THE’artificial intelligence is it reaching our intelligence? No, but it has made great strides in recent years. Language models (LLMs) such as ChatGPT, Gemini, and Claude have demonstrated amazing capabilities in text comprehension, problem solving, and content generation. The power of these models has reignited a decades-long debate onartificial general intelligence (AGI)that is, a system capable of think And learn like a human beingtackling complex tasks and transferring skills between different contexts. Currently, although language models are becoming better and better at solving structured problems, they are missing still of capacity Of abstraction And generalization necessary to arrive at the AGI.
What is Artificial General Intelligence (AGI) and why we talk about it
The AGI represents the “Holy Grail” of AI: a system capable of reason, plan And learn autonomously, like a human being. Just like the Holy Grail, so far it has been more legend than fact and the companies they do to competition for those who can reach it first. The term AGI entered public debate around 2007 and, in recent years, i progress in the language models of large dimensions (LLM) have led some researchers to seriously consider the idea that a form of AGI could be imminent and it could come from the LLMs. Unlike current AI, which specializes in specific tasks – such as playing chess or generating text – AGI could tackle a wide range of problems without needing detailed instructions. This achievement is not just a scientific curiosity: AGI could revolutionize sectors such as medicine, the fight against climate change and the management of pandemics using AI capable of “to reason”.
But where are we with the reasoning ability of LLMs?
ChatGPT has gotten much better at solving complex problems
The latest model behind ChatGPT released by OpenAI, ChatGPT o1solved the problem correctly83% of problems in an American mathematics competition (the AIME), while the previous one, GPT-4ohad only solved the 13%. The next model, ChatGPT o3not yet released to the general public, would seem to have resolved the issue 96.7%. This clear improvement is due to the integration within the models of an approach called chain-of-thought prompting (CoT)which involves showing an LLM an example of how break down a problem into smaller steps to solve it or in asking him to tackle the problem step by step.
However, despite these improvements, o1 It has its limitations and it does not yet constitute an AGI. Two different research groups have highlighted that he isn’t capable yet to solve problems planning solutions with many steps and that he is unable to abstract and generalize visual problems. Both of these tests were designed specifically to evaluate progress towards AGI. In fact, an AGI is expected to be able to solve problems by planning solutions with a large number of steps (currently performance decreases when there are more than 20 steps) and to be able to deduce an abstract rule from some examples and apply it to new cases, a task that humans perform with relative ease.
To abstract you need to build a “model of the world”
According to neuroscience, human intelligence derives from the brain’s ability to create a “model of the world“, one representation of the environment surrounding that allows you to plan, reason and imagine future scenarios. To achieve AGI, AI models will need to develop this ability, allowing them to generalize learned skills and address new challenges by simulating possibilities and predicting consequences.
Several studies have suggested theemergence of rudimentary “models of the world”. within the LLMs, which however they are not always reliable or they are not used to make decisions. For example, a Harvard research team trained an AI model to predict taxi turns on New York routes and achieved excellent results. Analyzing the model of the world created by AI, however, we saw that it contained roads with physically impossible orientations And overpasses over other roads. When the researchers tried to include unexpected detours, the model failed to predict the next turn, suggesting that Not is able to adapt to new situations.
AGI could still take at least ten years
The arrival of thegeneral artificial intelligence is still uncertain and the estimates of the experts vary considerably. Some say we may be just a few years away from achieving this goal, while others believe it will take us at least ten years or more. The lack of consensus reflects the complexity of the problem and the multiple challenges still to be addressed.
What is certain, however, is that in addition to researching how to build AGI, it is also necessary integrate safety into the design and regulation of AI systems. Research needs to focus on training models capable of ensure the safety of their behaviourfor example by calculating the probability that the model violates certain security constraints and rejecting potentially dangerous actions.
