Someone should notify Gary Marcus e LeCun: language models have just proven them wrong. The reductionist view that describes them as “mere predictors of the next word” is collapsing under the weight of new scientific discoveries; it is like defining a human being as “just an oxygen consumer”.
The truth is that these systems, before even churning out a single word, build an entire conceptual model customized for each query, activating hierarchies of specialized subnetworks that test the implicit logic of the conversation. This is not probabilistic prediction, but real cognitive emergency, a phenomenon that is strikingly reminiscent of what happens in the grey matter in our heads.
THEartificial intelligence modern, embodied in language models, is often dismissed with a shrug: “Well, in the end it just predicts the next word.” It’s a simplification that always makes me smile bitterly. It would be like saying that Michelangelo “just put stone on top of stone.” The complexity of what happens in the computational meanders of these systems deserves more respect and, above all, more scientific curiosity.
The researchers of anthropic e OpenAI have recently discovered something significant: within the neural architectures of their models there are specialized subnetworks that behave like the so-called “grandmother neurons” of the human brain. They are not metaphors, but real functional units that activate specifically to test complex concepts.
The Evolution of Models: From Predictors to Reasoners
It makes me smile how criticism of artificial intelligence has remained stuck in an outdated image of language models. It's like judging modern smartphones based on Nokia 3310s.
The first LLM (Large Language Models) were indeed more limited, focused mainly on the statistical prediction of linguistic sequences; systems that, while impressive, showed obvious logical and conceptual frailties. But the most recent models have made a remarkable evolutionary leap, to the point that it would be more accurate to classify them as LRM (Large Reasoning Models).
The difference? It’s not just quantitative but qualitative. LRMs don’t just predict; they build hierarchical conceptual representations that allow you to manipulate abstractions, test logical hypotheses, and generate new inferences. They can maintain coherence across long sequences of reasoning, identify contradictions, and even assess the plausibility of different conclusions.
It's as if we've gone from probability calculators to real thought simulators. Those who continue to criticize these systems as “mere statistical predictors” are essentially fighting against a ghost from the past, ignoring the evolutionary gulf that separates the first generations from current models.
The irony of chance
We take the irony as an example: a subtle concept that involves understanding the opposition between intentions and outcomes. It is not something that can be grasped simply by predicting words in sequence; it requires higher-level processing.
Both Anthropic and OpenAI have discovered these subnetworks that test the implicit logic of the query as “grandmother neurons”.
When one of the latest language models recognizes the irony in buying an alarm clock and still being late, it isn't following a predefined script. It's activating a neural network that specifically identifies the contradiction between the object's purpose (waking up on time) and the result achieved (being late).
This ability to grasp such subtle logical contradictions cannot emerge from simple statistical prediction. There is something much deeper at play; something that, frankly, should make us reconsider the limits we have imposed on our definition of “understanding.”
There is an emergent logic in linguistic models
Language models, I repeat, have reached a threshold of accuracy that goes well beyond simple probabilistic concatenation. They understand the logical function of words such as “because”, “but”, “despite” and use them correctly to build new inferences.
But here is the crucial point that is often ignored: even our biological neurons, if we want to be consistent in the analysis, would be nothing more than “probabilistic predictors of patterns”. The difference is not one of nature, but of organization and complexity. When we criticize language models as “just predictors of subsequent words,” we are applying a standard that we would never use to describe the human brain, despite the increasingly evident functional similarities.
These are not tricks, they are no longer statistical shortcuts; these systems They have developed, through training, the ability to self-organize neural networks to examine all aspects of inputs. Just like our brains, specialized structures are formed that emerge at a level higher than that of the single neuron.
This is just the latest step in a much more complex and fascinating process. The next time you interact with one of these systems, you might be reminded that behind that seemingly simple answer lies an entire universe of computation that increasingly resembles the way our own minds work.
