One theme around here is that banking is getting narrower. “Narrow banking” is the idea that you can separate (1) the business of taking bank deposits and facilitating payments from (2) the business of making loans. Traditional banks take deposits and make loans, which creates various problems; narrow banking would solve those problems, possibly at the cost of creating new ones. In narrow banking, deposits would be invested in the safest possible money (central bank reserves or short-term government bills), while loans would be made by funds with long-term equity financing. 

In this move toward narrow banking, the lending role is done by private credit, as we often discuss. Private credit firms raise funds from investors with long-term lockups, and the investors consciously bear the risk of the loans. 

But what about the role of deposit-taking? Who is going to take deposits and not make loans? Well, there are government money-market mutual funds, which do look a lot like narrow banks: They take money from depositors and park it in short-term government bills or, often, at the Federal Reserve. There was a thing called “TNB,” which proposed to take money from depositors and invest it solely in reserves at the Fed, though the Fed said no.

The hottest and strangest modern answer is stablecoins. A stablecoin is a thing that takes money from depositors, gives them back tokens that can be used in payments (an important job of deposit-taking banks), and invests the money in … well, there are all sorts of possibilities, but the standard answer these days does seem to be short-term government bills. (Or money market funds that invest in those bills.) The GENIUS Act regulating stablecoins in the US, for instance, mostly requires stablecoin issuers to keep their money in bills, repo agreements collateralized by bills, or central bank reserves, though they can keep some of it in bank deposits. 

Yesterday Andrew Bailey, the governor of the Bank of England,  published a Financial Times op-ed advocating for narrow banking via stablecoins:

The so-called backing assets (the other side of the balance sheet) should be free of financial risk in terms of credit, interest and exchange rate risk. This ensures that the value of a coin truly can be stable. 

This is not how banks work. Banks issue deposits, and the value of a $1 (or £1) deposit is stable at $1 (or £1). But the assets backing bank deposits are not free of financial risk, as we frequently notice. Obviously there is a whole social and regulatory apparatus to make bank deposits nonetheless effectively risk-free; transforming risky loans into risk-free deposits is an important bit of socially constructed magic. One could imagine extending that apparatus to stablecoins, but instead everyone seems to imagine taking a different approach, one where the stablecoins’ assets are risk-free. Bailey goes on:

Through banks, the system has combined the holding of money with the provision of credit, meaning that deposits directly support lending that underpins economic activity. This has come to be known as fractional reserve banking. In other words, most of the assets backing commercial bank money are not risk free: they are loans to individuals and to companies. Trust in money has been supported, and from time to time reinforced in the light of experience, by bank regulation, deposit insurance and resolution provisions.

The system does not have to be organised like this. It is possible, at least partially, to separate money from credit provision, with banks and stablecoins coexisting and non-banks carrying out more of the credit provision role. But it is important to consider the implications of such a change thoroughly before going ahead. ...

In the coming months, the Bank of England will publish a consultation paper on the UK’s systemic stablecoin regime, which will apply to those used in scale as money (ie for everyday payments or for settling tokenised core financial markets) and consider what standards they would need to meet. In doing so, we will set out that widely used UK stablecoins should have access to accounts at the BoE in order to reinforce their status as money.

Again, in the US, TNB — for “The Narrow Bank” — wanted to issue deposits and park its money in reserves at the Fed, but the Fed said no. It looks like if you tried that in the UK, and called it a stablecoin, the Bank of England might say yes.

The modern economy has a lot of niches with the essential form “you solve puzzles and get money.” You do not need to be an inspiring leader, you do not need to be a visionary strategist, you do not need to have your finger on the pulse of consumer demand, you do not need to be a charismatic salesperson or a hardy farmer or a skilled hunter. Just sit at a desk where someone will hand you challenging abstract intellectual puzzles, and you will have some fun using your creativity and math skills and pattern-matching abilities and sense of whimsy to solve them, and then you will hand in the solutions and get back bags of money. [1]

Maybe the highest-profile form of this in 2025 is artificial intelligence labs, where the puzzles are complex, the link between the puzzles and money is quite abstract, the puzzle-solvers are intrinsically motivated by their deep interest in solving them, and they all seem somewhat befuddled that Mark Zuckerberg wants to pay them $100 million. But AI researchers’ motivations, it seems to me, are not quite whimsical: There is a sense of mission, of social importance, of existential necessity, in building AI. The purest form of it might be in quantitative proprietary trading firms, where you solve puzzles and get money without any particularly strong sense of social mission. No kid dreams of growing up to make the spreads on cross-border exchange-traded-fund trades a bit tighter, but it turns out that doing that (1) involves solving interesting puzzles and (2) is quite lucrative. “There are few ways in which a man can be more innocently employed than in getting money,” said Johnson.

Who excels in these niches? Well, you’d expect it to be mostly people who love puzzles, people who have a deep intrinsic interest in math and puzzles and structure. Not necessarily businesspeople, not necessarily people who are interested in power and leadership and strategy; people who want to sit around and solve puzzles with their friends. Unworldly people, perhaps. Not necessarily people who are uninterested in money — if you like puzzles and don’t care about money there are lots of non-lucrative puzzles for you to solve — but people for whom the money is secondary and the puzzles come first.

But these niches are fragile. If solving the puzzles produces too much money, other people will notice, people who are not intrinsically interested in the puzzles but who really like money. Those people might be a bit less creative, a bit less whimsical, than the original puzzle-solvers, but they might be more practical and intense and careerist. They might be really good at, like, going on the internet and finding all of the previous puzzle questions that the prop trading firms have asked in interviews, and learning the solutions to those puzzles, so that when they are interviewed they can give the impression of (1) loving puzzles, (2) being good at them and (3) being more practical and dynamic and self-starting than the average puzzle-lover. And then they will get hired, and over time the business will shift from “whimsical puzzle-solvers who make hundreds of millions of dollars and stuff it in shoeboxes so they have more time for puzzles” to “ruthless businesspeople with firm handshakes and yachts.”

I wrote last year about Jane Street Group, the proprietary trading firm, which (1) for a long time had a reputation as a quite whimsical and puzzle-focused employer and (2) has recently developed a considerably more ruthless reputation. In particular, I discussed all the reports about how much money Jane Street was making, and said:

I  suppose one problem with this sort of press is that it is a potential negative for recruiting. I mean, in the sense that it is too good for recruiting. If Jane Street is an obscure haven for math nerds, it will mostly recruit math nerds and replicate its culture. If Jane Street is a famous financial services behemoth known for its minimal hierarchy and obscene profits, then it will attract a lot of resumes from people who love money more than they love math. And Jane Street’s distinction, perhaps, comes from hiring people who love math more than money and turning the math into money.

Today Bloomberg’s Sridhar Natarajan, Ava Benny-Morrison and Annie Massa have a profile of Rob Granieri, one of Jane Street’s four co-founders and the only one who still works there, who hits all the beats that you’d hope for in an old-school, I’m-just-here-for-the-puzzles-man guy:

“I sat down with a guy who had hair down to his waist in a ponytail and a backpack on,” [casino executive LuAnn] Pappas said, recalling her first meeting with him. “Someone you’d expect to see on a park bench reading Shakespeare.”

At his day job in downtown Manhattan — the casino of capitalism — Granieri slips back into the schlubby wardrobe that’s the norm at the 25-year-old firm. The soft-spoken libertarian guards his low-key stature so much that he often goes unrecognized within the company, where he officially has no title. His profile in the employee directory stands out because of its missing headshot. …

Unlike his peers who snap up trophy homes, Granieri has long preferred to live in rental housing to avoid the headaches of owning property. Still, a landlord once took him to court after Granieri missed reminders for an unpaid $10,372 bill.

Also, obviously, Burning Man. And “donations to psychedelic studies.” And accidentally funding a coup in South Sudan. Though “according to his high school yearbook, Granieri … harbored ambitions to make a lot of money.” You don’t actually get to where he is on pure whimsy.

Elsewhere in abstract ways to turn puzzles into money, I wrote the other day about Thinking Machines, the artificial intelligence startup founded by former OpenAI chief technology officer Mira Murati. Thinking Machines has raised $2 billion in funding at a $10 billion valuation, but, the Information reported, “even some of the company’s investors don’t have a very good idea of what it is working on.” I was pleased: There is a very hot and reasonably transparent market for AI talent, Thinking Machines employs a lot of high-end AI talent, so, I suggested, of course it is worth a lot of money, even without a product. The Information quoted one investor saying of Murati: “‘She was like, ‘So we’re doing an AI company with the best AI people, but we can’t answer any questions.’” Perfect pitch, I wrote.

That said, a couple of readers emailed to suggest another way to get to that valuation. Modern AI, you might think, is a winner-take-all or winner-take-a-lot-anyway business. The AI lab that builds the first superintelligence will take over the world, perhaps not metaphorically. Who will the winner be? Well, I don’t know, but if you are a venture capitalist who invests in AI, you might plausibly think:

1. The winner might be one of the AI labs you invest in.
2. It might not be Thinking Machines, sorry!, what with their lack of a product.
3. But there’s a significant risk it will be a company that you don’t invest in, perhaps because it’s a public company (like Meta or Alphabet) where you can’t invest, or just a private company (like OpenAI or xAI) that you didn’t happen to invest in.
4. Paying a few billion dollars to a few dozen top AI researchers so that they can not work at Meta or Alphabet could be a good trade. 

I don’t really buy this theory (you’d rather get the really good researchers to work at whatever lab you are betting on), but it’s fun to think about. More generally, if you are in a competitive business of solving puzzles for commercial purposes, it could be valuable to you to pay some really good puzzle-solvers to go off and solve some really non-commercial puzzles.

We talked last week about an experiment in quantum bond trading. HSBC Holdings Plc teamed up with International Business Machines Corp. to sprinkle some quantum magic on its corporate bond quoting algorithms, and allegedly it got “a 34% improvement in predicting how likely a bond will trade at a given price.” I enjoyed this as a special case of the general rule that all sorts of esoteric knowledge can be put to use most immediately and profitably in financial markets. But … how? Why do you need quantum computers to predict bond prices? Are the bonds … in two places at once? Are the bonds both a wave and a particle? “I can’t tell you the volatility, but I can tell you the price”? 

But here is the actual paper from HSBC and IBM and it is weirder than I expected. Basically they took a bunch of historical market data (bond trades and prices), used the data to train “common machine learning models,” had the models predict the likelihood of an out-of-sample historical bond trade occurring at some price given some market context, and then checked how good the predictions were. The baseline predictions were created by running the machine learning models on the market data, but then they also generated better predictions — 34% better — by running the training data through a quantum computer first:

Central to our trading data-specific exploration of a quantum enhanced approach is the notion of a feature transformation … enabled by a quantum device. … [This] represents an explicit feature engineering step, which is beneficial if we assume that, from the perspective of the learning algorithm ... at hand, the mapping [from the quantum-enhanced data to the predicted probability] is in some sense easier to estimate than the original [mapping].

Why would this help? Well, I am not a machine learning researcher and I don’t really know, but you could have might intuition like “shaking up the original data somehow makes the machine learning algorithm more robust or nuanced in its predictions.” [2] But in my defense the authors of this paper don’t know either. From the paper:

While this is a purely empirical observation of applying a heuristic method to a specific dataset without any generalization guarantees to other market environments or trading datasets, the results are still subject to many open questions. For instance, it is not understood how exactly quantum hardware noise affects our particular quantum circuit, and it is unclear how resulting noise-encoded feature vectors may benefit the analysis of noisy financial observables.

We are in pretty early days of building quantum computers, and it seems that the quantum computer hardware they used to transform the data was “noisy,” and that the noise is what actually improved the performance:

We observe a relative gain of up to ~34% in out-of-sample test scores for those models with access to quantum hardware-transformed data over those using the original trading data or transforms by noiseless quantum simulation. These empirical results suggest that the inherent noise in current quantum hardware contributes to this effect and motivates further studies. ...

The exact role that the intrinsic noise during the hardware execution of quantum circuits plays in these finally derived model performance gains is not understood.

I suppose the advantage of doing quantum computing research is that you don’t have to understand what you did to publish it. To be fair that also probably applies to bond trading so, you know, fine. But Scott Aaronson writes:

They see a quantum advantage for the task in question, but only because of the noise in their quantum hardware? When they simulate the noiseless quantum computation classically, the advantage disappears? WTF? This strikes me as all but an admission that the “advantage” is just a strange artifact of the particular methods that they decided to compare—that it has nothing really to do with quantum mechanics in general, or with quantum computational speedup in particular.

Indeed, the possibility of selection bias rears its head. How many times did someone do some totally unprincipled, stab-in-the-dark comparison of a specific quantum learning method against a specific classical method, and get predictions from the quantum method that were worse than whatever they got classically … so then they didn’t publish a paper about it?

Perhaps the main lesson here is “if you could use a quantum computer to make bond trading more profitable, that would be both lucrative and extremely cool, and banks would be interested.” Plus banks and proprietary trading firms have a lot of money to spend on computers, and are in an arms race with each other to trade more quickly and correctly, so if this worked you could sell a lot of quantum computers. 

My theory of “tokenization” — turning various financial or physical assets into crypto tokens that can trade on the blockchain — is that most people who talk about tokenization pretend that it is a computer technology, but really it’s a regulatory technology. When people say stuff like “we should tokenize real estate so that people can buy and sell fractional shares of buildings” or “we should tokenize private company stocks so that ordinary investors can buy shares of OpenAI,” they want you to hear “this is a technological breakthrough that will finally allow trading of these assets,” but that can’t be true. There have, for decades, been real estate investment trusts that allow people to buy fractional shares of (portfolios of) buildings. There have, for centuries, been stock markets that allow people to buy shares of companies.

It’s just that, in the US in 2025, there are rules — securities laws — about how you can offer shares of companies or buildings to public investors. Those shares are “securities,” and they are subject to a lot of disclosure obligations, and a lot of people would prefer not to comply with those obligations, sometimes for reasons of cost (the rules are quite burdensome) and sometimes for reasons of, you know, fraud. “Tokenization” carries the promise of starting over. A token, the theory goes, is an entirely new category, not a security, so it’s not subject to securities law. So maybe you can sell shares of buildings or OpenAI to the public without disclosure, if you call them tokens.

It has always seemed to me that this theory is completely wrong, and that a “tokenized” share of a company or building obviously is a security. This was also the opinion of the US Securities and Exchange Commission until recently. Now, uh, maybe not. So we are in a boom for tokenization, or at least for talking about tokenization.

If your theory of tokenization is “this is 100% a regulatory arbitrage and not at all a technology thing,” you will naturally spend less time on building technology and more time on building, uh, regulatory relationships. Anyway here’s this:

Zach Witkoff would like to make the Trump family’s real estate portfolio available as tokens on blockchain, giving access to a wider pool of investors.

“The Trump family has one of the most exciting real estate asset portfolios in the world,” Witkoff said in an interview at Token2049 in Singapore, speaking alongside Donald Trump Jr. “What if I told you that you could, you know, go on an exchange and buy one token of Trump Tower Dubai?”

Witkoff is the son of US special envoy to the Middle East Steve Witkoff and co-founder of World Liberty Financial, started together with the Trump family last year. The company already issues a stablecoin, USD1, and is planning to roll out features like lending and borrowing. He is also chairman of ALT5 Sigma, a publicly-traded treasury company holding WLFI, another token issued by World Liberty. 

Oh sure! Why would you want to start tokenizing real estate specifically with the Trump family portfolio?

One relatively new development in modern academic finance is that academics can now use artificial intelligence tools to analyze how cringe text is, which is useful in various contexts. Most classically, “if a company’s earnings call transcripts are cringe, is that a good or bad signal for the stock price,” but a good second-order question is “if a sell-side research analyst’s LinkedIn profile is cringe, are her stock recommendations likely to be good or bad?” To ask the question is to answer it (bad):

Based on analysts' LinkedIn profiles, we identify a robust negative correlation between the expressive tone conveyed in their self-presentations and forecast accuracy, particularly among male analysts, less experienced professionals, and those with fewer LinkedIn followers. We interpret this tone-performance inverse effect not as a behavioral bias, but as a strategic self-promotion behavior aimed at compensating for skill gaps and increasing visibility. … Low-accuracy but high-tone analysts achieve the highest career promotion rates relative to their peers.

That’s from the abstract to “Empty Vessels Make the Most Noise: Analyst Self-Promotion Behavior and Market Outcomes,” by Chun Liu and Shunzhi Pang. From the paper:

Based on annual earnings forecasts for 3,336 U.S.-listed firms between 2021 and 2023, we find that analysts with a more positive tone in their self-presentations demonstrate systematically lower forecast accuracy. This counterintuitive finding, which challenges the conventional wisdom linking self-assurance with competence, is termed the tone-performance inverse effect. The results remain consistent across alternative forecast accuracy metrics, time horizons, large language models, and model specifications, suggesting that it is primarily driven by analysts’ individual characteristics. The heterogeneity analysis shows that this paradox is more pronounced among: (1) male analysts, (2) individuals without formal training in economics or management, (3) professionals with limited LinkedIn social capital (measured by follower count), (4) junior analysts with less work experience, and (5) those affiliated with smaller brokerage firms. As these demographic subgroups are subject to heightened career insecurity and competitive pressures within the sell-side research industry, they appear to strategically employ inflated positive self-presentation as a compensatory impression management tactic, a behavioral pattern that directly corroborates the compensation for skill gaps mechanism.

I naturally looked for an appendix with examples of “high-toned” analyst LinkedIn profiles, but there wasn’t one, so one has to guess what that sounds like. (I can guess.)

Incidentally, I once stumbled upon a LinkedIn profile that appeared to belong to Rob Granieri. I recall that it gave his name as “rob g” and his occupation as something like “trader, proprietary trading firm,” with not much else. Warren Buffet doesn’t even have a LinkedIn! We are all familiar with the opposite sorts of LinkedIn profiles. Here I will boast that my own LinkedIn is flawlessly low-toned so, you know, count on me for stock picks.

OpenAI Valuation Reaches $500 Billion, Topping Musk’s SpaceX. A Meta Change on Publishing Research Causes a Stir in Its AI Group. Bridgewater Soars 26% to Lead Pack of Biggest Hedge Funds. Crypto Stockpiling Craze Cools After Red-Hot Summer. Musk Loses Bid to Move SEC Suit Over Twitter Stake to Texas. Insurance Executives Become Billionaires in