OpenAI’s o3 model: an AI that flirts with human intelligence

© Sebastian Voortman/Pexels

Current AI systems are burdened by several major limitations, which prevent us from fully reproducing the complexity of human thought, we had addressed the subject in this article. This does not prevent the leaders of the sector from continuing their efforts to one day develop THE model that will surpass all others and maybe one day win the IAG trophy(General Artificial Intelligence). A highly advanced form of artificial intelligence that aims to create machines that can understand, learn, and perform a variety of intellectual tasks in a human-like manner.

On December 20, a quiet announcement from OpenAI shook the scientific community: their new o3 model achieved a score of 85% on the ARC-AGI (Algorithmic Reasoning Corpus-Artificial General Intelligence) test, far surpassing the previous best performance of 55% achieved by other AI systems, and rivaling human capabilities averages.

Learning efficiency, a new paradigm of AI

The ARC-AGI test is fundamentally different from conventional artificial intelligence assessments. At the heart of its operation lies the concept of “sampling efficiency” or the ability of a system to deduce general rules from a minimal number of examples. Where ChatGPT requires millions of samples to establish probabilistic language models, o3 demonstrates a remarkable ability to extrapolate general principles from a few specific cases.

This test challenges the AI ​​to visual puzzles in the form of grids, where the system must identify the transformation rules based on just three examples (see below). In concrete terms, the AI ​​observes simple geometric patterns that transform according to precise logical rules: a square can move along a line, a shape can absorb another, or elements can combine to create new configurations. To succeed, the system must not only understand these transformations, but also extrapolate the underlying logic to apply it to new cases.

Illustration of an ARC-AGI test: each example shows a grid input and its transformation according to a logical rule. The challenge is for the AI ​​to identify this rule and apply it to produce the correct output grid, as in the case of #8217;test example. © ARC Prize

This methodology is similar to traditional IQ tests, but more specifically assesses the ability to adapt quickly to new situations. The major difference lies in the emphasis placed on the system's ability to formulate coherent hypotheses from a very limited number of observations. A fundamental skill of natural intelligence that AIs traditionally struggle to reproduce.

The cognitive mechanisms of o3

The technical innovation of o3 is based on its ability to identify what researchers call “weak rules »: subtle and often implicit correlations between data that, while not universal laws, allow the model to make predictions or decisions. It is as if o3 were able to detect “clues” in the data, allowing it to better understand and interpret the world around it. AI researcher François Chollet proposes an illuminating hypothesis to explain this: o3 would explore different “chains of thought“, analyzing various sequences of reasoning before selecting the most relevant one.