Monday Night Links

• The trillion-dollar cluster—+4 OOMs from the GPT-4 cluster, the ~2030 training cluster on the current trend—will be a truly extraordinary effort. The 100GW of power it’ll require is equivalent to >20% of US electricity production; imagine not just a simple warehouse with GPUs, but hundreds of power plants. Perhaps it will take a national consortium. (Note that I think it’s pretty likely we’ll only need a ~$100B cluster, or less, for AGI. The $1T cluster might be what we’ll train and run superintelligence on, or what we’ll use for AGI if AGI is harder than expected. In any case, in a post-AGI world, having the most compute will probably still really matter.) [Situational Awareness]
• South Bow (SOBO) on May 29 announced open season results for its Prairie Connector project, securing 20-year binding commitments from Hardisty, Alberta to US delivery points. The company also revealed several key project details, giving the market more visibility on a potential new Canadian crude egress path. The project would add 380 km (~236 miles) of new 36-inch pipe and use 150 km (~93 miles) of preserved 36-inch pipe, plus two existing pump stations. Based on a typical oil pipeline flow velocity of 3-10 ft/sec, a 36-inch line implies ~300-1,100 Mb/d of theoretical capacity. SOBO is targeting a final investment decision (FID) by mid-2027. [East Daley Analytics]
• A year and a half ago, our city’s one bookshop went up for sale. My wife and I bought it. The place had 20,000 books, a good music system that probably played 3,000 hours of Bach per year, and a black cat named Raven. It even had an entire room just for theology and philosophy books. (Steubenville, Ohio, is an unusual town.) Now we’ve been running the shop for a little while. Experientia docet, said our ancestors. Experience teaches. I prefer Vivaldi to Bach, so you’ll hear more Vivaldi around here now. Raven died and my children are now tending a litter of six kittens, grooming a replacement. And experience has taught us something heartening: Our customers have great taste in books. I write this because I hope it will be as great a consolation to you as it has been for us. People tell us all the time that civilization is finished, that the world is coming to an end. But then we look at our sales details and we smile. [John Byron Kuhner]
  
• In the early 1970s, the Bretton Woods Agreement—the post-world war pact that instituted a fixed-exchange rate regime for the major world nations— had begun to show its structural flaw. Finance ministers were finding it increasingly difficult to dictate the value of currencies relative the dollar in a world where value changes were constant and information and capital free flowing. On August 15, 1971, President Nixon announced an emergency economic package that sent a seismic shock through the entire financial world. On that day, the United States suspended the dollar's convertibility into gold thereby ending fixed exchange rates between currencies. The Chicago Mercantile Exchange was the first major futures exchange to recognize the market potential of the upheavals unleashed by this event. Supported by Nobel laureate Milton Friedman, the CME was the first exchange to assert that the principles of agricultural commodities futures could be applied successfully to financial instruments. Thus, currency futures—which began trading on the CME's International Monetary Market (IMM) on May 16, 1972—ushered in the era of financial futures, thereby forever changing the scope and utility of futures markets. One year later, the Chicago Board of Trade launched the Chicago Board Options Exchange (CBOE) and added a new dimension to the repertoire of risk management instruments. Three years later, Treasury bond futures at the Chicago Board of Trade made their debut and became the most actively-traded financial instrument. Within a decade, a vibrant new industry was born that subsequently opened the curtain on the index markets of the 1980s. The successes of these markets propelled the futures and options industry to unparalleled greatness. In the last decade alone, the volume in U.S. futures and options skyrocketed from 76 million contracts in 1979 to a record of 323 million contracts in 1989. These successes also prompted University of Chicago Professor Merton H. Miller, 1990 Nobel laureate in Economics, to nominate financial futures as "the most significant financial innovation of the last twenty years." [Leo Melamed]
  
• The Silicon Data Token Expenditure Index, which tracks total spending on large language model usage, has roughly doubled since late 2025 even as the price of a single token has fallen more than 90% since 2023. This is Jevons paradox in action. As tokens get cheaper, companies don't spend less but instead run more AI agents, automate more workflows and generate more code, pushing aggregate expenditure higher even as the unit cost of intelligence collapses. [Apollo]
• From our vantage point at KKR, the current U.S. policy mix looks meaningfully different from the 2000-2019 period, which was defined by low growth, low inflation, global interdependence, and tighter fiscal constraints. In the current environment, we expect our asynchronous global recovery thesis, characterized by rolling recoveries and rolling recessions across regions and sectors, to not only continue but also gain broader acceptance. If we are right, the long-term implications for asset allocation could be significant, including the need to own more Real Assets across portfolios. Indeed, in a world where dispersion is rising and traditional diversification is less dependable, we continue to favor return streams anchored in hard assets and collateral-based cash flows. These exposures do not eliminate volatility, but they can improve the controllability of outcomes through seniority in the capital structure, stronger recovery prospects, and cash flows tied to essential activity in the real economy. [KKR]
• If this were true—that an LLM, lacking consciousness, must instead constantly confabulate its own behavior and motives and reasoning—then, even if an LLM is extremely intelligent, we should expect errors to accumulate differently in LLMs vs. humans. And as philosopher Toby Ord has demonstrated, this is exactly what’s observed. In his analysis of METR’s data on AI task-length doubling times, Ord identified that the available data also fit there being a simple “half-life” for the success of an AI agent, and so have a “constant hazard rate” for long tasks. As Ord writes about what this suggests concerning an AI’s prospects for completing long tasks: "the chance of failing at the next moment is independent of how far you’ve come—just like how the chance of a radioisotope decaying in the next minute is independent on how many minutes it has survived so far." Ord even shows that humans, when their success rate at long tasks is analyzed, appear to deviate from an equivalent constant hazard rate. A strong contender as to why is because we have interpretable access to our own previous thoughts and actions, i.e., our consciousness, in a way that LLMs simply don’t have. [The Intrinsic Perspective] 
• By choosing to treat the 1930s as a form of extreme payback for extreme excess—“no generation exempt”—Sorkin stages morality play rather than history. He also helps set policymakers up for the kind of grand theatrical action they are inclined to take anyhow whenever markets turn down. In other words, another 1933- or 2008-style rescue: flooding the market with liquidity, and stringing up wrongdoers and even the better Wall Streeters, such as the Mitchells whom Sorkin seeks to rehab. The same subpar results are likely to follow. [Amity Shlaes]
• In terms of AI progress per resource, or per dollar, things are probably getting worse on most measures. This is what the pessimism about scaling laws is getting at. Measures of quality are increasing far slower than the exponentially mounting costs. So why were people like Ilya Sutskever and Dario Amodei so impressed by the scaling laws? The answer is that there was a lot of headroom in compute — a lot of room to scale those costs.  Even if the resources needed would be thousands of times larger than the largest ever ML experiment, they saw that this (1) was still within the feasible set of things a large company could fund and (2) would still be a good deal given the vast potential benefits. For on the high end of possibilities, we are talking about something that could potentially replace more than half of all human labour in the world, and perhaps bring forward scientific and technological advances that human labour would have taken centuries to reach. In other words, while the costs could escalate wildly, a deep-pocketed project might reach benefits of extreme value before it ran out of money. And those benefits might be worth more than enough to justify those (very high) costs. More prosaically, we can note that it is entirely possible that financial returns will scale very quickly as a function of the technical measures of AI quality. If so, then even though standard measures of AI quality scale poorly as a function of resources, the financial returns might still scale very well as a function of resources. Indeed, if they scale better than linearly, that would create a paradigm of increasing marginal returns which would explain a landscape with a small number of players, each investing as much as they possibly can. [Toby Ord]
• It will come as a surprise to most that a company so stiff and humorless as CME Group currently is, once used to run an aggressive and somewhat jingoistic ad campaign in the mid 1970s for the purpose of celebrating free markets.  As I recently obtained a second set of these somewhat rare prints from a retired former CME marketing employee, it's worth taking a deeper look into the story behind them. [Trading Pit Blog]
• Mr. Chairman, there are no commodity exchanges in Moscow; there is no Peking Duck Exchange in China; there is no Havana Cigar Exchange. The farmers of those countries have no need for a mechanism that offers risk transference, price projection, or price protection. In those countries, the governments establish the prices at which farmers can sell their products. Consequently, the farmers' primary risk is entirely removed. Alas, by removing the risk, that system also removes the incentive. The sorry history of such systems is that they have been abysmal failures. [Leo Melamed]

How much bitcoin would be too much to own?

Imagine if you had most of your net worth in Bitcoin. How might this have happened? Maybe you are a Bitcoin enthusiast who has been stacking and "hodling" for a long time. Or perhaps you recently remembered a forgotten password that unlocked a bitcoin trove. In either case, the assumption is that this is a large holding as a percentage of your net worth and you have a low cost basis, which would mean substantial capital gains taxes if you sold. What should you do?

Most people in this position resolve the question with a feeling. The bull feels that selling would be a betrayal of the thesis that made him rich. The accidental holder feels paralyzed by the tax bill. Neither feeling is a good answer. There is, however, a framework that produces an actual number, and it has been sitting in the mathematical literature since 1956. Before we get to it, we need to be honest about what Bitcoin is, and about the best case that can be made for it.

Not a First-Class Asset
I have a general rule for windfalls: invest them in first-class assets for the benefit of future generations, and consume only the income (at most). First-class assets are productive. A business generates earnings. Timberland grows wood. Farmland yields crops. An apartment building collects rent. You can hold these assets forever, live on their output, and pass the principal on intact. The income is a signal that the asset is real and not just a speculation.

Bitcoin produces no income. It pays no dividend, collects no rent, earns no royalty. Whatever you eventually get out of it must come entirely from a future buyer. This does not make Bitcoin worthless but it disqualifies it from the category of first-class assets you build a family fortune on, and it puts it squarely in the category of speculation. Speculation is a legitimate activity, but it obeys different rules, and the most important rule of speculation is position sizing.

The Best Case for Bitcoin
Let me give the bulls their due, because the strongest argument for Bitcoin is genuinely strong, and it was prescient of Murray Stahl to argue this at least as far back as 2019. It goes like this: add up the market value of garbage currencies like the Brazilian real, the Iranian rial, and the Russian ruble (just a few bad currencies that happen to start with "r") and you arrive at something well over $2 trillion of combined value. Why couldn't a new, better currency (with an algorithmically fixed supply and the ability to transfer virtually) rival any one of those currencies, or even all three together?

That argument did not require Bitcoin to dethrone the dollar. It did not require hyperbitcoinization, the orange-pilling of central banks, or the citadel. It required only that Bitcoin function as a more trustworthy store of value than the monetary product of a few chronically mismanaged states. For a citizen of Buenos Aires, Tehran, or Moscow, that proposition is no longer all that speculative.

So the bull case has a floor of plausibility. The question is its ceiling.

Bitcoin Already Worked
The bulls were right - past tense. Over fifteen years, Bitcoin grew from a hobbyist experiment worth essentially nothing to a market capitalization of about $1.2 trillion, approaching the entire value of euro banknotes in physical circulation. That is an extraordinary achievement in monetary entrepreneurship. And the contrarians who bought at $10 billion, or $100 billion, or $500 billion have been handsomely paid for being right.

If you study technological adoption patterns (see Vaclav Smil, or Crossing the Chasm), the penetration rate for a new innovation follows a logistic growth, or "S," curve. The early phase is explosive: that was the contrarians' decade, when Bitcoin's share of global physical currency went from essentially zero in 2011 to roughly 4 percent by 2017, a four-thousand-fold increase in penetration. Then growth decelerates as the addressable market saturates. From 2017 to today, Bitcoin's share of world M0 has grown only about four-fold, despite the same number of years. Plotted on a multi-year moving average (to smooth through the cycles), the trajectory has the unmistakable shape of a logistic curve: near zero through about 2014, a steep climb through 2021, and a clear deceleration since.

At roughly 15 percent of world physical currency today, Bitcoin is no longer in its hypergrowth phase. It is in the part of the curve where the rate of change falls and the asymptote becomes visible.

"Bitcoin worked" and "owning Bitcoin will keep working" are therefore different propositions. The contrarian trade that took Bitcoin from a curiosity to a major monetary unit has already been made; it is no longer on offer. The question now is whether the next move, from this side of the S-curve where saturation dynamics dominate, represents the same kind of opportunity. It doesn't. 

Being contrarian is not, in itself, a strategy. Being contrarian about the right thing at the right time is. Having been right about Bitcoin in 2013 does not generate forward edge in 2025, any more than having been right about Apple in 2003 tells you what to do with AAPL today. The Kelly framework, which we'll come to in a moment, is forward-looking by construction. It does not ask whether Bitcoin was a good bet a decade ago. It asks whether the bet currently available, at current size and against current failure probabilities, is one worth taking, and at what fraction. 

The Ceiling Is Smaller Than You Think
The S-curve raises a structural question the bull case usually papers over: an asymptote against what? Enthusiasts routinely make a category error here, and it is worth dwelling on because the error inflates the apparent upside by an order of magnitude.

The popular version of the moonshot thesis compares Bitcoin to the global money supply, which might be roughly $123 trillion by the M2 measure or even $150 trillion by broader gauges. Against numbers like those, Bitcoin's current market capitalization of roughly $1.2 trillion looks small, and the implied possible upside looks like a hundred-bagger.

But M2 is the wrong yardstick. The bulk of M2 consists of savings deposits, time deposits, and money market funds, all of which are instruments that pay their holders interest. People park money in a savings account because the bank compensates them for doing so. Bitcoin pays nothing. It cannot compete for interest-seeking money, and no mechanism exists by which it ever would. Comparing Bitcoin to M2 is comparing it to a market it cannot address.

The honest comparison is physical currency, the narrowest measure of money, M0: the banknotes and coins of the world. Cash pays no interest either. People hold it anyway, for liquidity, portability, privacy, and crisis insurance. These are precisely the virtues Bitcoin claims for itself. Cash is the incumbent Bitcoin is actually running against.

All the physical currency in the world amounts to something like $8 trillion. 

That is the ceiling. And against an $8 trillion ceiling, Bitcoin at $1.2 trillion has already captured roughly 15 percent of its total addressable market. This is a genuinely astonishing achievement for fifteen-odd years of work. More astonishing than the bulls themselves seem to realize when they reach for the M2 comparison. But it carries an uncomfortable implication: the remaining upside, in the best plausible case in which Bitcoin equals every banknote and coin on Earth, is about 6.7x. 

The Distribution of Outcomes
A 6.7x ceiling would be perfectly attractive if it were the likely or expected outcome. It is neither. It is the far tail of a distribution that includes zero. Curtis Yarvin framed the long run for Bitcoin very succinctly: "One day, everything in the world economy will be priced in Bitcoin. Or, one day, Bitcoin will be regarded as the biggest bubble that ever was. There are stable equilibria between these outcomes, but they are unusual and structurally exotic." 

He put forward three failure modes: Bitcoin could be actively killed by its enemies; it could lose its energy source; or it could be outcompeted by another candidate monetary standard.

None of these is far-fetched. States have banned more popular things than Bitcoin. Proof-of-work rests on an energy-economic equilibrium that regulation or technology could upend. And monetary standards are winner-take-most competitions in which the eventual winner (a central bank digital currency, a successor protocol, something unimagined) need not exist yet.

There is also now experimental evidence bearing on the upside scenario, and it is not encouraging. In 2021, El Salvador became the first country to make Bitcoin required tender, deploying both subsidies (a $30 signing bonus, free transactions, zero capital gains tax) and coercion (businesses were required by law to accept it). After four years, the experiment was dead: crypto remittances had shriveled from 4.5 percent of the total to under 1 percent, and by 2024 just one in two hundred Salvadorans was paying with bitcoin as often as every other week. 

JP Koning, the monetary economist who chronicled the experiment at his blog Moneyness, identified the structural reason it failed. There is a very big hurdle that prevented El Salvador's one-two punch of subsidies and coercion from working: Bitcoin is intrinsically ill-suited to perform as money. The stuff is innately volatile, so risk-shy individuals don't dare hold it or use it for payments. Risk-seekers can tolerate the volatility, but they expect to be rewarded by a dramatic price rise, so they refuse to spend their bitcoins, because they could miss out on the jump. The net result is that no one, neither society's risk-seekers nor its risk-avoiders, ends up paying with bitcoins.

This matters enormously for our probability estimates. The full moonshot (Bitcoin displacing the world's physical currency) requires Bitcoin to actually function as circulating money. A national government just spent four years trying to force exactly that outcome, under maximally favorable conditions, and failed. A rational holder should weight the M0-parity scenario accordingly: it remains possible, but the best natural experiment we have gotten so far argues against it.

So the shape of the bet is: a meaningful chance of losing everything, a meaningful chance of stagnation, and a tail chance of making roughly five to seven times your money. How much of your net worth do you put on a bet shaped like that?

Kelly Tells You the Number
In 1956, John Kelly of Bell Labs derived the formula for exactly this problem: given a bet with known odds and probabilities, what fraction of your bankroll maximizes the long-run growth of your wealth? Bet less than the Kelly fraction and you leave growth on the table. Bet more and volatility itself destroys you. Overbetting a favorable gamble, repeated over time, grinds even a winning hand down to nothing. The Kelly fraction is where aggression and survival balance.

For an asset with several possible outcomes, the procedure is to choose the fraction *f* of your portfolio that maximizes the expected logarithm of wealth:

E[ln(W)] = Σ pᵢ · ln(1 + rᵢ · f)

where pᵢ and rᵢ are the probability and return of each scenario. Solve for the *f* where the derivative equals zero, and you have your allocation. The inputs are estimates, but the discipline of writing them down is most of the value of the exercise. Let us write down three honest sets.

The skeptic thinks Bitcoin is a doomed experiment: a 65% chance it goes to zero, a 28% chance it merely treads water, a 7% chance it doubles by displacing one weak currency. Expected value is sharply negative, and the Kelly fraction comes out negative as well. The formula is telling you the bet is unfavorable: the correct allocation is zero. If these are your probabilities, you should own no Bitcoin at all.

The pragmatic bull believes Bitcoin survives and grows but stops short of conquest: a 45% chance of failure, a 35% chance of stagnation, a 15% chance it displaces all three weak "r" currencies (about $6 trillion, a 5x from here), and a 5% chance it reaches full M0 parity at $8 trillion (a 6.7x). Solving the optimization yields a Kelly fraction of roughly 15% of net worth, and at half-Kelly, the standard discount practitioners apply because real-world probability estimates are soft, about 7%.

The true believer is maximally, but still rationally, bullish: only a 30% chance of failure, a 25% chance of stagnation, a 25% chance of the $6 trillion outcome, and a full 20% probability that Bitcoin swallows the world's physical currency entirely (a probability the El Salvador evidence suggests is generous). Even granting these aggressive inputs, the Kelly fraction is roughly 50%, and half-Kelly says about 25%. Note what those numbers require: a one-in-five chance Bitcoin wins outright against every banknote and coin on Earth, and only a three-in-ten chance of total failure — assumptions that strain against the experimental record.

The defensible range is therefore wide but shaped. The skeptic case gives zero. The pragmatic bull, applying the standard half-Kelly buffer, gives about 7 percent. Only the true believer, with assumptions that don't really survive scrutiny, gets to roughly a quarter of net worth at half-Kelly. To justify the very large or "all-in" allocations of the biggest enthusiasts, you would need probability estimates so lopsided (near-certainty of success, near-zero chance of failure) that writing them down would embarrass you. That is the point of writing them down.

A Second Opinion: The Market Portfolio
Suppose you distrust the whole exercise of assigning probabilities. Fair enough, there is another framework, that requires no views at all.

If you believe markets are efficient, the portfolio you should hold is the global market portfolio: all the asset holdings in the world, weighted by market capitalization. This is the allocation of the investor with no edge and no opinions. What does it contain? World stock market capitalization is roughly $150 trillion. The world fixed income (bond) market is about equal. Global real estate is perhaps $400 trillion. Cryptocurrency, in its entirety (every coin and token on CoinMarketCap) is a couple of trillion, of which Bitcoin's share (its "dominance," in the argot) is about 58 percent.

Run the division and Bitcoin's weight in the total market portfolio comes to roughly 20 basis points. Granted, world real estate in its entirety is not an investable asset class, so one might reasonably narrow the universe. But even against world stocks and bonds together, Bitcoin's weight is well under half a percent. (And against stocks alone, still under 1 percent.)

Two entirely independent frameworks: one (Kelly) that assumes you do have views and rewards you for acting on them; one (the market portfolio) that assumes you have none and tells you to hold assets in proportion to their market valuations. The first, fed honest pragmatic-bull inputs and applying the standard half-Kelly buffer, says about 7 percent. The second says 0.2 to 1 percent. The defensible range for a Bitcoin allocation, then, runs from a fifth of a percentage point to perhaps 25 percent, with where you sit in that range determined by how confident you are that you see something the market doesn't and that you can assign probabilities the El Salvador evidence does not refute. What neither framework produces, under any inputs whatsoever, is a number anywhere near "most of my net worth."

"But the Taxes"
Now return to our holder with the low cost basis. Selling means realizing an enormous gain and handing a quarter or a third of it to the government. But the tax argument, examined closely, is an argument about which loss you prefer. Bitcoin has declined 80 percent or more from its peak multiple times in its short history. If it does so again, or if one of Yarvin's failure modes materializes, the holder who refused to pay a 25% tax will have donated far more than that to the market. Taxes are the price of having won. Drawdowns are the price of refusing to admit it.

One way to translate this thinking into action would be a target allocation consistent with one's probability estimates. Sell down to it over a multi-year horizon, into strength where possible, spreading the tax burden across tax years, and redeploy the proceeds into first-class assets.

The Answer
How much Bitcoin would be too much to own? Under the framework above, allocations beyond roughly a quarter of net worth require probability assumptions that the El Salvador evidence directly undermines. A defensible bull case, sized prudently at half-Kelly, puts the ceiling closer to 7 to 10 percent. The efficient-markets baseline says the neutral weight is closer to a fifth of a percent. The asset produces no income, faces live existential risks, has demonstrably failed its one national-scale audition as actual money, and has already captured something like 15 percent of its total addressable market. The glory days of hundredfold upside are history.

The Kelly Criterion gives you a framework. The framework says: even if you love Bitcoin, keep it to a fraction of your portfolio. Exactly which fraction depends on your honest assessment of the probabilities. But "most of my net worth" is not a defensible answer under a reasonable set of assumptions.

[This essay is a framework discussion, not personalized investment advice. Probability estimates are illustrative; readers should consult their own advisors regarding their specific circumstances and tax situations.] 

Thursday Morning Links

• It’s official: the world stock market is now as wild as it was during the tech-stock bubble. Since April 2026, we’ve had epic levels of return dispersion. AI excitement is driving extreme price moves, with some large-cap technology stocks up 50% to 100% in a single month. While it’s normal for small-cap stocks to sometimes rise 50% in a month, it’s not normal for the whole market to be meaningfully impacted by extreme winners. The chamber of dispersion has been opened, the beast of volatility has awakened, and the season of chaos is at hand. [Acadian] 
• The fact that short-lived strains are desirable for research because they allow it to go faster has polluted the literature because short-lived, convenient-to-research strains produce unreliable results for the very reasons behind their being short-lived. Pabis et al. therefore argued that, in order to demonstrate longevity benefits credibly, strains with long control lifespans in optimal conditions are needed. [Cremieux Recueil] 
• A short while later, the book ends suddenly and ignobly. It turns out that keeping the army alive one more season was all that was required. The real Spartan army is launching an operation against a Persian satrap in Asia Minor, and Xenophon’s troops get summoned to participate, initially as their own unit and then gradually getting absorbed into the main body of soldiers. In other words, it ends with an acqui-hire, as so many startups do. The dreams of great wealth and fame are over, and their independence is over too. It’s back to being a wagie at AWS. [Mr. and Mrs. Psmith’s Bookshelf]
• Rayonier was founded in 1926 as the Rainier Pulp and Paper Company in Shelton, Washington, with a headquarters office in San Francisco, California. Its name was inspired by Mount Rainier in Washington state. Its first mill opened the next year in Shelton Washington, on the Olympic Peninsula. The mill used Tsuga heterophylla (western hemlock) trees to create a premium bleached paper pulp. In September 1930, Rainier Pulp and Paper began working with the DuPont chemical company to produce hemlock pulp for the manufacture of rayon. Two additional pulp mills were constructed and began operation in the state of Washington at Hoquiam, Washington and Port Angeles, Washington. Rainier Pulp and Paper changed its name to Rayonier, a portmanteau of the words rayon and Rainier, in 1937, when it became a publicly traded company. [Rayonier]
• In response to the demand by residents for the board to stop the wind farm from being built, Supervisor Jason Whiting asked County Attorney Brad Carlyon if the board is legally authorized to prohibit the project. “Bottom line answer: No.” Carlyon further explained that the board’s authorities lie in land use and zoning. “You can’t deny it unless there is a health, safety, or welfare issue.” Chair Darryl Seymour also spoke on the work that was done to regulate incoming development. “The work that we did for two years to make it not easier, but to make it where we had more authority, which is already limited by the state. To make it more conducive and more neighborly to our people that we have.” The board then entered executive session to discuss the development agreement, after which the agreement was approved. [Painted Desert Tribune]
• The Rowlings, who own Omni Hotels & Resorts, in March paid $289 million for Justice family debt partly secured by the Greenbrier from a Virginia-based bank. They say the Justice family has siphoned revenue from the resort to pay bills at their coal mines and other businesses, letting the hotel fall into disrepair, skipping payments to contractors and employees, and risking its value as collateral. Jim Justice is credited with saving the Greenbrier when he purchased it out of bankruptcy in 2009. His daughter, Jill Justice, is president of the resort. Jim Justice said in an interview with The Wall Street Journal Friday that the resort has been profitable, retained its clientele and that he plans to continue investing in it. He called the allegations hurtful. “I’ve poured about everything I’ve got into the Greenbrier, and I’d do it again tomorrow,” he said. “I love that place beyond all good sense. It’s not just bricks and mortar to me.” [WSJ]
• This year, Google became the only large tech company to own a power company with the $4.75 billion acquisition of Intersect, a wind and solar developer that in recent years pivoted to building such projects to support data centers. Intersect has energy projects under development to supply multiple gigawatts of electricity, Google said. One gigawatt can power hundreds of thousands of homes. The company has also amassed a bench of energy-sector experts on its staff. “Google really has hired in-house to develop their data centers, and as a result, they have a more integrated approach that I would suggest is a lot more thoughtful,” said Jigar Shah, an energy entrepreneur who formerly directed the Energy Department’s Loan Programs Office. Having on-site power is becoming a strategic advantage for tech companies. Regulators and power officials in many regions are considering whether data centers built alongside new power sources should be afforded a faster connection to the grid because they would be less reliant on it. [WSJ]
• In Dickens's novels anything in the nature of work happens off-stage. The only one of his heroes who has a plausible profession is David Copperfield, who is first a shorthand writer and then a novelist, like Dickens himself. With most of the others, the way they earn their living is very much in the background. Pip, for instance, ‘goes into business’ in Egypt; we are not told what business, and Pip's working life occupies about half a page of the  book. Clennam has been in some unspecified business in China, and later goes into another barely specified business with Doyce; Martin Chuzzlewit is an architect, but does not seem to get much time for practising. In no case do their adventures spring directly out of their work. Here the contrast between Dickens and, say, Trollope is startling. And one reason for this is undoubtedly that Dickens knows very little about the professions his characters are supposed to follow. What exactly went on in Gradgrind's factories? How did Podsnap make his money? How did Merdle work his swindles? One knows that Dickens could never follow up the details of Parliamentary elections and Stock Exchange rackets as Trollope could. As soon as he has to deal with trade, finance, industry or politics he takes refuge in vagueness, or in satire. This is the case even with legal processes, about which actually he must have known a good deal. Compare any lawsuit in Dickens with the lawsuit in Orley Farm, for instance. [George Orwell]
• First smoking went out of fashion, then daytime drinking. Negroni-drinkers were dying out, or reforming. Wine began to be drowned out by absurd fizzy water. Then the food itself was threatened, partly because business lunches were becoming more businesslike, partly because interviewees were becoming more self-conscious. “Even I have never managed a two-bottle lunch,” laments the FT’s arts writer Peter Aspden, who is thought to have conducted more Lunches than anyone else. “There’s such reluctance now to be seen to be pigging out in any shape or form, and hardly anyone drinks over lunch.” [Financial Times]
• The word embedding approach is able to capture multiple different degrees of similarity between words. Mikolov et al. (2013) found that semantic and syntactic patterns can be reproduced using vector arithmetic. Patterns such as "Man is to Woman as Brother is to Sister" can be generated through algebraic operations on the vector representations of these words such that the vector representation of "Brother" - "Man" + "Woman" produces a result which is closest to the vector representation of "Sister" in the model. Such relationships can be generated for a range of semantic relations (such as Country–Capital) as well as syntactic relations (e.g. present tense–past tense). [word2vec] 
• The distributional hypothesis in linguistics is derived from the semantic theory of language usage, i.e. words that are used and occur in the same contexts tend to purport similar meanings. The underlying idea that "a word is characterized by the company it keeps" was popularized by Firth in the 1950s. The distributional hypothesis is the basis for statistical semantics. Although the distributional hypothesis originated in linguistics, it is now receiving attention in cognitive science especially regarding the context of word use. [Distributional semantics]

Thursday Night Links

• It is not good news, then, that over the past 10 years, private equity has focused on taking companies private that are both more levered and less profitable than similarly-sized businesses (which we have already established are junky to begin with), all while paying a higher price for them than their size-peers. These companies bank on the same set of factors—that interest rates will remain sufficiently low and the economy sufficiently robust—for them to be able to generate enough cash flow to pay down their debts. Different types of shocks—an economic slowdown, a further pickup in real interest rates if inflation proves sticky, a widening of credit spreads if private credit continues to sour—can all be enough to meaningfully hamper the profitability of these businesses, and to do so in a correlated manner. [GMO]
• The leading indicator of all previous technology transitions has, however, been the growth of the new, which always precedes the destruction of the old.  The transition to mobile telephony wasn’t declared a failure because the global number of landlines continued to grow until 2006, 33 years after the invention of the mobile phone. Similarly, peak US horse population did not occur until 1920, over 34 years after the first internal combustion vehicle hit the roads. [Michael Liebreich]
• Historically, the world seems highly stochastic and slightly cyclic, with vaguely regular cadences of empires, companies, and other organizations rising and falling with their own characteristic half-lives. The fundamental problem here is that at the ‘object level’ almost all the returns to success are positive. This means that naively success should generally beget more success in an exponential fashion. It requires an extremely powerful force to offset this default tendency. Regression to the mean is certainly not powerful enough for this. However, such a force must exist because, in practice, the world looks dramatically different from the simple exponential model. There is a constant random-seeming churn of states, empires, companies, wealth etc vs a world where the rich get richer indefinitely and thus everything ends up controlled by the winners at time t=0 forever after. We do not all live in the grand one-world Sumerian empire which simply grew faster and more successfully than its challengers during the hinge of history in 3000BC. Similarly, the East India Company no longer controls global trade; the Ford motor company does not have an eternal monopoly on cars; IBM does not reign supreme in computers. Similarly with wealth, despite its undeniable exponential properties, compound interest seems to be pretty weak in practice. The descendants of Croesus and Crassus do not lord over us all today. This means that there must not only be positive returns to scale and success but also negative returns. Larger states become harder to govern and more prone to civil strife. Larger companies become slow, unwieldy, and lose that indefinable ‘taste’ that made them stand out in the first place. There also appears to be some kind of phenomenon of increasing ‘entropy’ and decay in organizations or states that have been highly successful for a long time. Exponential systems are always poised on the edge of a knife destined to either explode or collapse, however the world largely seems to stabilize these systems or at least prevent explosions (there are a fair few collapses). Understanding and explaining these dynamics has to be one of the fundamental questions of sociology. [Beren Millidge]
• I think it has become clear to me that there is a universality class of ‘architectures sufficient to scale to AGI’, of which transformers are the first discovered element, and that there are a vast array of other possibilities within this same class, many existing on significantly better pareto curves than the classic transformer. I definitely think an order of magnitude of flop efficiency in both training and inference is easily achievable through superior architectures, and likely this will actually be several OOMs of gains over the next five years or so. The combination of this process plus vast software and hardware efficiency improvements will inevitably drive the cost of intelligence and AGI down to very low levels which, if alignment can be managed well, is good news since it vastly increases the utilons per capita achievable with the cosmic endowment. [Beren Millidge]
• Even in this idealized scenario, inheritance tax and the division of the fortune among children mean that almost all fortunes are barely able to compound long enough to be maintained even with relatively low TFRs (2 children). Beyond this idealized scenario, the real world also contains many negative ‘shocks’ which can tip a fortune from a regime of steadiness and growth towards exponential diminishment and exponentials are very unkind on the downward slope. The real stock market, even when highly diversified, does not return 7% annually smoothly forever. It goes through huge bust periods where valuations can drop by 50% or more and can remain below all-time-highs for decades at a time. These times can often be coupled with periods of high inflation which are also devastating for large existing fortunes. In theory, many down-markets can be weathered if you just reduce spending to maintain the same withdrawal fraction. However, in practice achieving sudden 50% or more budget cuts can be challenging, even for the wealthy. Many ‘big ticket’ items such as mansions, yachts, etc have high fixed maintenance costs which cannot easily be reduced in a downturn without selling the underlying asset at precisely the wrong time. [Beren Millidge]
• Goodspeed shows that British and American expansions do not resemble Dorian Gray, looking beautiful but hiding an inevitable accumulation of malinvestments (objectively bad investments that are destined to fail) and distorted decisions (mistaken economic decisions taken on the basis of bad regulation or flawed prices) that make a correction inevitable. If they did so, he argues, one would expect that as expansions get longer they get more and more likely to end. In his data, however, the relationship between the age of an expansion and the probability of death is essentially zero. Nor do measures of increased investment during the boom correlate with the severity of a downturn. Nor do longer expansions have longer recessions after them. This is why recessions remain essentially unpredictable. Any perceived regularity is likely to be a statistical illusion. Goodspeed shows that attempts to forecast recessions such as inversions of the yield curve (where long-dated government bonds have lower interest rates than short-dated ones) or the Sahm rule (which says a recession is likely underway if the unemployment rate spikes high above its recent lows for three months) are overfitted to US data and don’t work for the UK. [Works in Progress]
•  In 2024, medical spending as a share of GDP was 15% below forecasts made in 2010 and only marginally higher than in 2010. Relative to expectations, the savings were nearly $1 trillion in 2024. In light of this prolonged period of slower growth, we ask the question: has the United States bent the health care cost curve? We start with a model of medical spending that incorporates both the development of new technologies and the equilibrium use of those technologies. Technology development and incentives around its use may add to or subtract from spending growth. We then examine the drivers of the spending slowdown empirically. We attribute slower medical spending growth to five factors: the development of technologies that simultaneously improve health and lower cost; long-run supply elasticities that exceed short-run elasticities; improvements in population health; reimbursement changes that reduce demand and make demand more price elastic; and reductions in the rate of price increases, likely driven by some of these same demand factors. Because these cost slowing mechanisms are expanding over time, we conclude that the United States has bent the health care cost curve, though not as much as it could be or will need to be bent. [NBER] 

Memorial Day Links

• Realistically, frontier LLMs are already strongly superhuman at basically all tasks requiring crystallized knowledge and pattern-matching reasoning, and they are increasingly getting better at fluid intelligence via long CoT reasoning. We are only a small number of breakthroughs away from AGI. When these breakthroughs happen is hard to predict (it could be tomorrow!, or a few months ago in secret, or in 5 years), but once they happen AGI will be here much sooner than we will expect. All the infrastructure for both training and inference is primed and ready. The data-centre buildout will likely reach its climax in the next 2-5 years, but my opinion is that the scale we have right now is sufficient for AGI and probably has been for the last few years. There is thus a massive compute and infrastructure overhang being created that will allow AGI to scale incredibly rapidly once created, at least for a while. Industry trends though are relatively closely (although not entirely) vibe-following and hence we may see a cooling of interest and excitement about AI at ironically precisely the wrong time. Broadly, this is because society is slowly assimilating LLMs into its worldview and understanding. This is going to cause some waves and change a lot of jobs but clearly not going to be utterly revolutionary in the way AGI would be. It will be e.g. like electrification or some large 19th century industrial revolution technology but not the creation of a god. People who are used to LLMs then forget the original promise and dream of AGI. But the AGI is still lurking there in the future, not so distant, and coming closer. [Beren Millidge]
• Scientists don’t yet know why the drugs might be having this effect. One theory is that GLP-1s reduce cancer risk indirectly—by driving weight loss and improving metabolic health, both of which are independently linked to lower cancer rates. The other is more direct: receptors for GLP-1, the hormone these drugs mimic, appear on the surface of some tumor cells, raising the possibility that the drugs are acting on cancer biology itself. [WSJ]
• My current, very rough, timelines are something like 5-15 years for AGI and then 10-20 years after that for widely rolled out and accessible indefinite life extension (ILE) for currently existing humans 4, accompanied of course with a large amount of other classic transhumanist technology. In the conservative case this is a 35 year lag time from today to ILE. Assuming an average life-expectancy of 80 then somebody age 45 today has a pretty good chance of reaching ILE and somebody age 35 has a very good chance. Age 45 (birth date 1980) pretty much demarcates the millennial generation from gen X. [Beren Millidge]
• Perhaps a more intuitive way to think about this is in terms of ‘zooming in’ to a fractal. As we progress down the power-law slope we zoom into increasingly small and detailed features. We can think of a neural network as similar to a telescope observing a region of space. Having a larger model is like having a larger telescope, it allows us to ‘zoom in’ to resolve finer detail. Similarly, training for longer is similar to having a longer ‘exposure time’ which also lets us resolve finer details for a fixed lens size. Due to the natural structure of the dataset, these linear effects get transmuted to power-law decreases in the loss. [Beren Millidge]
• We see that the three income categories had very different experiences during the March 2026 energy price shock. Low-income households increased their nominal gas spending by the least (12 percent). However, this was accomplished because they cut their real gas consumption the most (7 percent). On the other hand, high-income households increased their nominal gas spending by the most (19 percent) in a large part because they reduced their real gas consumption the least (1 percent). Middle-income households had intermediate increases in nominal spending and decreases in real consumption at gas stations. Thus, the K-shaped consumption pattern in both nominal and real gasoline spending was strongly evident in March 2026. These divergences in the response to an energy price shock are not unique to the month. Four years ago, energy prices rose and remained elevated during the spring and summer of 2022 when the Russia-Ukraine war disrupted energy markets. The magnitude of the initial Russia-Ukraine gasoline price shock was broadly similar to the current one, but it lasted longer to date and ramped up over time. As we see from the panel chart below, between January and July 2022, nominal gas spending rose more for high- and middle-income households than it did for low-income households, while real gas consumption declined less for high-income households than it did for middle-income and low-income households. Notably though, while directionally similar, the magnitudes of the gaps (both for nominal and real spending) were noticeably smaller than the corresponding gaps we see in March 2026. [Liberty Street Economics]
• Although this all seems hyper speculative, these kind of dynamics can be important for modelling the far future. Specifically, it implies that once a K3 civilization is established in a galaxy it is extremely hard to dislodge. This means it is likely that the universe will end up mostly cleanly partitioned between full galaxies of colonizers with perhaps only short wars early on in a galaxy’s lifetime to determine which colonizer faction will be victorious in galaxies where colonizers both reached it at roughly the same time. Beyond that, the fact that intergalactic warfare seems likely to be expensive, interminable, and bring little prospect of gain to the aggressor, makes it likely that such wars are rare in general and only pursued for ‘irrational’ purposes. For instance, once the colonization phase is complete, a pure paperclipper would be unlikely to invade neighbouring non-paperclipper galaxies since it would have to expend far more energy to do so than it would gain, and hence would end up with less paperclips overall than if it just paperclip-ified its own galaxy and then sat there for the rest of time. This also implies it is very important for early civilizations to colonize to maximize their share of the cosmic endowment because it is hard to fight back and claim inhabited territory if you are late to colonizing. Another thing of importance to note is just how extremely vulnerable large fixed objects like planets or stars are in this universe. Anything on a predictable orbit can be sniped with impunity and complete surprise from millions of light years away. This means that humanity is super vulnerable in its early stages right now and also even when spread across multiple planets and building its first Dyson sphere. Any K3 or less civilization in the vicinity (if they see us) can directly destroy us at this point, which makes moving some sensible fraction of civilization away from obvious large orbiting bodies and into freely moving habitats far away from any obvious orbits an immediately important backup strategy. Unsurprisingly, it also means that we should invest heavily into trying to determine if there are any K2 or K3s out there in our intermediate neighbourhood (e.g. perhaps in the local group) since if so not only will they likely have colonized almost all galaxies we could reach first but also that as soon as they detect meaningful civilization around earth, then they could disintegrate it with extremely powerful beams with ease. In a variant of the anthropic argument, the fact that this has not happened yet provides some evidence against hostile local K3s existing. [Beren Millidge]

Crises & Bailouts

We started the year bullish. It seemed to us that it would be difficult to have a recession with wholesale gasoline below $2 per gallon and interest rate cuts along the way. We did note one caveat at the end:

Stagflation is the big concern that we see. If the economy is weak and inflation is high, things can get very ugly. The Fed can't come to the rescue by printing. Perhaps that is what Warren Buffett is thinking about, the 1970s stagflation parallels. But it is hard to have stagflation with a plummeting oil price! 

We now have the opposite of a plummeting oil price, and interest rate cuts are no longer in the cards. In fact, we are on the verge of stagflation.

Trump is begging Iran for a deal to reopen the Strait of Hormuz and Iran is refusing. They have him cornered. By closing the strait they are forcing the world to burn through several million barrels per day of reserves. Trump threatens but doesn't dare attack Iran any further because of the damage Iran can do by retaliating against the Gulf states' energy infrastructure. 

We have not had a credit cycle or financial crisis since 2008. It seems to be the case that you can prevent something like 2008 if you are willing and able to do bailouts.

In 2008 the government was caught off guard by the real estate speculation and how that interacted with opaque mortgage securitizations. There was also a false sense of security because real estate prices had never declined on a nationwide basis. (We had a variant view because we had a front row seat to the excesses of the sunbelt real estate bubble.) So, the government was "unwilling" to bailout, but mainly because they did not realize how bad it was at first. Ultimately they nationalized the problem with the Fed balance sheet.

Now the government is certainly willing; look at Powell's bailout of the economy in 2020. Trump & Co. would happily run the printing press even to prevent a 5% market decline. But you can reach a point where the bailout that is needed cannot be delivered because of its size in relation to GDP. The U.S. debt/GDP has doubled since the GFC (60% -> 120%).

One of the things that stays with us from Trillion Dollar Triage was the head of the NY Fed's market desk saying that what was "most frightening" during Covid was that "Treasury yields were spiking higher at the same time that equity markets were plummeting."

The warning sign for another financial crisis would be if bad things happening (whether geopolitical or equity market declines) made bonds go down.

We are seeing some of that with the Iran war. The ten year bond yield (TNX) had fallen below 4% prior to the Iran war, and is now 4.62%. On Friday (May 15th), the S&P was down 1.2% and the ten year yield was up 8.5 bps. (Bond price down almost 1%.)

Our guess of Buffett's thinking is that he has been predicting stagflation. Inflation-beneficiary "croupier" businesses at 30x earnings will not work if higher interest rates cause their valuation multiples to get cut in half. 

Monday Morning Links

• A cyclical business that produces average returns over a full cycle will give investors worse total returns, because at the peak it won't return quite enough capital, and because the equity slice of the capital structure can't always survive the worst part of the cycle. The equity in some high-cost miner might 100x during a good cycle, but that 100x return may be enjoyed mostly by former bondholders who got most or all of the equity when it reorganized: shareholders in Warrior Met Coal have seen their holdings compound at 33% since it went public in 2017. But "Warrior Met" was a new holding company formed by the first lien debtholders of Walter Energy, whose shareholders were zeroed in 2016. Another issue, related to the cycle, is Engineering Ego: people who go into the business of solving complex technical challenges around getting ore out of the ground are going to be excited if there's a uniquely extreme way to do it. [The Diff] 
• Recent leaked US intelligence assessments, as reported by Western media, estimate that Iran has regained access to 90 percent of its underground missile storage and launch facilities, many of which remain at least partially operational, that were buried due to US-Israeli airstrikes. Iranian forces have likely also reestablished communications between units and commanders and begun restoring force morale—both of which were disrupted during the war due to US-Israeli airstrikes and had significantly degraded Iran’s ability to conduct operations to achieve its objectives. Recent US intelligence assessments also estimate that Iran still has about 70 percent of its mobile launchers and 70 percent of its prewar missile stockpile, including both ballistic and cruise missiles. US forces buried some of these assets during combat operations, which rendered the buried assets combat ineffective. Any military force, including the Iranians, would use the time and space granted by the ceasefire to reconstitute itself. The restoration of underground missile storage and launch facilities means that Iran was degraded operationally and then prepared itself for a new round of fighting. US forces have surely also prepared for a new round of fighting. [Institute for the Study of War]
• Sodium ion batteriess could ease supply constraints and price volatility linked to lithium-based batteries by expanding the range of viable chemistries and reducing dependence on lithium mining and processing. Sodium is far more abundant than lithium – around 1,000 times more abundant in the Earth’s crust and roughly 60,000 times more abundant in the oceans. SIB manufacturing primarily relies on soda ash (sodium carbonate) as the main sodium precursor. Soda ash is widely used across several industries, and is far more abundant than lithium, making it less susceptible to resource availability concerns and price volatility. Natural soda ash resources globally are estimated at 47 billion tonnes, while reserves are estimated at 25 billion tonnes. Soda ash can also be, and already is, produced synthetically from salt and limestone, both virtually unlimited resources. Cost could still be one of the competitive advantages of SIBs over LIBs due to a number of factors. The first of these is the abundance and accessibility of sodium, a material that is considerably cheaper than lithium. [International Renewable Energy Agency]
• Well into his seventies, Wilson would march into the Princeton Club and order a half dozen martinis, to be prepared not sequentially but simultaneously—six shining glasses in a bright row, down which Wilson would work, all the while talking and thinking at a rapid pace. To the end of a long life, he kept on making the only thing he thought worth making: sense, a quality almost entirely lacking in American literature where stupidity—if sufficiently sincere and authentic—is deeply revered, and easily achieved. Although this was a rather unhealthy life in the long run, Wilson had a very long run indeed. But then he was perfect proof of the proposition that the more the mind is used and fed the less apt it is to devour itself. When he died, at seventy-seven, he was busy stuffing his head with irregular Hungarian verbs. Plainly, he had a brain to match his liver. [Gore Vidal] 
• American Catholicism is in steep decline. In 2000, 2.6 million American children attended Catholic schools. In 2025, only 1.6 million did. In 2001, more than a quarter-million Catholic weddings took place; in 2024, around 107,000 did. In 2001, more than a million infants were baptized in the Church. In 2024, fewer than half a million were. These numbers presage a future in which the Catholic Church will be much smaller and poorer than it currently is. Given these facts, talk of a Catholic revival—spurred this spring by a surge in adult conversions, many in urban and university churches—is misplaced. [First Things]