Scaling Factor

Large language models are transforming the way we work, but that does not mean they are reproducing human intelligence. This distinction is fundamental to understanding which jobs are likely to change and which skills may be replaced by automation.

Large language models (LLMs) are extraordinary tools. They can write code, analyse documents, synthesise information, and automate tasks that required hours of work just a few years ago. Underestimating their capabilities would be a mistake. But it would be equally misguided to regard them as an artificial version of human intelligence.

This is essentially the distinction emphasised by Yann LeCun, one of the pioneers of deep learning: developing increasingly powerful language models does not necessarily mean moving closer to human intelligence. Language is only one part of how human beings learn, reason, and interact with the world.

Language Is Not Intelligence

Human intelligence develops through an integrated system of perception, action, and experience. We see, touch, hear, move through space, and observe the consequences of our actions. We gradually build a model of the physical and social world long before we can describe it in words.

Language plays a role in this process as a tool for communication and for organising knowledge. It is neither the only source of our learning nor necessarily the fundamental mechanism through which we understand reality.

A young child learns to distinguish objects, anticipate movement, recognise people, and understand spatial relationships without needing to formulate linguistic descriptions of what they observe. Children also learn through mistakes, movement, and direct contact with their environment.

An LLM, by contrast, learns primarily from large quantities of data by predicting sequences of tokens. It can develop remarkable reasoning and generalisation capabilities, but its learning process is fundamentally different from that of a child exploring the world.

The difference is not merely quantitative. Above all, it concerns the kind of experience from which learning emerges.

A Child and a Language Model Do Not Learn in the Same Way

LeCun has repeatedly highlighted the central role of perceptual learning and world models in the development of intelligence. In their early years, human beings acquire enormous amounts of visual information, learning physical regularities and causal relationships without having to translate them into words.

Comparisons between a child and an LLM in terms of the amount of data they process are often used to illustrate this point. However, the units of measurement must be distinguished: text tokens and visual data are not directly comparable. Estimates suggesting that a four-year-old child has received an amount of visual information comparable to that used to train a language model depend on the assumptions adopted and should not be treated as an established scientific equivalence.

The essential point remains valid: human intelligence develops through multimodal, embodied experience that cannot be reduced to the quantity of text processed.

This does not rule out the possibility that LLMs may evolve towards forms of intelligence different from human intelligence, perhaps by integrating vision, action, memory, and interaction with the environment. Rather, it means we should not assume that simply scaling up language models will automatically reproduce human cognitive abilities.

AI Will Not Eliminate All Jobs in the Same Way

This distinction also has practical implications for the labour market.

LLMs perform particularly well on tasks in which language, code, or other symbolic representations form a significant part of the problem. Programming and mathematics are important examples: in both fields, some of the work can be expressed through rules, symbols, procedures, and verification criteria.

Here too, however, we should avoid jumping to conclusions. The performance of models in theorem proving, for example, requires rigorous verification: a proof that appears convincing may contain errors, unjustified assumptions, or missing steps. Formal verification systems help distinguish a correct solution from one that merely seems plausible.

In software development, AI assistants can already speed up code writing, debugging, and system maintenance. But this does not automatically mean that programmers will disappear. Producing code is only one part of the job: developers must understand requirements, design architectures, assess trade-offs, manage complex systems, and take responsibility for outcomes.

It is therefore not reasonable to infer from LLM capabilities that the IT workforce is on the verge of being eliminated on a large scale. The impact depends on workflows, how companies are organised, and whether the output can be verified.

Where AI Agents Can Actually Replace Work

A different scenario emerges in repetitive, structured, and narrowly defined activities, such as many system administration and Identity and Access Management (IAM) operations.

Managing accounts, assigning permissions, applying configurations, checking policies, and carrying out standardised procedures are tasks that can, at least in part, be described through explicit rules and verifiable workflows.

In these contexts, an AI agent does not necessarily need to understand the world as a human being would. It needs to know the relevant state of the system, have access to the appropriate tools, comply with the defined rules, and complete a task with a verifiable result.

The advantage comes from combining three elements: a relatively bounded environment, clearly specified objectives, and procedures that can be standardised sufficiently. When these conditions are met, an agent can chain operations together, check results, and manage a significant part of a process without continuous human intervention.

This could profoundly change the work of junior IAM professionals. Tasks currently performed manually—such as user provisioning, handling access requests, or applying standard configurations—could be carried out directly by AI agents, with human supervision focused on exceptions and more complex cases.

This does not mean that every system administration task is easy to automate. Real-world enterprise environments contain dependencies, legacy systems, exceptions, and unforeseen problems. But when the scope is well defined, automation can move beyond simply providing assistance and become a genuine substitute for some duties.

The most important consequence may be a shift in the skills required: less time spent manually executing procedures, and more attention devoted to design, oversight, and exception management.

Guardrails Become a Central Issue

The more autonomously AI agents can operate, the more important it becomes to define the limits within which they are allowed to act.

In an IAM system, for example, an agent might create an account, change a permission, or revoke access. A mistake would not simply be a matter of output quality: it could create a vulnerability, disrupt a service, or enable unauthorised access.

That is why it is not enough to ask an agent to behave correctly. Technical and organisational guardrails must be designed to constrain what it can do.

Essential measures include least-privilege permissions, separation of duties, human approval for high-risk operations, activity logging, independent checks, and mechanisms for stopping or rolling back actions. Wherever possible, these limits should be enforced directly by the infrastructure rather than relying solely on instructions given to the model.

The question, then, is not simply how intelligent an agent is. It is also how effectively its power can be bounded, its actions controlled, and responsibility for its decisions assigned.

Not a Replica of Humanity, but a Transformation of Work

LLMs do not necessarily need to become human-like to have an enormous impact on society. They can be exceptional tools without sharing the same cognitive architecture, experiences, or learning processes as people.

Indeed, their distinctiveness is part of their strength: they can perform specific tasks at speeds and with processing capabilities that augment human work, or automate entire processes when those processes are sufficiently structured.

The decisive distinction is therefore not between artificial and human intelligence in the abstract. It is between tasks in which AI acts as a multiplier of human capabilities and tasks in which it can take over the work itself.

In some sectors, we will see primarily an increase in productivity. In others, including certain administrative and operational activities, some duties may disappear or be radically redesigned.

We do not have to wait for machines to become human before they can replace some human work. They only need to be effective enough in a bounded environment, with clear objectives and controllable outcomes.

That is precisely why understanding the limitations of LLMs, designing reliable agents, and defining effective guardrails may prove more important in the near term than pursuing a machine that thinks exactly as we do.