modelthinking

The weakness of artificial neural networks is that the power to learn anything comes at a cost: Learning can require a significant amount of data. By contrast, learning from small amounts of data is one of the key characteristics of human intelligence: A child can learn a new word after hearing it used just once. Bayesian models of cognition, our

One of Aristotle’s key insights was that the form of an argument is what makes it valid, rather than its content.

These three frameworks—rules and symbols, neural networks, and Bayesian models—offer complementary perspectives on the mind. Each framework highlights a different kind of mathematics that is ultimately going to be critical to understanding the Laws of Thought.

In the context of value investing, this expectation error seems to be an overreaction to negative news, on average; for momentum, the expectation error is surprisingly tied to an underreaction to positive news (some argue it is an overreaction, which cannot be ruled out, but the collective evidence is more supportive of the undereaction

As we will discuss, the speed at which mispricing opportunities are eliminated depends on the cost of exploitation. Putting aside an array of transaction and information acquisition costs, which are nonzero, the biggest cost to exploiting long-lasting mispricing opportunities are career risk concerns on behalf of delegated asset managers. The

Machines for breadth, human intuition for depth

Sean Monahan, co-founder of K-Hole and writer of the 8Ball newsletter, recently argued that, with the advent of AI, the human trend forecaster will become obsolete. I disagree. What I think will happen instead is that LLMs will raise the standards for our trend research. Machine and human

Instead, what I think we’re missing is a sense of imagination, creativity, and genuine curiosity. Without these elements, the act of research itself has lost its potency, our findings and interpretations reduced to a shallow performance, where the same ideas are recycled and repackaged in an endless feedback loop.

Tuning in to weird signals

In the world of trend research, we are traditionally taught to be on the lookout for ‘weak signals’ - small, low-frequency indicators that hint at a potential shift in culture. And while scanning for these weak signals is a fundamental part of the process, it’s easy to get stuck in a pattern of only noticing evidence

On that template, AI looks like it has already finished the first act. The outer ring has cracked: neoclouds, “AI infra” smaller caps, and nuclear or power-adjacent trades are down 50% or more from their highs. That’s your paper-railway phase. The core of the story, Nvidia and a handful of megacaps, is still treated as the “safe” way to own AI, but