Vedrana Pribičević: Large Language Models Are Not Intelligence

Written by: Vedrana Pribičević, ZŠEM

The intensification of the debate on artificial intelligence and its economic effects in the Croatian public space requires a more precise delineation of terms. In the dominant discourse, large language models are often treated as a synonym for artificial intelligence as a whole, and such reduction obscures the key difference between statistical sequential models and systems with a persistent internal representation of the world.

In technical terms, LLMs (large language models) are models for predicting the next token conditioned by context. They do not possess explicit causal structures and stable models of physical and social reality. In this sense, their ‘intelligence’ is a function of statistical regularity, not understanding, therefore further increasing the size of the model yields diminishing returns in terms of persistence, reliability, and long-term planning.

World Models

It is precisely this limit that Yann LeCun warns about, who, after many years of work at Meta, has redirected his operational focus to the development of architectures based on world models, believing that further scaling of language models does not solve the fundamental problem of understanding. His thesis is that without an internal representation of the laws of space, time, and causality, there is no transition to systems capable of stable predictive planning and autonomous action in complex environments.

The world model is therefore a mechanism that allows for the simulation of outcomes before action and the optimization of behavior in the real world. Another doyenne of AI, Fei-Fei Li, shares the belief in the need for such a shift by working on embodied systems of artificial intelligence and visual-spatial perception as prerequisites for modeling the dynamics of the environment.

Both have founded companies: Li raised one billion dollars in the latest funding round, and LeCun’s company is expected to have an initial valuation of three billion dollars. The half a million citations that LeCun and Li have together deserve the trust of investors.

The economic effects of artificial intelligence largely depend on how quickly world models will develop. LLMs primarily act as augmentation technology in which information is intensively used, increasing the efficiency of administrative and cognitive tasks, reducing transaction costs, and accelerating data processing, thus their macroeconomic effect is incremental and concentrated in services.

World models, especially in combination with physical embodiment in robots and autonomous systems, have the potential to change the very production function. A system that can plan in space and time, anticipate consequences, and adjust actions enters the realm of the real economy – energy, logistics, manufacturing, and infrastructure – at which point artificial intelligence becomes a complement to capital equipment, not just a software tool.

The Narrative of Inevitability

However, the broad industrial application, even of LLMs, faces institutional limitations. The European AI Act sets a framework for risk and liability classification, but operational technical standards are still being developed within CEN-CENELEC and similar initiatives.

Without clear mechanisms such as tracking the origin of data and models (chain of custody), measurable performance, and precise allocation of responsibility, the integration of AI systems into critical infrastructure, the financial system, or public administration remains limited. Industrial diffusion requires standardized protocols and regulatory predictability; otherwise, application remains partial and experimental.

Additionally, the current dynamics of artificial intelligence bear the characteristics of a typical hype phase. In such phases, not only is technology spreading, but also the narrative of its inevitability. Media discourse, investment capital, and technological marketing mutually reinforce each other, creating the perception of immediate and universal transformation. Valuations then reflect expectations of future monopoly or infrastructural dominance, rather than current productive effects.

The history of general-purpose technologies shows that between technological breakthrough and macroeconomic transformation lies a period of adjustment in which standards, regulatory frameworks, and complementary skills are built. Industrial application therefore depends on reliable and interoperable systems integrated into global value chains, not on media narratives.