Just finished watching Geoffrey Hinton’s interview on The Diary of a CEO — a rare, candid conversation from one of the pioneers of AI.

Takeaways

1. Real AI risks are already here Hinton emphasizes that the pressing dangers aren’t sci-fi superintelligence, but things we see today: algorithmic manipulation, the amplification of bias, and disinformation at scale.

Cartea, Álvaro and Jin, Qi and Shi, Yuantao, The Limited Virtue of Complexity in a Noisy World (April 02, 2025). Available at SSRN here or here.

Summary: In this paper, the authors analyse the role of model complexity in the context of predicting asset returns and portfolio construction. In particular, they aim to address the significant question of whether adding a large number of predictive features ultimately harms performance. Their work aims to bridge two views: the traditional econometric one, favoring parsimonious models (i.e. Occam’s razor), and the more modern machine learning findings that highly the fact that so-called “overparameterized” models can perform well under proper regularization ( double descent phenomenon).

The authors set up a framework where investors predicts excess returns using a large number of features, but these features are “contaminated” by noise (which can arise from data collection gaps, computational approximations, or other infrastructure limitations). They examine how this affects the Sharpe ratio of a timing strategy, which the investor seeks to maximize, and the out-of-sample R-squared of return forecasts. Given the high-dimensional setting, they use ridge regression and apply Random Matrix Theory classical results to characterise the asymptotic behaviour of these metrics, focusing on the case where true features are independent (a quantitative investing).

Natural History Ensemble, no. 11 (1596–1610) by the talented Flemish draughtsman Anselmus Boëtius de Boodt (1550–1632). Check out more of Anselmus Boëtius de Boodt’s works in this Rawpixel Gallery.

Source: Public Domain. Original from the Rijksmuseum.