OpenAI claims its new o3 model approaches human intelligence

OpenAI’s o3 model approaches human intelligence based on its 85% ARC-AGI test score, which exceeds the record of 55% set by a previous AI. The results are excellent for complex mathematics and mark a significant step towards artificial general intelligence (AGI) which targets large AI research laboratories.

The o3 model would come closer to human intelligence

However, let us contextualize the nature of the ARC-AGI test. This involves evaluating “sampling efficiency” of an AI system, that is to say its ability to adapt to a new situation from a limited number of examples. Unlike ChatGPT-4 which requires millions of examples of human text to establish probability rules on the most likely word combinations, a true general artificial intelligence is able to learn from a small number of samples.

The ability to effectively adapt, or “generalization”represents an essential element of intelligence. The ARC-AGI test assesses faculty via grid problems: the AI ​​must identify the pattern transforming an initial grid into the final grid based on just three examples.

The method used by OpenAI is unknown but the results show that its new o3 model has great adaptation capacity. François Chollet, French AI researcher and designer of the test, explains that AI uses different “reflection chains” which describe the resolution steps then selects the “best” according to a heuristic rule. This is a method similar to how Google’s AlphaGo analyzes possible movement sequences.

Advertising, your content continues below

We can compare the thought chains to programs that correspond to the examples given. The heuristic could choose the simplest or most general rules. And like AlphaGo, it’s possible that an AI has been trained to evaluate different reasoning sequences.

However, questions remain. If OpenAI’s o3 works this way, the underlying model may not be as advanced as its predecessors. Concepts learned may not be suitable for generalization. So, progress would mainly be found in the identification of more effective chains of thought based on specialized heuristics.

We still have to wait for the release of the o3 model

OpenAI has not shared more information beyond a few press presentations and preliminary testing with a limited number of researchers, laboratories and institutions specializing in AI security. It will therefore be necessary to carry out a thorough evaluation of the capabilities of the o3 model, its failure and success rates.

We still have to wait for the release of the o3 model to determine if AI is really as adaptable as a human. In this case, the impact on the sector could be major with a new era of accelerated self-improving intelligence which will require new criteria for evaluating AGI and serious reflection on its governance. As a reminder, one of the experts in this technology estimates that it has a 10 to 20% chance of destroying humanity in three decades.

Otherwise, despite this impressive result, its influence will be limited.

Advertising, your content continues below