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]