Peter Berezin

Growth in average hourly earnings fell to a 5-year low in August. However, BCA's wage survey indicator, which tends to be leading indicator, suggests that wage growth could pick up in the coming months. https://t.co/isy5RvX2Fc

This is one of the most important charts for bond investors. It shows that the correlation between equity returns and bond yields has turned negative. Long-term bonds are no longer a good hedge against declines in stock prices. This has reduced the demand for bonds, pushing up yields.

Why are bond yields rising? A key reason is that interest payments on the federal debt have exploded and are expected to rise even further. https://t.co/hPRbwwfaMl

The volume of data transmitted over the Internet has grown by 42% per year over the past quarter century, and yet spending on Internet infrastructure has stagnated. Efficiency gains in data transmission have been a phenomenal, especially in fibre optics. As the chart below shows, similar efficiency gains are occurring with large language models. Could we end up with an AI-powered economy without the need for trillions of dollars of spending on data centers?

The “most hated bull market in history” also features the most optimistic long-term earnings growth expectations in history. The math ain’t mathing on that one. https://t.co/J4XLEoDvDX

People who argue that compute will remain scarce indefinitely should look up “Turing machine”. Computation is completely universal: It can be done in a vacuum tube, transistor, or even by a polarized photon within a quantum computer. https://t.co/XmbasvnWmx https://t.co/Cjqs7UbcEu

Coinbase burning more tokens but spending less on tokens by shifting to cheaper open-source models. Many other companies doing the same. https://t.co/X91Bwr1exM

In the initial phase of the dotcom bust, tech stocks crashed while the rest of the market went up. We may be entering a similar phase now. Ironically, this means that a good hedge for stocks may be.... other stocks. https://t.co/zrptfUj8w9

OK, so China is less than 12 months away from having a Mythos-level model. Since Chinese models are generally open source, this means that Russia, North Korea, and whoever else are also within striking distance of having access to a model that cracked most NSA systems in just a couple of hours. Cybersecurity is going to be a huge investment theme in 2027.

Much of the popular investor narrative around AI presupposes that there aren't diminishing returns to LLM development. If there are diminishing returns, then the economic advantage of having a 6-month lead for frontier models will shrink over time. This would be a huge problem for the frontier labs. People who use LLMs don't interact with other users of the LLM, so being the first to have a large installed base of users is unlikely to produce the sort of natural monopolies one sees in, say, social media. What you end up with is a fairly competitive environment for AI systems. Great for the consumer, but not so great for the Open AI or Anthropic.