X, claimed in its May listing documents that “enterprise apps”, which would include its products and services backed by AI, make up a market worth $22.7tn — a little more than the entire annual revenue of the S&P 500.

With such a large theoretical market to play for, a little goes a long way. Even the mighty Microsoft’s revenue is less than 2 per cent of what Musk sees as the industry’s overall opportunity. If Anthropic can muster 3 per cent, it would stand to make nearly $700bn of revenue. Pop that on a multiple of 10 times and discount it back assuming it takes three years to achieve, and out comes a valuation of $4.5tn.

Qwen3.8-Omni-Flash and Qwen3.8-LiveTranslate

Sun, 20 Sep 2026 00:00:00 GMT

Interesting model launches from Qwen. Omni Flash especially seems promising as a workhorse model for simple text classification and voice transcription tasks. Just started testing it in a voice dictation app I built myself to transcribe my voice notes and so far so good.

some audio tasks and so far so good.

qwen.ai ↗QwenQwen offers comprehensive functionality spanning chatbot, image and video understanding, image generation, document processing, web search integration, tool utilization, and artifacts. qwen.ai ↗QwenQwen offers comprehensive functionality spanning chatbot, image and video understanding, image generation, document processing, web search integration, tool utilization, and artifacts.

We were sloppy we’ll before AI

Sun, 20 Sep 2026 00:00:00 GMT

If AI is slop, then so are humans!

x.com/kmcnam1/status/2101356137396265106 ↗

Notes from messing around with vision, text, and audio models

Sat, 19 Sep 2026 00:00:00 GMT

The bulk of my vibe-coding experiments over the last few months have involved audio and video, or rather audio and vision, and I’ve been experimenting pretty heavily with multimodal models.

So this is just a small guide based on what I’ve used.

If you have some project where you need vision capabilities for text extraction, or models for speech-to-text workflows — recording voice, transcribing audio, cleaning up audio, voice journaling, whatever — some of this might be useful.

Most of this is based on actually using these models in different workflows for months, rather than looking at benchmarks. And my use cases are also slightly weird, especially when it comes to vision.

Vision

If the documents you’re trying to digitise are modern, and they involve modern English, I think pretty much all the major vision models will work.

At that point, the more important question is cost.

If you have a heavy workload, cost becomes a problem very, very quickly.

The newer DeepSeek Flash model seem very promising here. It has vision capabilities and it’s ridiculously cheap.

The Gemini Flash-Lite family is also very cheap and has been pretty reliable for me. The Qwen Flash family is reasonably reliable. Qwen 3.7 Flash in particular seems promising. GLM 5.3 Flash is another one worth trying.

The reason I keep mentioning Flash models is simply that they are really, really cheap.

If you need a workhorse model for a heavy workload, that matters much more than people realise. If you start throwing thousands or tens of thousands of pages at frontier models, the cost gets prohibitive, to put it lightly.

The Gemini Flash family is slightly costlier but also very good.

So if your use case is basically extracting text from normal modern PDFs, my recommendation is: start cheap. You almost certainly don’t need the smartest model on earth to read a clean PDF.

Old books are where all of this becomes annoying.

I have a weird interest in digitising very old books, which is what I do with Akshara, and I’ve also been digitising old collections of letters.

I have wasted enough dollars on this particular problem to have opinions.

There is no best model.

Of all the models I’ve tested, the Gemini Flash family has probably been the most consistently accurate and reliable for old books, and occasionally Gemini Pro is useful if the page is particularly horrible.

The big problem with Gemini is recitation blocks.

Every now and then Gemini will simply refuse to extract the text from a page because it has triggered some safety filter or some other dumb and stupid thing. So you end up with a model refusing to reproduce a page from a 120-year-old book.

In those cases, Qwen is a reasonable backup. You will sometimes sacrifice some accuracy, but it gets the job done.

In terms of accuracy, Gemini generally ranks much higher than Qwen for the material I work with. But with some books it is more or less a wash.

I would not use DeepSeek as my primary model for old books.

It has a tendency to aggressively modernise the text. Archaic spellings become modern spellings and things like that, which is obviously terrible if the entire point is to preserve what was actually printed on the page.

Qwen has a similar problem sometimes.

I’ve also tested Kimi and GLM Turbo for old books. They’re reasonable backups. I wouldn’t use them as my primary models.

The broader problem I’ve found with a lot of these Chinese models is that there isn’t one failure pattern. There are several.

Some models modernise the text. Some are bad at preserving old punctuation. Some are bad at archaic spellings. Some flatten or drop diacritics. With old Indian books, transliteration itself can become a problem because the book may be using some older or idiosyncratic Sanskrit transliteration scheme and the model will helpfully “correct” it into something resembling modern IAST.

Which is useful if you want normalised text.

It is a disaster if you want a faithful transcription.

They also fail differently, which makes this more annoying. One model may preserve the spelling but screw up the punctuation. Another may preserve the page structure but quietly modernise words. Another may start dropping diacritics.

So if your requirement is digitising genuinely old documents, you need to test these things properly rather than assuming “OCR accuracy” is one number.

For modern documents, I think this distinction matters much less.

If the PDF is clean and contemporary, pretty much all the major vision models are going to be in the same broad neighbourhood. At that point I would optimise for cost.

And for heavy workloads I really would not recommend frontier models unless you have a very specific reason to use them. The cost becomes silly. Flash models usually get the job done.

You can also do hacky things around pricing. Keep an eye out for API discounts, temporary free models, OpenRouter, OpenCode, and so on. Sometimes there are new or anonymous unreleased models floating around and if you have some huge one-off workload you can abuse these things while they last.

I also tried some of the newer Qwen models through Qwen Cloud. The models themselves can be quite good. The API, at least in my experience, has been unreliable enough that I would not build some enormous unattended workflow around it without fallbacks.

So, more or less: for normal PDFs, use the cheap Flash models and save your money.

For horrible old books, Gemini is still my first choice, and Qwen and some of the other Chinese models are useful backups when Gemini suddenly develops moral objections to OCR.

My experience here is obviously skewed because my main use case for vision is unusually stupid: digitising old books and old letter collections.

I also briefly started another project where I wanted to extract the text from a gigantic collection of PDFs. I very quickly discovered that this was going to become expensive, so that project has taken a back seat.

Text

The same thing is true for text.

If your workflow involves a lot of summarisation, rewriting, extraction, text generation, classification, whatever, I think at this point pretty much any decent Flash model will get the job done.

I was running an experiment where I would take raw stock-exchange filings and transform them into almost instantaneous news stories about the companies involved.

For that, I used DeepSeek Flash and Xiaomi MiMo V2.5 quite a bit.

Both were perfectly decent.

I particularly like DeepSeek Flash. Its prose capabilities are surprisingly good.

If your benchmark for prose is Shakespeare or Hemingway, then sure, use the frontier models.

But if your benchmark is “pretty good”, which I think covers most bulk text-generation workflows, then the Flash models are more than enough.

And they’re dirt cheap.

Once you’re calling a model thousands and thousands of times, this difference matters. The frontier models become prohibitively expensive for things that are frankly not difficult enough to justify them.

This also means you don’t necessarily need one multimodal model to own your entire audio workflow.

If you prefer keeping transcription and editing separate, use a cheap ASR model — Microsoft MAI-Transcribe-2, OpenAI GPT-Transcribe, whatever — to create the raw transcript.

Then pass the transcript through something cheap like DeepSeek Flash, GLM Flash or Gemini Flash-Lite to clean it up, remove filler, structure it, rewrite it, summarise it, whatever you need.

For heavy workloads, this can be much, much cheaper.

Audio

I’ve taken heavily to voice typing.

In fact, because I write regularly, I’ve almost stopped manually typing. Typing now feels slightly archaic to me. Pressing all those disgusting keys on a computer. Horrible.

But I also think voice typing unlocks a slightly different kind of writing.

It is not necessarily a one-for-one replacement for the regular disgusting form of writing where you sit in front of a keyboard.

A lot of my writing is very short form: fleeting thoughts, observations, musings. Voice typing is perfect for that.

A lot of my writing is also exploratory. I’m trying to work through an idea without necessarily knowing what the idea is yet. I’m perfectly fine putting out the half-baked thought first and then refining it later.

Voice works very well for that too.

The workflow that makes sense to me is: capture everything first.

Raw ideas, unfinished thoughts, repetitions, weird digressions, whatever.

Then worry about turning it into writing later.

Basically, your phone becomes one of those little Dictaphones journalists used to carry around.

ChatGPT works perfectly fine for this. Gemini is a close second.

But if you want to build your own app, which I have done for myself, there are a few different ways to do it.

One is to just use a multimodal model that can take the audio directly and do transcription, cleanup and structuring in one pass.

The Gemini Flash and Flash-Lite models are very good at this.

You can basically tell Gemini: listen to this, transcribe it correctly, remove filler and repetitions and all the other speech debris, and give me a cleaned-up post.

It works really well.

If you don’t care about keeping the raw transcript, this is probably the easiest workflow.

If you want a backup to Gemini for this sort of thing, the Qwen Omni family seems promising too. It can take the audio, understand it, and give you the transformed text in one pass.

The other option is to keep the raw transcription and the editing workflow completely separate.

I actually like doing this because sometimes I want all my raw thoughts preserved in their original ugliness.

Then you just use an ASR model — Automatic Speech Recognition, speech-to-text, whatever you want to call it.

Gemini 3.5 Transcribe has worked very well for me.

Microsoft MAI-Transcribe-2 came out recently. I tested it a little. Seems promising. Very cheap.

OpenAI’s GPT-Transcribe family is also worth trying.

There is also GPT-4o Transcribe Diarize if your use case involves multiple speakers.

Qwen’s ASR models are decent and reliable.

Grok Voice Transcribe 2 came out literally while I was writing the original version of this note. I haven’t tested it properly yet, so I have no opinion on the accuracy, but again it is dirt cheap.

And at these prices, transcription itself is increasingly not something you need to spend too much time worrying about.

I know people whose entire use case is recording meetings and then turning the meetings into transcripts, notes and summaries.

If you have multiple speakers, look for speaker diarisation, which is basically the model figuring out who spoke when. Some of these transcription models have it built in now.

Another failure pattern worth mentioning is vocabulary.

Very specific words can get mangled repeatedly. Indian words, names, technical jargon, model names.

For some reason Qwen itself is a good example. Q-W-E-N. Alibaba’s Qwen.

A surprising number of transcription systems manage to get that wrong.

A better prompt often fixes some of this.

And depending on the ASR system, you may also have things called custom vocabulary, hotwords, keyword hints, phrase boosting, etc. The terminology varies, but the idea is basically the same: give the transcription model a list of weird words it should expect to hear.

If your workflow has a lot of specialised vocabulary, use that.

And then there is the third option: run everything locally.

I’ve also used local speech models on my laptop, mostly OpenAI’s Whisper and NVIDIA’s Parakeet family.

Again, my use case is voice typing.

I built myself a little desktop app called Yawp.

Or rather Claude Code and Codex built it for me because I was too lazy to type.

It works only on Linux.

The workflow is basically similar to Wispr Flow: press and hold a button, speak, let go, and it transcribes the audio and pastes the text into whatever input surface I’m using.

It works reasonably well.

I should point out that my testing setup here is idiotic.

I have been too lazy to buy a proper microphone, so from about 50–60 centimetres away I either shout at my laptop or I lean over and speak like a moron directly into the built-in microphone.

So the accuracy is a little lower than I would like.

I strongly suspect this is a microphone problem rather than some deep indictment of speech recognition.

Right now the workflow is press, hold, record, transcribe, paste.

I haven’t tried live streaming the text while I’m speaking, similar to Wispr Flow.

That’s probably the next experiment.

If your workflow involves transcribing a lot of audio and you want to do it for cheap or almost free, local audio models are absolutely worth trying.

Whisper is the obvious one.

NVIDIA Parakeet is another.

Cohere also has an open-source model called Cohere Transcribe. I haven’t tested it, but it seems promising.

And as I finished writing this post, Qwen 3.8 Omni Flash came out. It looks like another promising multimodal/audio model.

I have not tested it yet, so I have no useful opinion beyond that.

Also, for people who actually know what they’re doing technically, there is an enormous open-source ecosystem here, both for audio and vision.

If you’re technically savvy, you can probably get these models to do a whole lot more than anything I’ve described here.

I am a complete non-technoob.

There are probably things you can build with all of this that I can’t even dream of.

Languages

One giant caveat to everything I’ve said about audio: I have barely tested any of this with non-English speech.

English is the primary language in which I write, so almost all my experience is English.

I did one brief test in Kannada using a Qwen Omni model.

It failed miserably.

That is obviously nowhere near enough testing for me to make some grand statement about multilingual transcription.

So if your primary workflow is Kannada, Hindi, Tamil, Telugu, or some other language, assume none of my audio recommendations have been properly tested for your use case.

The part I still find crazy

The other thing I keep coming back to is how fucking crazy it is that all of this is possible in 2026.

I have absolutely no technical skills.

And I don’t mean this in the fake tech-bro sense where somebody says “I’m not technical” and then starts casually writing Python.

I barely know Excel.

And somehow I am getting these coding tools to build desktop apps, utilities, little workflows, websites, whatever random thing I feel like making.

That is insane.

I know people keep having AGI-this, AGI-that, ghanta-loda debates about all of this.

To say AGI is here would obviously be misleading.

But I also think that for normies some threshold was crossed a while ago. For me it was probably around Claude Sonnet 4.

And I think a lot of people are dramatically underestimating the utility of these tools because they are using the wrong frames to think about them.

Or they’re simply not experimenting enough.

Or they are listening to some mouth-breathing idiot with a very confident opinion about LLMs.

Even though I have been fucking around with these things heavily, I still keep finding an insane amount of utility in them.

They have become a deep part of a lot of things I do, both at work and outside work.

A lot of the things I do now simply would not have been possible for me without them.

Which is why it continues to surprise me when people have these utterly idiotic debates about whether LLMs are useful.

And I still hear “they’re just next-token predictors” or “stochastic parrots” thrown around as if that settles something.

In 2026 this boggles my mind.

Okay, sure. It is a stochastic parrot.

The stochastic parrot transcribes my voice, cleans up my notes, reads 120-year-old books, helps me digitise old letters, turns stock-exchange filings into news, writes code, builds me a Linux desktop app and lets me make all sorts of weird side projects that I would otherwise have absolutely no ability to build.

At some point I stop caring about the insult.

The one thing I would highly recommend is: use these tools heavily.

And not just the chat websites.

Get Claude Code. Get Codex. Fuck around.

Build something stupid.

Automate something annoying.

Take one of those side projects you’ve had sitting in your head for five years because you never knew how to build it and just try.

You’ll be surprised by how much you can do.

But I think the more interesting surprise is the kind of things you suddenly realise you can do.

Especially if you haven’t allowed your brain to die and you still have an imaginative streak, this is a golden age for side projects.

There has probably never been a better time to fuck around and find out.

These tools are really, really, really good.

If you’re not using them, I think you’re missing out.

Disclaimer

This note was voice typed using ChatGPT. I also used ChatGPT to clean up the raw notes, structure them, edit them, and fact-check model names and links.

Why Japan matters

Thu, 17 Sep 2026 00:00:00 GMT

Source: imf.org

IMF on why Japan matters:

The Japanese government bond market is among the world’s largest and deepest and remains an important safe haven for investors when conditions deteriorate. Combined with Japan’s open capital account, sizable international investments, and growing foreign participation—including in mutual funds and hedge funds—this makes for strong global financial ties. As a result, developments in the Japanese debt market can affect other markets, just as shocks in the United States and euro area affect conditions in Japan and beyond. For example, the IMF’s Yan Carrière-Swallow and coauthors show that shocks to Japanese sovereign yields tend to transmit to global sovereign debt markets, with spillovers strengthening as the participation of Japanese investors in the domestic market rises.

That’s why Japan’s fiscal outlook matters not only for domestic policymakers but also for investors and governments around the world.

The lesson from Japan is that debt must be viewed in an appropriate context. Gross debt is an important indicator, but it is not the whole story. Assets, maturities, investor attributes, currency denomination, market structure, and institutional credibility all shape fiscal sustainability.

Reading list on open source AI

Wed, 16 Sep 2026 00:00:00 GMT

Source: interconnects.ai

To read:

Nathan Lambert @natolambert · September 13, 2026 ↗I made a list of the 50+ best articles and resources I’ve read and relied upon in the last few years on open models and open-source AI. Happy reading :)

Hidden chokepoints

Tue, 15 Sep 2026 00:00:00 GMT

Source: bbc.com

One of India’s biggest digital vulnerabilities sits on a six-kilometre stretch of Mumbai’s Arabian Sea coast.

Around Versova, a densely populated neighbourhood on the city’s northwest, at least 13 of ++India’s 18 intercontinental submarine cables++ come ashore within six kilometres of one another.

Together they carry as much as 95% of India’s international bandwidth to Europe, Africa and West Asia.

“This network faces a critical vulnerability due to severe geographic concentration,” says Anwesha Sen of Takshashila Institution, a think tank.

The price of Trump’s stupidity

Mon, 14 Sep 2026 00:00:00 GMT

Source: wsj.com

Donald Trump is a man who lives by a singular philosophy: “Aaj kuchh tufani karte hain” (Let’s do something daredevil today).But sadly, his deep desire to be disruptive is completely at odds with what the world actually needs. Today, the entire planet is paying a steep price for his almost orgasmic fixation on chaos—reflected directly in surging global energy prices.His volatile tendencies have set the entire Middle East ablaze, recklessly destabilizing a region that was already incredibly fragile.We have reached a point where reality is merely parody masquerading as truth. Welcome to 2026.

Iranian officials have said the war won’t end and shipping routes won’t reopen until all of the conflicts in the region are resolved. Gulf states that had hoped for a quick end to the war are now grappling with the realization that the U.S. has pulled them into a conflict that could run for years with no clear strategy for ending it.“This isn’t just a Yemen story,” said H.A. Hellyer, a senior fellow at the Royal United Services Institute for Defense and Security Studies, a think tank in London. “Washington’s approach to Iran has rested on grinding it down through direct military and economic pressure while assuming the wider region stays manageable. That assumption was always fragile.”The chessboard is becoming more complex at a time when the U.S. military is already straining to shepherd tankers through the Persian Gulf and enforce its blockade of Iranian ports.

CNBC · cnbc.com ↗@LCO.1: ICE Brent Crude (Dec'26) - Stock Price, Quote and News - CNBCGet ICE Brent Crude (Dec'26) (@LCO.1:Intercontinental Exchange Europe) real-time stock quotes, news, price and financial information from CNBC.

What does it mean to think in the shadow of AI?

Sun, 13 Sep 2026 00:00:00 GMT

Source: x.com

As always another thoughtful and prescient post by Ruxandra Teslo.

If we want humans at all to be in charge in any way over the world, we need them to continue exercising their intellect and gaining mastery over intellectual domains. If we deplete the institutional and cultural resources that underpin that, in a generation or two our minds will atrophy and no human will be in charge of anything (or the world will be horrendous).

What GDP Doesn’t See

Sun, 13 Sep 2026 00:00:00 GMT

Source: x.com

Yet another interesting debate on the limits of GDP, this time triggered by David Deutsch posting about fixing his dishwasher with ChatGPT, and Jesús Fernández-Villaverde riffing on what that means for GDP, welfare and taxation.

Jesús Fernández-Villaverde @JesusFerna7026 · September 12, 2026 ↗David Deutsch (@DavidDeutschOxf) has a nice post today explaining how he fixed his dishwasher with the help of ChatGPT. We have all been there over the last couple of years. In my case, from fixing a light fixture in my kitchen to repairing a leak in my irrigation system.…

Amen to this:

David Deutsch @DavidDeutschOxf · September 12, 2026 ↗Give up trying to measure knowledge or creativity. Correct the restraints and constraints and brakes and sabotage and punishments instead. rohit @krishnanrohit · September 13, 2026 ↗You're also decreasing GDP when you make a sandwich or mend a shirt. But the creativity which increases the productivity shows in *tomorrow's* GDP, because well collectively make and sell more goods and services.

This is one of those debates that keeps coming back in economics because everybody knows GDP has obvious limitations, and yet nobody has really managed to replace it.

The funny thing about GDP is that for something so crude, it is annoyingly useful.

Lant Pritchett has made this point quite well. Economists obviously don’t think GDP is a direct measure of wellbeing, and he readily acknowledges all its limitations. But, as he puts it:

“GDP per capita turns out to be a handy and available proxy.”

His broader point is that sustained increases in productivity and income tend to be associated with improvements in most of the basic things we actually care about: poverty, health, education, sanitation and so on.

So GDP is simultaneously a terrible measure of lots of things we care about and an extremely difficult measure to beat. That line is mine, by the way, not Pritchett’s.

But its limitations manifest themselves in wonderfully weird ways.

Suppose a country is at war and starves large parts of its civilian economy of resources to ramp up the production of missiles, shells, guns, tanks and every other kind of armament. All of that military production counts towards GDP. Measured GDP can therefore rise even while ordinary people are consuming less and, quite plausibly, becoming worse off.

This isn’t some new critique invented by people arguing about AI. Simon Kuznets, one of the fathers of modern national-income accounting, was wrestling with almost exactly this problem nearly 90 years ago.

In his famous 1934 report to the US Congress, he warned:

“The welfare of a nation can, therefore, scarcely be inferred from a measurement of national income.”

There is a useful discussion of that warning and the evolution of national-income accounting in this NBER paper.

But Kuznets went much further than the usual “GDP is not welfare” disclaimer.

In a 1937 discussion published by the NBER, he wrote:

“It would be of great value to have national income estimates that would remove from the total the elements which, from the standpoint of a more enlightened social philosophy than that of an acquisitive society, represent dis-service rather than service. Such estimates would subtract from the present national income totals all expenses on armament, most of the outlays on advertising, a great many of the expenses involved in financial and speculative activities, and what is perhaps most important, the outlays that have been made necessary in order to overcome difficulties that are, properly speaking, costs implicit in our economic civilization.”

You can read the original NBER chapter here.

That is a remarkable passage for something written in 1937.

Kuznets wasn’t saying that armaments, advertising or financial activity somehow weren’t economic activity. He was asking a harder question: does every dollar of measured economic activity represent an equivalent increase in human welfare?

Clearly not.

There is also a historical irony here. Part of what makes GDP look strange as a measure of human progress is that national accounting was shaped partly by a very different practical problem: how much can an economy produce? During a war, the capacity to produce tanks, aircraft, ammunition and ships matters enormously even if household consumption is being squeezed.

Diane Coyle, in her excellent GDP: A Brief but Affectionate History, gives another wonderfully absurd example.

A widower employs a housekeeper and pays her a salary. Her work is market activity and therefore contributes to GDP.

He marries her.

She continues doing exactly the same household work, except now she isn’t being paid for it.

GDP falls.

Nothing necessarily happened to the quantity of useful work being done, and their welfare need not have fallen by a rupee. An activity simply crossed the accounting boundary between market production and unpaid household work. There is a nice discussion of the example in this IMF review of Coyle’s book.

Which brings us back to AI and David Deutsch’s dishwasher.

In the old world, your dishwasher breaks. You call a technician. He fixes it. You pay him ₹2,000. That transaction is part of measured economic activity.

In the new world, the dishwasher breaks. You show ChatGPT the problem, it figures out what’s wrong, tells you what to do, and you fix it yourself.

You have a functioning dishwasher. You have saved ₹2,000. You may even have learned something.

But the repair transaction has disappeared.

Strictly speaking, GDP doesn’t necessarily fall because David Deutsch repaired his dishwasher. GDP is simply lower than it otherwise would have been if he had paid someone to perform the same service.

And this is where I think AI makes an old measurement problem much more interesting.

The same thing already happens with books. If Jesús Fernández-Villaverde buys a book, the production and sale of that book are captured in GDP. The author gets paid, the publisher gets paid, Amazon gets its cut, etc.

What isn’t captured is the full value of the knowledge he acquired.

Maybe a $20 book teaches him something that saves him hundreds of hours over his lifetime. Maybe an idea changes an important decision and makes him enormously better off. GDP doesn’t suddenly record the value of that knowledge merely because it originated in a market transaction.

AI could make this wedge between what we pay and what we get much, much larger.

For a few dollars a month, or perhaps eventually almost nothing, you could have something that helps you repair appliances, learn maths, understand a difficult paper, debug code, translate documents, learn history, analyse data, write, think through decisions and do dozens of things for which you previously needed either another person or years of accumulated expertise.

A lot of the value might therefore show up as higher consumer surplus rather than proportionately higher measured GDP.

There is actually an interesting attempt to measure precisely this sort of gap.

Erik Brynjolfsson and his co-authors have spent years trying to estimate the value of digital goods that conventional national accounts largely miss. The basic idea is simple: instead of asking what you paid for something, ask how much you would have to be paid to give it up.

In one set of experiments, respondents valued access to search engines at a median of about $17,530 a year, email at $8,414 and digital maps at $3,648. Because many of these services are free at the point of use, much of that value is consumer surplus rather than expenditure captured directly in GDP. The study is here.

Brynjolfsson and his co-authors later proposed a supplementary measure called GDP-B, with the “B” standing for benefits: not a replacement for GDP, but an attempt to account for welfare created by new and free goods.

And a larger 2023 study covering ten digital goods across thirteen countries estimated more than $2.5 trillion a year in consumer welfare, roughly 6% of those countries’ combined GDP. Interestingly, it found proportionately larger welfare gains among lower-income people and lower-income countries. That paper is here.

This raises a broader question that goes well beyond AI.

What is the economic value of all the things we get for free?

Greg Ip @greg_ip · September 12, 2026 ↗That AI might make us better off in ways GDP can't show is not an interesting or important insight. It has ALWAYS been true that much, even most, welfare improvement is unpriced. The lives saved by childhood vaccines, for example, never shows up in GDP (while the health…

Take Google Search, YouTube, email, maps, Wikipedia or any number of free digital services. Obviously the companies behind many of these products are not economically invisible. They employ people, buy servers, sell advertising, make profits and pay taxes, and all of that shows up somewhere in the national accounts.

But that is not the same thing as measuring the value these products create for their users.

And yes, big technology companies create harms too. Algorithmic manipulation, privacy loss, effects on communities and politics, mental-health concerns and all the rest are real questions. If you wanted to do some impossible moral accounting of the whole thing, you would have to put the benefits on one side and the negative externalities on the other.

My point is narrower: the value of a free service to its users is not zero merely because its price is zero.

Now take the idea one step further and think about open-source software.

Linux is free, and in one form or another it sits underneath an absurd amount of the world’s digital infrastructure. The same is true of PostgreSQL, Python, Git, OpenSSL and thousands of other projects.

The economic activity built using open-source software eventually shows up everywhere: cloud computing, banks, websites, software companies, salaries, profits, taxes and so on.

But where exactly is the value of Linux itself?

Imagine two otherwise identical worlds.

In one, every company pays “Linux Corporation” $100 billion a year to use Linux.

In the other, exactly the same software does exactly the same job but is freely available as open source.

Measured market transactions would be larger in the first world. It isn’t obvious that humanity would be better off.

In fact, we would probably say the opposite. Making an enormously useful input essentially free has made society richer by freeing resources for other things.

This is perhaps another version of the housekeeper problem. Diane Coyle’s housekeeper moves from paid work to unpaid work. Linux moves something from a proprietary market transaction into a commons. David Deutsch moves dishwasher repair from a paid service into something he can do himself with freely available knowledge.

AI could cause millions of activities to cross that boundary at once.

There is another place where critics of GDP are on even stronger ground: the environment.

Imagine a factory that burns enormous quantities of coal and produces a million widgets. The widgets are sold, workers get paid, the company earns profits and the government collects taxes. All of that economic activity shows up in GDP.

What doesn’t automatically appear as a deduction is the carbon released into the atmosphere, local pollution or the damage those emissions may impose on people somewhere else or decades later.

Or take an extreme example. Clear a large part of the Amazon, sell the timber, build factories on the land, employ thousands of people and raise their wages. GDP could increase substantially. Material living standards in the region might improve too.

But something else has happened at the same time: an enormous stock of natural capital has been destroyed.

The accounting analogy here is useful. GDP is a little like looking at a country’s income statement without looking closely enough at its balance sheet. You can make this year’s income look better by consuming an asset.

The UN itself uses almost exactly this forest example when explaining its System of Environmental-Economic Accounting: cutting down all of a country’s forests could increase GDP in the short run through timber production while destroying natural wealth and imposing enormous long-run costs.

Economists and statisticians are perfectly aware of this problem. The UN now has a formal framework for ecosystem accounting. There are also attempts at “Green GDP”, which deduct resource depletion and environmental degradation from conventional economic measures. The World Bank publishes Adjusted Net Savings, which takes net national savings, adds education expenditure, and subtracts things such as energy depletion, mineral depletion, forest depletion and carbon and particulate pollution damages.

But once you try to do this properly, the measurement problems become brutal.

What is a rainforest worth? What is the value of an extinct species? What number do you put on a stable climate, clean air or an ecosystem that prevents floods? What discount rate should we use for damage imposed on someone born seventy years from now?

There probably isn’t a neat answer.

And then there is one final weirdness. Environmental damage can sometimes generate measured economic activity when we cause it and more activity when we repair it. Pollution creates output. Treating some of the illnesses associated with pollution creates output too. A natural disaster destroys homes; rebuilding them generates output.

None of this means GDP is fraudulent or useless. It means GDP isn’t keeping a moral ledger in which it records factory output on one side and subtracts damaged lungs, depleted aquifers or lost forests on the other.

There have been endless attempts to go “Beyond GDP”. A 2024 review in The Lancet Planetary Health went through 65 different alternative metrics developed across roughly five decades of work. Despite all of these proposals, their integration into actual policymaking and public discourse remains limited.

That, in a way, is the punchline.

I realise some of this can sound like the usual rant about the inadequacy of economics or the absurdity of reducing everything to GDP. That isn’t really my point.

GDP is useful. Income growth is useful. GDP per capita is useful. Part of the reason these measures have survived for so long is that they often do a remarkably good job of summarising a vast amount of economic reality in a single number.

The problem, I think, is that most of us don’t stop to ask what exactly is being counted, how it is being counted and, perhaps most importantly, what isn’t being counted at all.

There is a famous line, usually misattributed to Einstein, but apparently due to the sociologist William Bruce Cameron:

“Not everything that can be counted counts, and not everything that counts can be counted.”

The attribution is discussed here.

In a peculiar way, I think we all suffer from a kind of economic blindness.

When we talk about human progress, we naturally reach for the broad numbers that are easiest to see: GDP, GDP per capita, income growth, productivity and so on. Most of the time these are useful shortcuts. But they are still shortcuts.

They can tell us an enormous amount about market production. They are much worse at telling us how much knowledge has become freely available, how much consumer surplus has been created, how much unpaid work is being done, how much natural capital has been destroyed, how much pollution has been imposed on somebody else, or whether all of this has actually translated into better lives.

This doesn’t mean GDP has served its purpose and should be discarded. If anything, it means we have reached the limit of what we should ask it to tell us.

At the very least, economic progress should probably be thought of as an ensemble rather than a scoreboard: production and income, yes, but also health, education, distribution, leisure, environmental wealth, sustainability, consumer welfare, perhaps access to knowledge and other things that are becoming increasingly important in a digital economy.

The difficult part, of course, is that the moment you move beyond GDP, measurement becomes much messier. There is no obvious exchange rate between a forest and a factory, between an extra year of healthy life and another percentage point of income, between free access to human knowledge and the harms created by the platforms providing it.

Perhaps that messiness is precisely the point.

GDP survived partly because it gives us an extraordinarily clean answer to a relatively narrow question.

Human progress is not a narrow question.

Adam Butler @GestaltU · September 5, 2026 ↗We desperately need new indices to help measure the evolution of human welfare over time. The progenitor of GDP specifically warned it should never be used to measure gains in welfare. Worse, GDP is always ‘deflated’ to account for inflation, using indices like CPI and PCE.…

This short note was written with the help of ChatGPT, based on a back-and-forth about the economics, the historical context and the underlying sources.

There is still Alpha to be had in good training data.

Sat, 12 Sep 2026 00:00:00 GMT

Source: huggingface.co

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], without algorithmic modifications beyond well-established practices. Instead, our efforts are concentrated almost entirely on what the model is trained on rather than how it is optimized: we invest in large-scale, automated pipelines for data synthesis and environment construction. Concretely, the pipeline (i) synthesizes diverse, verifiable training tasks together with their reference solutions and reward signals, (ii) procedurally constructs and scales interactive Agent environments in which trajectories can be collected and evaluated at low cost, and (iii) applies rigorous filtering, deduplication, and difficulty calibration to ensure data quality and curriculum balance. We find that, under a fixed and unremarkable optimization procedure, systematic improvements in the scale, diversity, and verifiability of synthesized data and environments account for essentially all of the observed gains. This observation echoes a broader lesson: at the current stage, the marginal return of engineering the data and environment pipeline substantially exceeds that of algorithmic novelty in post-training.

DeepSeek V4.1 Flash Vision is surprisingly good

Fri, 11 Sep 2026 00:00:00 GMT

Source: deepseek.com

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 its vision capabilities should be really useful for modern documents and PDFs, where it can chew through large volumes of material at ridiculously low cost.

DeepSeek currently prices its V4 Flash vision API like this, per 1 million tokens:

Rate| Input, cache hit| Input, cache miss| Output Off-peak| $0.007| $0.22| $0.66 Peak| $0.014| $0.44| $1.32

Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday. Everything else is off-peak, at half the peak rate.

Images are converted into tokens and billed as input. DeepSeek says images use up to 384 tokens each, which makes the economics particularly interesting for large-scale document processing.

I think this makes DeepSeek V4.1 Flash a viable model for large-scale digitization projects. It’s cheap enough that workloads involving tens or hundreds of thousands of pages start looking much more realistic.

I’m a little excited about this because I’ve had a bunch of large-scale digitization project ideas sitting around, and costs like these make some of them much more viable.

“DeepSeek API documentation” (https://api-docs.deepseek.com/)

“DeepSeek API pricing” (https://api-docs.deepseek.com/quick_start/pricing/)

We are the stories we tell

Tue, 08 Sep 2026 00:00:00 GMT

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:

Jim O'Shaughnessy @jposhaughnessy · September 7, 2026 ↗"To poison a nation, poison its stories. A demoralised nation tells demoralised stories to itself." ~Ben Okri

Feelings that defy language

Mon, 07 Sep 2026 00:00:00 GMT

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, pass things down, argue, explain, warn, comfort, persuade, remember.

And despite the fact that we’ve been doing this for God knows how long, it still surprises me that there are so many feelings and emotions that seem to run up against the limits of language.

There’s this beautiful line by Tracy K. Smith that I keep coming back to:

“Poetry is the language that sits really close to feelings that defy language.”

And I think about this quite a bit because I suffer from debilitating sleep issues, which have gotten worse in recent months.

The thing with not getting enough sleep is that it doesn’t just make you tired.

It messes with your emotional balance. Your baseline changes. And the longer it goes on, the worse it gets.

There are days when I’m unable to think clearly. There are days when I’m unable to express myself clearly. And there are days when I don’t even know how to describe what exactly is happening inside my head.

This is quite scary, at least to me.

Whenever I speak to friends or colleagues about this and they ask me, “How do you even function without sleep?”, I end up using metaphors.

I feel like a zombie.

I feel underwater.

I feel like I’m moving through sludge.

And I don’t say these things because I’m trying to make it sound dramatic. I say them because I genuinely don’t know a clean way to describe how I feel using plain old English.

Maybe I just don’t know enough words. But I don’t think that’s the problem.

This is just one example.

I have a lot of friends who struggle with mental health issues, and sometimes when I ask them how they are feeling, I can see the struggle on their face.

They are trying to find the words.

And often they can’t.

Forget explaining what caused the feeling. Sometimes it is hard enough to describe the feeling itself.

I relate to this.

There have been times when I’ve felt so blue that the only way I’ve been able to describe it is as drowning, with the weight of the water increasing every minute.

Or it feels like somebody is pushing me underwater with a force much stronger than me, and I’m unable to pull myself out.

Of course, neither description is quite right.

But “sad” isn’t right either.

“Low” isn’t right.

Even words like “anxious” or “depressed” can sometimes feel too broad, too clean, or too inadequate for whatever messy thing is actually happening inside your head.

I love words. A ridiculous amount of my life revolves around the written word and the spoken word.

So I tend to pay attention to how people speak. The words they choose. The phrases they invent. The metaphors they reach for. The way they sometimes pause halfway through a sentence because whatever word they were looking for simply isn’t there.

And I’m reasonably certain that we are doing this all the time.

We are straining at the limits of language.

I don’t think this has much to do with vocabulary or grammar or whether somebody is articulate enough.

I think the full spectrum of human experience is simply much larger than the meanings that our humble words can comfortably contain.

Maybe this is one reason why we invented metaphors.

Maybe this is one reason poetry exists.

Maybe this is also one reason talking about mental health is so difficult. Before another person can understand what you’re feeling, you first have to somehow convert this shapeless thing happening inside you into words.

And that conversion is imperfect.

Sometimes wildly so.

Which is a fascinating paradox, at least to me.

Language is the thing that lets us share our inner lives with one another.

And yet our inner lives are often much larger than language.

Sweden’s crime problem

Mon, 07 Sep 2026 00:00:00 GMT

Source: ft.com

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.”

India after Gandhi

Sun, 06 Sep 2026 00:00:00 GMT

Source: amzn.in

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 up, and it’s really good—riveting, to say the least.

Reading the book made me realize that if most Indians knew how this great nation came to be, how it persisted, and how it had to navigate perilous realities in its early days, they would have a stronger sense of pride. They’d have a stronger sense of togetherness and camaraderie with their fellow Indians, instead of being the sheep that we are in a lot of instances, played by the powers that be to further their own narrow political interests.

The book is also a reminder of the importance of knowing history, because it shows just how small our thoughts can be, and the emptiness of a lot of our ideas and ideologies.

Visualizing the OpenAI-Hugging Face incident

Sat, 05 Sep 2026 00:00:00 GMT

Not bad

Generative History @HistoryGPT · September 3, 2026 ↗To illustrate a talk I’m giving, I asked ChatGPT Image 2 to depict the hugging face incident from the agents’ perspective and what I got was a strange creation myth:

Here’s what I got:

Research on the wealth effect

Fri, 04 Sep 2026 00:00:00 GMT

Source: ft.com

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.

The bear case for AGI

Fri, 04 Sep 2026 00:00:00 GMT

Source: x.com

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:

Philippe Lemoine @phl43 · September 4, 2026 ↗This is a good post and I think people should take seriously the points it raises and think more about them. I would just like to make additional comments on the observation that LLM intelligence seems very different from human intelligence, because it's something I have been…

I got some links for you

Thu, 03 Sep 2026 00:00:00 GMT

File this under mad genius:

Aristo @aristomarinetti · September 2, 2026 ↗The same person invented Aspirin and heroin 11 days apart .

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:

Ethan Mollick @emollick · May 12, 2023 ↗Digital humanities is about to be wild. I pasted in the Catalog of Ships from the Iliad, and asked GPT with Code Interpreter to map the origins of the Greeks. With hiccups, it extracted solid information from the text, found modern locations & started building an interactive map

Been testing flash and it seems very promising:

Demis Hassabis @demishassabis · September 2, 2026 ↗Introducing Gemini 3.8 Flash, another upgrade in under a month! And the new 3.8 Flash Cyber pushes the frontier of cyber defense. Relentless progress 🚀 More on the models & our cyber security work in the blog - happy building! https://t.co/C5h14qANIV

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

Kyle Chan @kyleichan · September 3, 2026 ↗This is quite stunning. China’s vast electric vehicle fleet can be flexed during an oil supply shock and offset the equivalent of the UK’s entire oil consumption.

Agentic payments are coming to UPI

Tue, 01 Sep 2026 00:00:00 GMT

Source: reuters.com

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/

AI is NOT guzzling water!

Mon, 31 Aug 2026 00:00:00 GMT

Source: lnkd.in

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.

Omawrite

Mon, 31 Aug 2026 00:00:00 GMT

Source: github.com

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

Sonder

Sun, 30 Aug 2026 00:00:00 GMT

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 defy language.”

It’s a phrase that I often keep thinking about.

Whenever we try to convey something and run headfirst into the limits of the language that we use, we often try to find shortcuts. Somehow we muddle through. Somehow we make do with the words that we know.

But some special people rebel against these limits and try to create new words.

John Koenig seems like one of those people.

Koenig is the creator of The Dictionary of Obscure Sorrows, a project devoted to inventing words for emotions and experiences for which we don’t quite have words.

The reason why I know of John Koenig is because of one word, a neologism that he coined called sonder.

I first came across the word in this Substack post by Olivia Jarrell, and I’ve been thinking about it ever since.

It has stuck in my mental craw because it’s such a brilliant word. It’s deep and capacious.

Sonder

The realization that every random person, every stranger, is living a life as vivid and complex as your own.

It is the realization that behind every ordinary face lies a person who has his or her own ambitions, routines, worries, inner monologues, fears, hopes, dreams and anxieties.

Beautiful word, isn’t it?

Sonder.

There wasn’t a single trigger for writing this post. In a way, this has been many, many, many months in the making.

The reason why I’ve been thinking about sonder is because of some of the conversations that I have been having with people really close to me.

These are people I have known for a long time, people I otherwise thought I knew. And yet, some of the conversations I’ve had with them recently spectacularly shattered the illusion that I knew these people, or that I was good at judging people.

The kind of vivid interiority these people possessed, I could scarcely have guessed, let alone imagined.

These conversations made me realise that people are far more complex, far more layered, far more dense, far more amorphous than I otherwise believed.

As obvious as it sounds, it’s something we often forget.

Or, I don’t want to speak for others. It’s something that I often keep forgetting.

I often have very simplistic notions of a lot of the people I come across, and I constantly keep forgetting that every person has a rich inner life, a rich life of their own.

In writing about this, it does seem like we engage in a kind of lossy compression when it comes to comprehending people.

And lossy compression is good.

If you had to sonder every random stranger you came across, your brain would probably explode. You’d have a nervous breakdown.

Evolution endowed us to be a little judgmental. It probably knew this, which is why we are the kind of people who jump to conclusions and coast on sense perceptions and fleeting impressions.

But lossy compression is not always good.

Sometimes you need to see the raw files to understand someone.

And that is probably more often the case with the people you love and the people closest to you.

We need to ingest the full raw data files of these people, because lossy compression leads to lossy understanding.

I think this is also why the word sonder resonates with me so much, and why I haven’t forgotten it unlike many of the other new words that I’ve been trying to learn.

It speaks to this inner desire in me: the desire not to forget that there is far more to people than the outer appearances on which we base our judgments and construct narratives about the people we come across.

Behind every person’s face lies an entire universe to which we are not privy. They’re all stumbling and struggling through their lives, trying to deal with all their hopes, anxieties, fears and countless other emotions for which there probably aren’t good words in the English language.

Every person is dense and fibrous in that sense.

Nobody is simple.

And I’m pretty sure nobody has a simple life either.

There is another side to this realisation.

Most people we bump into in the course of life barely register or leave a mark. We tend to forget them because we are all stuck in our own heads.

I mean, I don’t know about you, but at least I am.

We are all stuck in our heads, making up stories, writing autobiographies, writing this continuous, never-ending autobiography of our lives, and constantly authoring things about who we are, what we need to do, why we do the things that we do, where we have to go, and everything else on the fly.

Once you appreciate the meaning of the word sonder, I think we shrink a little in our own heads.

We become a much smaller part of the picture because everybody else is preoccupied with navigating their own rich, vivid, complex, messy, dark, wonderful, delightful and thorny universes.

They don’t have time to think about us.

It’s a little similar to the spotlight effect. We often think that we are the centre of the world when, in reality, nobody gives a shit because they’re all busy with their own shit.

Now, sondering doesn’t necessarily turn you into one of those clichéd people who suddenly decides that life is beautiful, everybody is complex and therefore we should all be kind and understanding.

I don’t know.

It’s still entirely possible to sonder and be a murderous sociopath.

And strangely, the fact that even a murderous sociopath can sonder gives me some hope.

Not because I want more murderous sociopaths, obviously. But wouldn’t you rather have a sondering murderous sociopath who is aware that other people have rich, vivid inner lives than an unsondering, sonder-ignorant murderous sociopath?

Sorry for the stupid and juvenile digression.

But that’s probably the point I’m trying to get at.

Realising that other people have their own rich universes doesn’t magically make you good. It doesn’t necessarily make you kinder, wiser or more understanding.

At the bare minimum, it helps you see people differently.

And isn’t that enough?

The rise of agentic commerce

Sat, 29 Aug 2026 00:00:00 GMT

Source: x.com

Grok bot can now buy things for you:

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

When everyone take the fastest route

Sat, 29 Aug 2026 00:00:00 GMT

Source: nature.com

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. Individuals optimizing for things that are good for them can often lead to collective insanity.

Or put another way, you can have a system where nobody is doing anything stupid. Everybody is doing what they think is sensible, but collectively it leads to a stupid outcome.

What I’m referring to here are ideas like externalities and the tragedy of the commons. At an individual level, an action might be perfectly rational, but once enough people behave the same way, the outcome for the overall system can be worse.

Anyway, I didn’t know that GPS systems like Google Maps and Waze essentially optimize at the level of the individual driver. They try to get you from point A to point B through what they think is the optimal route. But when you aggregate all these individually optimal decisions, you can end up sending a lot of traffic onto the same optimal routes, which then creates congestion and gridlock.

Air-traffic networks, by contrast, use network-aware routing, where routing decisions take the health and efficiency of the overall network into account.

What Google did is really clever. They took about 100 congested road segments across 10 U.S. cities, and then changed the route recommendations for about 2% of the trips in those regions. So instead of recommending the same route that Google Maps would typically suggest, which would have meant sending more drivers onto the same congested roads, it nudged a small fraction of them towards relatively underused roads with lower traffic, while keeping their travel times as close as possible to the fastest route.

In effect, traffic was taken away from a small number of already choked arteries and dispersed across a much larger number of roads that had more spare capacity.

And that tiny intervention was enough to produce statistically meaningful improvements: higher speeds on the targeted roads, slightly lower travel times across affected trips, and estimated reductions in carbon emissions.

Very interesting paper.

Hat tip to Yohan:

Yohan @yohaniddawela · August 28, 2026 ↗For six months, Google Maps sent a slice of drivers in ten American cities down deliberately slower routes. About 30 seconds slower on average, and under 2% of trips were touched. The experiment ran on roughly 100 of the most congested road segments in each city, switching on…

How does the preference for sons impact the mental well-being of women in rural India?

Fri, 28 Aug 2026 00:00:00 GMT

Source: aces.illinois.edu

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 included the youngest woman. They collected socioeconomic and demographic data, as well as information on mental health using a survey that screens for the risk of depression and anxiety.

They found the strongest effects on the mother-in-law’s mental health; on average, they had a 15% lower risk of depression or anxiety when their daughter-in-law had a firstborn son. These results are equivalent to other findings measuring the effect of inheritance reforms or pension programs that provide added security for older women, Herrera-Almanza noted.

The researchers did not find any effects for the daughters-in-law overall, but when they broke down the results by age group, they found that risk of mental health consequences increased as women grew older.

“For the mother-in-law, having a grandson is a relief at any age. For the daughter-in-law, as she gets older and her reproductive horizon is closing, it is much more stressful not to have a son,” Herrera-Almanza said.

Small language models are all we need

Fri, 28 Aug 2026 00:00:00 GMT

Source: calv.info

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.

Google’s Pragmatic Pivot?

Thu, 27 Aug 2026 00:00:00 GMT

Source: blog.google

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 reviews—at least going by the vibes on Twitter more than anything else.

Personally, I’ve never used Google models for coding all that much in my vibe coding experiments. My usage has been very limited because they more or less suck for someone like me who’s not a coder. Tools like Claude Code and Cursor often just get what you want with some poking and prodding. Gemini, on the other hand, is just really bad on most things when it comes to coding.

That being said, Gemini still has a lot going for it on vision tasks. I’ve found it to be the best and most reliable model for historical archival projects—like my projects digitizing old books at akshara.ink and collections of letters at Paper Lanterns—and even for modern documents. Even the cheapest models like Gemini 3 Flash Preview and Gemini 3.1 Flash-Lite are really good at this, and they’re very cost-effective. While there is some chatter online about the latest Flash series models regressing a bit on vision tasks, they still perform really well.

There has also been a lot of talk that Google seems to have ceded the race to build the latest frontier models, and instead is focusing on areas where it can actually add value—especially things like document processing in the enterprise that rely heavily on vision.

Today, they also launched the Gemini 3.5 Transcribe model. I put it to the test, and it’s quite effective. Voice has become my primary mode of writing and typing; I’ve built myself several voice-first productivity utilities, including a personal journaling app. For a long time, Gemini 3.1 Flash-Lite has been my go-to workhorse model for that. Now that I’m playing around with 3.5 Transcribe, it seems pretty good, and I’m planning to switch it across all my utilities.

At the same time, I saw two more things that make me lean further into this theory: that Google has decided not to play the game of building the biggest model, but rather focus on niches where it can monetize quickly with cheap workhorse models that are easy to adopt.

First was a tweet about GlucoFM, a foundation model for continuous glucose monitoring:

GlucoFM (Continuous Glucose Monitoring Foundation Model) A lightweight, self-supervised model developed by Google Research for continuous glucose monitoring (CGM). Instead of treating glucose data as a single stream, it uses a dual-stream architecture that separates slow background metabolic baselines from short-term transient spikes (like meals, exercise, or sensor noise). This allows it to predict downstream metabolic outcomes—including diabetes risk, insulin resistance, and beta-cell dysfunction—without requiring large-scale, heavily labeled datasets.

Then I saw a post on LinkedIn about Google DeepMind rolling out two models for Indian agriculture:

Agricultural Foundation Models (AnthroKrishi / India & Global Deployment) Two landscape-scale AI models built by Google’s AnthroKrishi team to turn remote sensing and satellite data into farm-level digital public infrastructure:

Agricultural Landscape Understanding (ALU): Automatically maps individual field boundaries, farm sizes, and water bodies from satellite imagery (now integrated directly as a layer in Google Earth).

Agricultural Monitoring & Event Detection (AMED): Tracks in-season crop activity, sowing/harvest timelines, and health markers across 15-day refresh cycles to power credit scoring, crop insurance, and state-level farm advisories (like Telangana’s ADeX and Karnataka’s water management).

So it seems credible that Gemini has decided not to play the game of burning billions just to win the frontier model war, and instead is focusing on areas where it has built deep, practical capabilities. Or at least, that’s my reading as an amateur observer of what AI companies are up to.

Seek not more answers but better questions

Wed, 26 Aug 2026 00:00:00 GMT

Source: kiriakakis.net

What a beautiful thought: seek better questions.

sunil pai @threepointone · August 25, 2026 ↗.@dillon_mulroy shared this wonderful story with me https://t.co/mq1hJlSdIq

Chinese models are the most popular AI models by token volume

Tue, 25 Aug 2026 00:00:00 GMT

Source: scmp.com

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.

On AI doom

Mon, 24 Aug 2026 00:00:00 GMT

Source: x.com

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

Gold is a hedge against debasement

Sun, 23 Aug 2026 00:00:00 GMT

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 much any CPI measure and look at its correlation with gold—it’s weak, to say the least.

For the longest time, this left me confused about what exactly gold was doing in a portfolio. It was clearly doing something. Gold has diversification properties. Adding it to a portfolio of equity and debt can improve the risk-return characteristics of the portfolio. But I didn’t have a satisfying answer for why.

Then I came across something from Corey Hoffstein—I forget whether it was a post or a tweet—that gave me a much better way of thinking about it.
Gold is better understood as a hedge against debasement, rather than a hedge against inflation. Anyway, here’s an excerpt from this post on Corey’s site:

Gold’s tendency to move opposite the dollar comes from where it sits, outside the system that issues it and beyond the reach of any government or central bank. Bitcoin’s link to the dollar runs the same direction but far looser, so its claim rests on that shared independence rather than on any price relationship. Both are non-sovereign currencies, money no government issues and no central bank can expand.

That can sound like an inflation hedge, but the claim is narrower. It may provide no protection when prices rise for reasons unrelated to money, like an oil or food shock. What it guards against is debasement, the value a currency loses when its supply expands or its credibility erodes.

Corey Hoffstein 🏴‍☠️ @choffstein · August 23, 2026 ↗whenever i see someone talking about the downfall of the US hegemony, erosion of trust in the federal government, and the inevitable debasement of the US dollar...

I recently recorded a podcast with Abid on the topic as well:

YouTube · youtu.be ↗The Real Truth About Buying Gold in India! | Cash & Copium Ep 2Zerodha Subtext on YouTube

Pep Guardiola on players he admires

Sat, 22 Aug 2026 00:00:00 GMT

Source: youtu.be

Pep’s favorite football players:

Welcome to AI slopulism

Fri, 21 Aug 2026 00:00:00 GMT

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 concerned than excited about AI. This is up from 31% in 2021.

As for the share of people who say they are more excited than concerned, it has fallen from 25% in 2021 to just 11% today.

So, things aren’t looking good.

And then there is the issue of data centers.

Heatmap has been asking Americans whether they would support the construction of a data center near where they live. And the answers are, again, depressing.

Hunter📈🌈📊 @StatisticUrban · August 20, 2026 ↗Insane decline. Robinson Meyer @robinsonmeyer · August 20, 2026 ↗We’ve got exclusive new polling on local data center development at @heatmap_news. Over the past year, we’ve asked Americans whether they would support or oppose a data center being built near where they live. We haven’t changed the wording. When we first polled the question…

A year ago, Americans were roughly evenly split. Today, 75% say they would oppose one.

Across political lines, support has basically cratered. Data centers are now 43 points underwater among Republicans, 65 points underwater among independents, and 75 points underwater among Democrats.

One of the tweets that triggered me into voice-writing this post was this one from Théo Jaffee at Google DeepMind, who called this the “anti-data-center psychosis,” which seems like a pretty good encapsulation of the attitudes of a lot of Americans right now.

LM Sacasas @LMSacasas · August 21, 2026 ↗I've been trying to find an apt way of describing the tone that characterizes so tech-booster rhetoric (see below), and I think I've landed on "hysterical condescension."

But what is also interesting to me is the divergence in attitudes towards AI in the US and elsewhere.

At least among parts of the educated elite in the US, there seems to be a lot of widespread pessimism about AI. And while the data is still noisy, we are perhaps starting to see the initial sparks of things going badly for some people.

Whether the old truism that technological shifts ultimately create more jobs than they destroy holds this time remains to be seen.

Just because it has been true in the past doesn’t necessarily mean it has to be true in the future.

And even the smooth statement that “technological shifts create more jobs than they destroy” ignores all sorts of compositional effects and adjustment issues.

A job can disappear in one industry, in one part of the country, for one kind of worker, while another job gets created years later somewhere else, requiring completely different skills. Both can eventually become part of one reassuring aggregate statistic.

When large parts of the labour force are turned into one aggregate number, we tend to miss the lived realities underneath it.

But anyways.

At the same time, I’ve been reading quite a bit about attitudes towards AI in China, and they seem surprisingly positive.

Bloomberg had an interesting piece on this recently, looking at why AI optimism seems to be so much higher in China than in the US.

Maybe some of this gets exaggerated. Cross-country polling is messy, and there are obvious reasons to be careful about comparing attitudes between China and the US.

But nonetheless, things don’t seem anywhere near as rotten in China as they currently do in America.

And that is very interesting considering Chinese AI labs themselves are making progress at a remarkable pace. It isn’t as if the potentially disruptive technology is happening somewhere else. China is right in the middle of the same technological shift.

The Chinese government also seems, at least from what I’ve been reading, more proactive and sanguine about trying to anticipate the effects of AI and using policy to forestall some of the potential downsides.

So why are Americans so pessimistic about AI?

I don’t have a good answer.

It is actually very hard to suss out what exactly is animating this backlash.

I don’t think this is simply AI populism. It is probably a much sloppier mix of things. AI slopulism, if you will.

The backlash against data centers and AI also feels like it may be a symptom, or an instantiation, of a deeper economic, social, and political malaise rather than something as straightforward as: people heard AI will take their jobs, and now they hate AI.

There are probably a lot of things getting compressed into the word “AI.”

Distrust of Big Tech. Anger at billionaires. Housing. Power prices. Water. Jobs. Economic insecurity. Social media. A generalized feeling that technological change keeps happening to people rather than for them. And probably many other things.

So I want to be careful about monocausal reasoning here.

That being said, I do think one reason for the backlash is the AI companies themselves.

In the initial years of the AI boom, some of the loudest people building these systems spent an extraordinary amount of time shouting and screaming at every opportunity that AI would automate vast swathes of jobs, devastate parts of the middle class, radically reorder society and, at the extreme end, perhaps even become a civilization-ending technology.

And then they kept making the technology better.

Listen to Dario Amodei for three minutes, and it’s hard to come away feeling particularly hopeful about the future of humanity.

YouTube · youtube.com ↗Anthropic CEO on Safety, Job Displacement and Anthropic's $350B Valuation | WSJWSJ Events on YouTube

If the people building AI keep telling everybody, over and over again, that vast numbers of jobs are going to disappear and that large numbers of people could be economically displaced, why would it be plausible to expect Americans to have a positive view of AI?

In that sense, I think the AI companies have done a phenomenally good job of shooting themselves in the foot and painting themselves as the villains.

Again, I’m not saying the shrill rhetoric from the heads of AI companies is the reason Americans are turning against AI. It isn’t.

But it does feel like a definitive factor.

And all of this feels especially consequential because we seem to be living through an age of extremes, in almost every sense of that phrase, including politics.

My worry is that the extreme reaction that some Americans are developing towards technology, including AI, will lead to first-order and second-order reactions that are very hard to foresee but will nonetheless produce some depressing and perhaps tragic outcomes.

If I put on my sci-fi hat for a second, I don’t think it is outside the realm of possibility that we eventually start seeing violence directed at data centers and other AI infrastructure.

Andy Masley @AndyMasley · August 20, 2026 ↗In all seriousness I think this represents both a disturbing society-level inability to process basic facts that involve numbers, a lurch toward degrowth on the left and right, and a failure of mainstream journalism to go beyond just nodding along to the "concerns" of everyday…

We have seen versions of this dynamic before with other causes. Protest movements begin with demonstrations and symbolic stunts, and at the fringes, some people escalate.

And in the case of AI, we may already have had a forewarning sign.

In April, a man allegedly threw a Molotov cocktail at Sam Altman’s home and was arrested shortly afterwards while allegedly attempting to break into OpenAI’s headquarters carrying kerosene, a lighter, and an anti-AI manifesto.

I don’t want to overgeneralize from one deeply unusual incident. That would be stupid.

But it does make the tail risk feel a little less like science fiction.

And I don’t really foresee the underlying tensions getting better if AI progress continues at its current pace while public attitudes keep moving in the opposite direction.

To my very untrained and amateurish political eye, it increasingly feels like there is a real constituency in America for some form of anti-technology politics.

What form that eventually takes is hard to know.

But AI already seems well on its way to becoming an election issue, and politicians will inevitably figure out how to package some combination of anxiety about jobs, data centers, Big Tech and economic insecurity and put it on the ballot.

Maybe this all blows over.

Maybe attitudes change dramatically as people use AI more and find it useful.

Maybe the economic gains become widely distributed enough that today’s fears look silly in retrospect.

Or maybe this is the beginning of something much larger.

I don’t know.

This is still a very preliminary and messy observation.

But it is definitely a trend worth watching.

Quotes

Thu, 20 Aug 2026 00:00:00 GMT

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

― Agatha Christie, The A.B.C. Murders

T.M. Krishna on symbols

Wed, 19 Aug 2026 00:00:00 GMT

Source: thewire.in

AI is still underhyped

Tue, 18 Aug 2026 00:00:00 GMT

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, regular usage is still surprisingly low. Depending on how you define usage and which survey you look at, only a minority of people seem to be using these tools frequently. And in places like India, we don’t even have particularly good surveys to begin with.

More importantly, the people who use these tools the most are still largely what we’d call early adopters.

The adoption among the normies, so to speak, isn’t all that much.

And if you exclude tech workers, power users, and people who spend their entire lives online talking about AI, I’d imagine the numbers drop quite sharply.

I was reminded of this recently when a friend called me to ask about some financial course he was thinking of taking.

I told him that he’d probably be better off using ChatGPT to learn whatever he wanted to learn instead of paying for some overpriced, pointless course which, more often than not, will have some idiotic instructor who has very little clue what he’s talking about.

He found this a little hard to believe.

He asked me: isn’t a course better because it has some structure?

And I said, well, ChatGPT can give you all the structure you want.

But more importantly, the fact that you don’t need structure is precisely what makes these tools so useful for learning.

You can ask whatever you want. You can ask stupid questions. You can ask the same question ten different ways. You can ask it to explain something like you’re five or like an idiot. You can jump from one topic to another, go down some random rabbit hole, and come back.

It felt like he couldn’t quite picture what I was talking about.

And this is precisely why I think AI is underhyped.

I’m not even talking about the fancy stuff.

Forget agents, autonomous companies, AGI, and whatever else people are verbally ejaculating about on Twitter on a given day.

I’m talking about what Zvi Mowshowitz calls mundane utility.

Just boring, everyday usefulness.

Using an LLM instead of Googling something.

Asking basic questions.

Summarizing something.

Writing a response to an annoying email.

Writing some basic ad copy.

Creating an image for social media.

Translating text.

Transcribing audio.

Understanding some concept you don’t know.

Learning how to use Excel.

Taking some incoherent brain dump and turning it into something vaguely readable.

These are all incredibly mundane things.

But that’s the point.

I don’t think most people realize just how much utility these tools can offer ordinary people on an ordinary day.

And yes, the absolute user numbers for things like ChatGPT and Gemini are already enormous.

But those big numbers can be misleading.

There is a massive difference between having opened ChatGPT a few times and having internalized the idea that whenever you’re confused, stuck, curious, trying to learn something, trying to write something, or trying to avoid some irritating piece of grunt work, you can just ask an LLM.

These are still very, very early days.

And if you’re sitting there thinking you’re too late to AI, that everybody else has gotten some enormous head start, that other people are already “AI natives” or “AI experts,” and that you’re somehow behind in your career because you haven’t figured all of this out yet—

all of that is utter bullshit.

You’re not late.

In fact, I suspect it’ll be very hard to be late for quite some time.

Because most people probably won’t use these tools heavily. They’ll continue doing whatever they’ve always been doing in their personal and professional lives because it works well enough.

So regardless of what you think about AI, regardless of whether you’re optimistic or pessimistic about it, and regardless of whatever everybody else is saying online, I’d recommend doing something much simpler.

Use these tools heavily.

Try them for everything.

Personal things. Professional things. Serious things. Stupid things.

See where they work for you and where they don’t.

And then make up your own mind.

You might be surprised by just how useful they are.

A 17GB miracle

Mon, 17 Aug 2026 00:00:00 GMT

Source: simonwillison.net

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.

Rare earths are neither financially valuable nor financially viable

Mon, 17 Aug 2026 00:00:00 GMT

Source: ft.com

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?

Bhuvan: notes

Thank you, Perry

The real change India needs

When facts change I change my mind. What do you do sir?

Twitter is annoying, frustrating, but still awesome

William Hazlitt on Jeremy Bentham

The times, they have changed.

The rage economy

AI videos are crazy

On brain rot

Gandhi vs Ambedkar

Sukumar Sen and India’s First Election

An act of faith

The mathematician in charge

Building the machine

The vote

A forgotten founder

Further reading

A tale of two Chinas

Indian government sucks at industrial policy

Sales of weight loss drugs in India

Begum Aizaz Rasul

How much foreign investment did India attract?

NVIDIA Nemotron 3 Diarization

Galbraith’s Antidote to Economic Dogma

Coding from my phone is absurd

Mancur Olson on collective action

How to value companies when you’re a cokehead

Qwen3.8-Omni-Flash and Qwen3.8-LiveTranslate

We were sloppy we’ll before AI

Notes from messing around with vision, text, and audio models

Vision

Text

Audio

Languages

The part I still find crazy

Links

Vision and general-purpose models

Audio and transcription

Disclaimer

Why Japan matters

Reading list on open source AI

Hidden chokepoints

The price of Trump’s stupidity

What does it mean to think in the shadow of AI?

What GDP Doesn’t See

There is still Alpha to be had in good training data.

DeepSeek V4.1 Flash Vision is surprisingly good

We are the stories we tell

Feelings that defy language

Sweden’s crime problem

India after Gandhi

Visualizing the OpenAI-Hugging Face incident

Research on the wealth effect

The bear case for AGI

I got some links for you

Agentic payments are coming to UPI

AI is NOT guzzling water!

Omawrite

Sonder

The rise of agentic commerce

When everyone take the fastest route

How does the preference for sons impact the mental well-being of women in rural India?

Small language models are all we need

Google’s Pragmatic Pivot?

Seek not more answers but better questions

Chinese models are the most popular AI models by token volume

On AI doom

Gold is a hedge against debasement

Pep Guardiola on players he admires

Welcome to AI slopulism

Quotes

T.M. Krishna on symbols

AI is still underhyped

A 17GB miracle

Rare earths are neither financially valuable nor financially viable