OBEN
Generative artificial intelligence (AI) was recognized with two Nobel Prizes in 2024, a remarkable achievement given the technology’s youth. The Deep Learning revolution that enabled generative AI began in 2012 when a deep neural network called AlexNet won the ImageNet computer vision challenge. Geoffrey Hinton of the University of Toronto, who led the team, was awarded the Nobel Prize in Physics alongside John Hopfield, a fellow contributor to the algorithmic foundations of artificial intelligence.1
The Nobel Prize in Chemistry was awarded to three researchers who applied generative AI to predict the three-dimensional structure of proteins. Among the recipients was Demis Hassabis, whose team at Google DeepMind developed AlphaFold, a deep learning system that determined with high accuracy the structure of all 200 million known proteins. It was a short time from AlphaFold 2 winning the protein structure prediction challenge CASP14 in 20202 and the publication of the AlphaFold Protein Structure Database in 20213 before the Nobel Prize Committee recognized this work.
Quantification has long been inspired by Newtonian mechanics, which emerged in the 17th century and laid the groundwork for new quantitative methods. In their quest to describe the movements of celestial bodies from imperfectly measured observations, Carl Friedrich Gauss and Adrien-Marie Legendre developed least squares regression around the turn of the 19th century.4 It wasn’t until Sir Francis Galton studied biological inheritance in the 1870s that least squares regression was used to quantify causal relations.5 These structural regression models of hypothesized causal effects have long been the backbone of predictive modeling in insurance, especially with the rise of generalized linear models (GLMs) in the 1980s.
The advent of machine learning has challenged the dominance of structural models in predictive modeling. Structural models generate predictions based on quantification of manually specified relations, and these relations are typically interpreted as causal. Causal interpretation is associated with the stability of the estimated relation, a link that may be stronger in the physical world than in social and economic settings. Machine learning, on the other hand, is agnostic to casual relations and makes predictions solely based on connections.6
In regression analysis, the number of explanatory variables cannot exceed the number of observations, and the functional form of their influences must be manually specified.7 In contrast, machine learning is not constrained by the number of variables that can contribute to the prediction, and the form of these contributions does not need to be specified in advance. Traditional machine learning relies on humans to extract the attributes that represent the raw data before processing by the algorithm.8 Deep learning, however, performs automatic feature extraction, reducing the need for human intervention.9
At their core, deep learning systems like OpenAI’s GPT‑4 for text generation and DALL‑E 3 for image generation create content by averaging over existing data, a method rooted in regression models. Gauss and Legendre developed ordinary least squares to draw a line through the center of a data cloud, where the center is defined by the arithmetic means of the variables on both sides of the regression equation. Besides the mean, other important statistical averages include the median (the central value of ordered data) and the mode (the most likely value).
In conventional text generation, a large language model predicts the next word by invoking the mode, choosing the word most likely to occur in the given context. Similarly, when an image generation model is asked to predict a cat, the system generates an “average cat.”10
Reasoning capabilities extend the functionalities of generative AI models beyond merely averaging over existing data to generating new data. To build intuition, let’s consider a popular syllogistic argument:
All men are mortal.
Socrates is a man.
Therefore, Socrates is mortal.11
The first two lines are premesis, which we can consider as existing data. The third line, a conclusion, is derived through reasoning, representing new data.
AlphaFold 2, in predicting the 3D structures of proteins, incorporates elements that parallel human spatial reasoning.12 There are also mathematical reasoning models, most notably Google DeepMind’s AlphaProof and AlphaGeometry 2. These two models together solved four out of six problems at the 2024 International Mathematical Olympiad (IMO), achieving a score in the high end of the silver medal category.13 Among large language models with advanced reasoning capabilities are OpenAI’s o3 and DeepSeek’s R1.
The predictive power of deep learning has significant implications for the role of quantitative methods in decision-making. For example, AlphaFold solved the 50‑year-old problem of protein folding without understanding the folding process itself.14 Similarly, the Artificial Intelligence Forecasting System (AIFS), a deep learning system for weather forecasting introduced by the European Centre for Medium-Range Weather Forecasts (ECMWF) in February 2025, outperforms the most accurate physics-based models for many target variables without understanding meteorology.15
Just as these generative AI systems lack an understanding of biology or physics, we do not fully comprehend how these models arrive at their predictions.16 Deep learning systems provide predictions without the opportunity for causal interpretation (“why”)17 and related counterfactual inspection (“what if”). However, causal interpretation is crucial for forming narratives, which help us organize information and communicate it coherently. Narrative plays a key role in corporate decision‑making.18
Generative AI can be viewed as an abstraction layer. In software engineering, abstraction layers are designed to hide the inner workings of subsystems. Also in science, abstraction layers simplify complex systems. For instance, chemistry serves as an abstraction layer over physics, allowing us to understand chemical processes without fully grasping the underlying physics.19
Consider how we drive motor vehicles without fully understanding the physics that powers them. When we learned to drive, we formed an abstraction layer. We verify this abstraction based on predictive benefits. For example, we predict that stepping on the accelerator will make the car move faster, and stepping on the brakes will slow it down. When driving an unfamiliar car, our predictions may be slightly off, and we adjust our abstraction accordingly.
The Nobel Prizes for AI marked a milestone, not only in scientific discovery but also in decision-making. Generative AI systems provide powerful predictions that resist causal interpretation and the formation of accompanying narratives. These systems require us to abstract from the underlying causal forces and focus on verifying the predictive benefits. Abstraction is a well-known concept in both science and everyday life.
