Are World Models set to transform AI? - Polytechnique Insights

π Digital Are World Models set to transform AI? On September 2nd, 2026 4 min reading time Jean Ponce Professor at École normale supérieure-PSL and Global Distinguished Professor at New York University Key takeaways Large language models (LLMs) are leading us down a dead end, as they lack any understanding of physical phenomena. The next step involves world models (WM), which are models trained on data relating to physical environments. World models, as envisaged by Yann LeCun, are predictive models, but operate in an abstract (or latent) space, represented by a dynamic system designed to predict the causes and effects of an action. To achieve this, we must successfully replicate the ‘common sense’ that humans and animals acquire instinctively, including, amongst other things, an understanding of gravity or the permanence of objects. World models are designed to be built using a different architecture; this is the case with Yann LeCun’s model, called JEPA, which is structured around three neural network components: a target encoder, a context encoder and a predictor. Large lan­guage mod­els (LLMs) need no intro­duc­tion. Omni­present in our lives, AIs such as Chat­G­PT, Claude and Gem­ini pos­sess remark­able cap­ab­il­it­ies for pre­dict­ing the sequence of words in a text based on stat­ist­ic­al cal­cu­la­tions. How­ever, in the view of a grow­ing num­ber of lead­ing fig­ures in AI, they are lead­ing us down a dead end. Since they lack any under­stand­ing of phys­ic­al phe­nom­ena they fail to grasp the caus­al struc­ture of the real world, which lim­its their ‘intel­li­gence’. Fur­ther­more, the stat­ist­ic­al gen­er­at­ive mod­el comes up against the real­ity that is riddled with unpre­dict­able details. The world of AI is there­fore ready for the next step: world mod­els (WM). These mod­els, trained on data relat­ing to phys­ic­al envir­on­ments, are expec­ted to have applic­a­tions extend­ing far bey­ond con­tent gen­er­a­tion, par­tic­u­larly in the field of robot­ics. In recent months, there has been a flurry of announce­ments. Yann LeCun, who left Meta in 2025, foun­ded the start-up AMI ( Advanced Machine Intel­li­gence Labs ) to focus on this new approach. At Google Deep­Mind, Demis Hassabis also believes that LLMs lack crit­ic­al cap­ab­il­it­ies – not­ably reas­on­ing and a phys­ic­al under­stand­ing of the world – which world mod­els could pre­cisely provide. Fei-Fei Li, founder of World Labs, also con­siders ‘spa­tial intel­li­gence’ to be the next fron­ti­er in AI. For his part, Yoshua Ben­gio is advoc­at­ing for a safer AI sys­tem 1 incor­por­at­ing a world mod­el that gen­er­ates the­or­ies to explain the data. He argues that, without a coher­ent mod­el describ­ing how the world works, LLMs are unable to pre­dict the con­sequences of their actions. Abstract representation of the world World mod­els , at least as envis­aged by Yann LeCun, are pre­dict­ive mod­els, but in an abstract (or lat­ent) space. The idea is that, if you want to plan a robot’s move­ments, for example, it may not be neces­sary to pre­dict the evol­u­tion of an image down to the pixel, as LLMs do. “If we cre­ate a mod­el with an abstract rep­res­ent­a­tion of the dynam­ic sys­tem, it will be bet­ter able to pre­dict the rep­res­ent­a­tion of sub­sequent states,” explains Jean Ponce, Pro­fess­or of Com­puter Sci­ence at École Nor­male Supérieure. In short, it is as if the mod­el were con­struct­ing an intern­al “men­tal” space and could pro­ject itself into it. Thus, rather than mak­ing prob­ab­il­ist­ic pre­dic­tions about the next ele­ment (or token ) in a sequence, these sys­tems learn to mod­el the world across sev­er­al levels of abstrac­tion and to pre­dict how the over­all ‘pic­ture’ will evolve under dif­fer­ent con­di­tions and con­straints. This enables them to ignore irrel­ev­ant details and focus on the inform­a­tion needed to com­plete a task. For example, if we want to pre­dict the move­ment of a vehicle in a video at time t+1, it is unne­ces­sary to focus on its col­our or signs of wear on the body­work. Hence, com­pu­ta­tion could be sig­ni­fic­antly stream­lined. “The fact that this is more effi­cient has not yet been clearly demon­strated, but it is plaus­ible,” argues the expert. Whilst there is no ‘offi­cial’ defin­i­tion of what a world mod­el is, Yann LeCun’s aim is nev­er­the­less to rep­lic­ate the ‘com­mon sense 2 ’ that both humans and anim­als acquire instinct­ively, includ­ing, amongst oth­er things, an under­stand­ing of grav­ity and the per­man­ence of objects. The research­er aims for the level of ‘a cat’s intel­li­gence’, as he repeatedly states – that is, the abil­ity to nav­ig­ate a new situ­ation, to pre­dict future states of one’s envir­on­ment, but also to fill in miss­ing information. Data efficiency Whilst LLMs improve by feed­ing on ever-increas­ing amounts of data, world mod­els are designed to be built using a dif­fer­ent archi­tec­ture, which avoids ‘wast­ing’ resources on attempt­ing to pre­dict unpre­dict­able details or tra­ject­or­ies pixel by pixel. The data required to train them is mul­timod­al, but primar­ily video. Just as a baby (or a cat) observes the world and devel­ops an instinct­ive under­stand­ing of the laws of phys­ics, ana­lys­ing videos enables an under­stand­ing of space. The archi­tec­ture pro­posed by Yann LeCun—which is still a neur­al network—is called ‘JEPA’ (short for joint embed­ding pre­dict­ive archi­tec­ture ) and can be trained in a self-super­vised man­ner. JEPA is struc­tured around three neur­al net­work components: A tar­get encoder, which trans­forms, for example, a video into an abstract rep­res­ent­a­tion (the embedding); A con­text encoder, which pro­cesses a trun­cated por­tion of the input data (for example, an image with parts masked out); A pre­dict­or, which pro­jects the abstract rep­res­ent­a­tion of the masked or future portion. “There may be uncon­trolled sys­tems, such as a fall­ing stone or a raging hur­ricane, and we pre­dict what will hap­pen in a few moments’ time. But the aim is also to have con­trolled dynam­ic sys­tems, for example a robot, which will move not only under the influ­ence of grav­ity, but also in response to the com­mands we send it. If we tell it: ‘apply such-and-such a velo­city at this angle, at this joint,’ it will be able to trans­late that into action,” explains Jean Ponce. Towards planning? Due to their struc­ture, LLMs struggle to for­mu­late multi-step reas­on­ing. Con­versely, if world mod­els are able to make pre­dic­tions, they could, in the­ory, plan sequences of actions. This is why they are of par­tic­u­lar interest in robot­ics and logistics. As their learn­ing is not guided by a spe­cif­ic task, they can adapt to a chan­ging envir­on­ment. “We sep­ar­ate the learnt part – that is, the rep­res­ent­a­tion, the sim­u­lat­or – from the final task. The hope is that this will enable us to use rel­at­ively simple and gen­er­al algorithms to plan the task, which is itself spe­cified simply by show­ing the robot, for example, a photo of the ini­tial state, a photo of the goal, and pos­sibly a few com­mands in nat­ur­al language.” Whilst it is clear that world mod­els are gen­er­at­ing sig­ni­fic­ant enthu­si­asm – by mov­ing AI bey­ond its encyc­lo­paed­ic know­ledge and enabling it to act in space-time – it is still too early to know wheth­er they will be cap­able of ‘reas­on­ing’, or even of achiev­ing a form of gen­er­al arti­fi­cial intelligence. Will world mod­els live up to their prom­ises? Will they be more effect­ive, truly cap­able of incor­por­at­ing the char­ac­ter­ist­ics of the phys­ic­al world? “Only time will tell. What’s import­ant is to think about some­thing oth­er than LLMs,” says Jean Ponce. Célia Chaboud 1 https://​arx​iv​.org/​a​b​s​/​2​5​0​2​.​15657 ↑ 2 https://​open​re​view​.net/​p​d​f​?​i​d​=​B​Z​5​a​1​r​-kVsf ↑