world models, ai 2027/2040, and what if europe is key

tldr at bottom

so ever since i got into ai safety, everyone has been telling me to read ai 2027 and ai 2040, and i finally got around to both of them this past weekend.

my first intuition was basically, "hmm, this seems plausible, even if the whole story depends on a few primary assumptions that could very easily break in different directions."

ai 2027 is compelling because none of its individual steps feel completely insane when you read them one at a time. coding agents keep getting dramatically better. ai labs are obviously incentivized to point those agents at their own research. each step follows.

then you zoom out. the intelligence explosion needs several uncertain transitions to happen in the same direction, at the same time, without some deeper bottleneck interrupting the curve.

one housekeeping note before i get into any of this. i am going to argue with ai 2027's original takeoff model, and the ai futures project has since published a scenario with later dates. plan a's counterfactual is fully automated ai r&d in 2030 with superintelligence by the end of that year, rather than the 2027 explosion in the earlier scenario. its authors say the shift is meant to reflect uncertainty across a portfolio of scenarios, rather than a simple revision to one median forecast. i still think the original model is the right thing to argue with, because the structure of the claim is what matters, and that structure did not change when the dates moved.

the assumptions they made

i ended up reducing the whole takeoff story to four assumptions that feel especially important to me.

1. agent horizons keep scaling until models can reliably complete work lasting weeks or months.
2. coding skill transfers into research skill, including experiment design, research taste, and scientific judgment.
3. automated research creates huge r&d speedups, which then recursively improve the agents doing that research.
4. no major architectural wall appears before current systems become reliable long-horizon autonomous researchers.

the first three are basically visible inside the forecasting framework itself, which forecasts a superhuman coder and then models a software-driven intelligence explosion through large ai r&d progress multipliers. the original takeoff forecast gave a median of roughly one year between the superhuman coder milestone and artificial superintelligence, conditional on that coder arriving.

the fourth assumption is the one that kept bothering me afterward. it is much harder to capture inside a smooth extrapolation of coding task horizons.

if the architecture itself starts failing once agents need persistent state, causal understanding, or planning across thousands of changing dependencies, then the curve could flatten for reasons that have almost nothing to do with raw coding competence.

and this is where i remembered that i have been reading about world models recently, and following yann lecun's research for a hot min, which immediately made the entire timeline feel much stranger.

then i remembered world models

the world model thesis starts from a pretty intuitive complaint about current ai, once you ignore the terminology and think about what a genuinely autonomous agent has to do.

an autoregressive model receives a history and predicts what should come next. that objective has somehow turned out to be insanely powerful across language, code, images, and increasingly complicated forms of tool use.

the harder question is whether repeatedly predicting the next useful token or action eventually gives you an agent that maintains a stable understanding of an evolving world across weeks, months, and millions of intermediate observations.

imagine an autonomous ai researcher halfway through a three-month project. twenty experiments have failed. five assumptions have changed. several collaborators have modified the codebase. one weird result from six weeks ago suddenly becomes relevant again.

that agent needs more than excellent local reasoning. it has to preserve a coherent belief about what is true, what remains uncertain, which interventions caused which outcomes, and which parts of its earlier plan should now be discarded.

lecun's proposed architecture treats this kind of persistent internal world representation as central to autonomous intelligence, with hierarchical planning operating over learned representations rather than requiring the system to regenerate every unpredictable detail of future observations.

that distinction matters. the world contains an absurd amount of entropy that an agent should usually ignore, while planning depends on preserving exactly the parts of reality that determine which actions remain possible.

if i move a cup across a table, i care about its location, orientation, stability, and relation to nearby objects. i do not care about predicting the exact future arrangement of every reflection and pixel around it.

jepa-style training tries to predict useful representations in latent space, which gives the system an opportunity to model the underlying state and dynamics without spending most of its capacity reproducing irrelevant sensory detail.

why world models could actually solve the planning problem

the planning argument gets much clearer when you write down what an action-conditioned world model is trying to learn.

the system first maps its observations into some latent state that represents the information needed for future prediction and decision making. it then learns how that state is expected to change when the agent takes a particular action. that means the model can internally test candidate actions before committing them in the real environment.

a planner can therefore ask which sequence of actions leads from the current latent state toward a desired future state, updating the plan whenever new observations show that its internal prediction was wrong.

this is much closer to model predictive control than ordinary chatbot reasoning. the agent predicts consequences, acts, observes reality, updates its state, replans.

the important part is that this has already moved beyond a philosophical proposal, even though current systems remain extremely far from the general version lecun ultimately wants.

v-jepa 2 post-trained an action-conditioned latent world model on less than 62 hours of robot video, without task-specific training or reward labels, then planned zero-shot on franka arms in two different labs. one caveat matters more than it first looks: for the longer-horizon pick-and-place results, the paper hands the planner a series of visual subgoals. the flat planner was doing short hops between waypoints a human decomposed for it.

that gap is exactly where the more recent work lives. hierarchical latent world models learn higher-level macro-actions and slower-timescale dynamics, so the long-horizon model produces subgoals for the short-horizon model instead of a human producing them. on real-robot pick-and-place from a single goal image, hwm reports 70 percent success where the flat v-jepa 2-ac planner gets zero, with up to four times lower planning cost across the long-horizon settings they evaluate.

i actually think this result makes the world model thesis more convincing precisely because it also shows how unfinished the problem is.

the base world model still suffers from compounding prediction errors and exploding action search when the horizon gets long. researchers are having to build explicit temporal hierarchy into the planning process rather than assuming scale magically removes those difficulties.

that looks much more like the architecture of a persistent agent to me. long-term goals live at a slower level while lower-level models handle the immediate details required to reach the next subgoal.

if something like this turns out to be necessary for serious autonomous research, then ai 2027 has a potentially enormous missing interval between "superhuman coder" and "superhuman ai researcher."

maybe that interval lasts two years. maybe twenty. maybe someone collapses it next month. the existence of the bottleneck changes the shape of the forecast either way.

the obvious objection: just write things down

here is the strongest counterargument, and i want to give it a fair hearing rather than a footnote.

a robot arm needs a latent world model because the state of a table is not written down anywhere. a research environment is different. the state is already symbolic and already external. git history, experiment logs, wandb dashboards, a running markdown file, the notebook where you record what broke and why. an agent that writes things down and reads them back has a world model that lives outside its weights, in a format that happens to be exactly the modality autoregressive systems are best at.

if that works, the bottleneck dissolves. long-horizon research agency becomes a scaffolding problem, and scaffolding is famously the thing that improves fastest because it does not require anyone to retrain a foundation model.

my honest reply is that scaffolding gives you storage. it does not give you calibration.

writing "experiment 14 failed, loss diverged at step 3k" into a log is cheap. what an autonomous researcher needs is a model of which interventions cause which outcomes, how confident to be in a result given how it was measured, and which of its beliefs should shift when a new observation arrives. that structure has to be learned from interacting with the environment. you cannot log your way into it, in the same way that a lab notebook does not make an undergrad a good experimentalist.

so the crux, for me, is whether causal and predictive structure over a research environment can be acquired through next-token prediction over text about that environment, plus enough rl. i genuinely do not know. i notice that the thing failing in current long-horizon agent evals is rarely retrieval and often something closer to judgment about consequences, which i take as weak evidence for the world model side. weak evidence, though. this is the part of my argument that could just be wrong.

why world models might be friendlier to interpretability

this is the part i actually care most about, because world models seem structurally more conducive to interpretability even though they are absolutely not automatically interpretable.

by structurally more conducive i mean that the architecture gives interpretability researchers stable intermediate objects to investigate, instead of forcing every safety question through one giant end-to-end mapping from context into behavior.

a monolithic agent policy maps an enormous history directly into an action. that leaves interpretability researchers reconstructing which internal features represented the state, which encoded possible futures, and which computations actually drove the decision.

a world model architecture gives you more explicit interfaces between those functions. perception produces a latent state. dynamics predict future states under actions. planning selects trajectories using those predictions.

that decomposition creates specific questions a method can target. what does the model think exists? which variables stay stable through time? what consequences does it predict from an intervention? where does its prediction diverge from reality?

there is already empirical evidence that jepa representations can become spatially structured, semantically coherent, and temporally consistent. v-jepa 2.1, for example, reports strong dense object-tracking results alongside its global video performance. those properties matter, because an interpretable world state is much more plausible when meaningful entities and dynamics persist inside the representation instead of being entangled with transient surface details. there are also papers explicitly trying to make world model latents correspond to physical quantities, which suggests interpretability can be part of the representation design problem rather than a post hoc explanation exercise.

now two caveats, and the second one is the one i would want a critic to raise so i can say i already agree.

first, ordinary jepa representations can still be dense and opaque enough that researchers have proposed sparse jepa variants specifically to improve legibility. a world model gives you nothing for free. the latent state can still be a gigantic vector whose internal organization makes no sense to a human.

second, and more important: modular interfaces are a hypothesis about the computation, not a guarantee about it. calling one module a planner does not mean planning happens there. the "state" can smuggle policy. the "dynamics" model can absorb perceptual work that the encoder failed to do. we already know from mechanistic interp that the boxes people draw on architecture diagrams routinely fail to carve the computation at its joints. verifying that the interfaces mean what they are labelled to mean is itself interpretability work, and it is work nobody has finished.

the claim that survives both caveats is narrower, and i think still worth a lot. world models give safety researchers an architectural place to demand transparency, because the system already needs representations for state, transitions, actions, uncertainty, and planning in order to function reliably. if those interfaces are deliberately shaped around sparse, causal, object-centered, or otherwise legible representations, interpretability becomes part of the architecture that enables capability rather than a forensic exercise performed afterward.

this is roughly the bet i am making in my own research, where the whole point is buying legibility through structural constraints on a vision-language-action model instead of trying to reverse-engineer an opaque policy after the fact. the cost is real. you give up some freedom in the function class, and you have to argue that what you bought back is worth it. having spent a while inside that tradeoff, i am more optimistic about legible-by-construction than i was, and much less optimistic that structure alone does the work.

one more thing i should flag rather than let someone flag for me. "safety becomes a source of capability" is a close relative of arguments that have historically been used to justify racing, and it is a prediction rather than an observation. interpretability has not yet clearly produced a durable commercial advantage at the frontier. i am claiming it plausibly could in a world where systems have to be trusted with irreversible physical actions. that is a claim about the future, which is the kind of claim that is usually wrong.

the europe thought experiment

right, this is the speculative part, and i want to be honest that it is speculative.

once i started thinking about world models as a possible architectural bottleneck, i wondered whether they could also scramble the geopolitical map that ai 2027 and most frontier ai conversations take for granted.

the current story is overwhelmingly organized around the united states and china. that makes sense when the central strategic ingredients are enormous compute clusters, frontier labs, semiconductor supply chains, capital, and the ability to scale existing architectures quickly. europe has serious research and several strong companies. the strategic conversation still reflects a frontier concentrated in american and chinese firms, which ai 2040 also treats as the central relationship requiring intervention.

so here is the naive version of my thesis: an architectural transition rewards different strengths, lecun's new lab is in paris, therefore europe.

that version is wrong, and i want to kill it before someone else does.

if the next bottleneck rewards physical-world data, robotics supply chains, manufacturing, and industrial deployment, the obvious beneficiary is china, not europe. china has the robot fleets, the component supply chain, the deployment volume, and an enormous amount of embodied ai research. the united states has world labs, deepmind's genie line, nvidia's cosmos work, physical intelligence, and every hyperscaler with a robotics ambition. the world model paradigm is not european. transformers being invented at google did not make attention american property, and jepa being lecun's idea does not make world models french.

the conditional version is more interesting, and it is the one i actually want to put out as food for thought.

world models favor whoever holds the specific advantages that matter once intelligence acts continuously inside physical systems that can hurt people. that is a different list from the one that decides who trains the biggest language model. it includes safety-critical engineering practice, certification and assurance regimes, industrial automation depth, precision manufacturing, and regulatory machinery that already knows how to decide whether a machine is allowed near a human being. europe has an unusually strong claim on that list. the eu spent decades building certification culture for aviation, medical devices, industrial machinery, and automotive safety, and everyone treated it as a tax on innovation. in a world where the frontier capability is an autonomous system operating in a factory or a hospital, the assurance stack stops being a tax and starts being the product.

that is the version of the argument i would defend. it is conditional on world models mattering, conditional on physical deployment being where the value lands, and conditional on europe actually converting a regulatory asset into an engineering one, which is a thing europe has historically been bad at.

on the concrete facts, ami labs is based in paris with teams in new york, montreal, and singapore, co-founded by lecun with alexandre lebrun as ceo. it raised $1.03 billion at a $3.5 billion pre-money valuation in march 2026. its own description calls it a lab focused on world models with persistent memory, reasoning, planning, controllability, and safety. that is enough capital that the research program is hard to dismiss as an academic side path.

update, july 30: on compute, i want to be careful, because this is where european sovereignty takes usually age badly. the european commission launched a call for up to seven ai gigafactories, with up to €10 billion in eu and national funding intended to unlock at least €20 billion more in private investment. the €30 billion headline is an ambition attached to a tender, not money in the ground. for scale, amazon alone expects about $200 billion of capital expenditure in 2026. so the honest read is that europe is building a floor, not parity, and the gigafactories matter for who gets to do serious research on the continent rather than for who wins a scaling race.

none of this means europe suddenly has an openai equivalent sitting in paris. ami still has to prove its architecture produces systems whose capabilities justify the bet. the interesting part is narrower: europe now has the beginning of an alternative path, where a well-capitalized lab and new public compute could converge with industrial strengths that a different paradigm happens to reward.

where this challenges ai 2027 and ai 2040

both scenarios are ultimately built around a world where frontier capability stays concentrated enough that the decisive relationship is american and chinese.

ai 2027 gets there through rapid automation of ai research, where a superhuman coder contributes to large r&d speedups and the resulting systems move toward vastly superhuman capabilities inside roughly a year.

my disagreement is that an architectural bottleneck around persistent world state and long-horizon planning could interrupt that transition exactly where the scenario needs continuity most. a superhuman coder could be economically transformative while remaining incapable of autonomously running the open-ended research process required to discover and validate an entirely new intelligence architecture. that leaves a much stranger intermediate world. ai is already extremely powerful, recursive takeoff has not materialized, and several regions have time to pursue genuinely different technical approaches.

ai 2040 is interesting here because it uses a later default timeline and because plan a explicitly describes itself as a recommendation rather than a prediction. it assumes fully automated ai research would otherwise arrive around 2030, then imagines the us and china agreeing in 2029 to avoid that race, scaling within the human range, pausing at top-human-expert level in 2035, and reaching superintelligence in 2040 with far more safety work done. to be fair to it, plan a wants the eventual equilibrium to be much more multipolar, with dozens of companies globally catching up while research becomes public.

i like a lot of the motivation. secret racing and extreme concentration of frontier power create terrible incentives for everyone involved, and saying so concretely is more useful than saying it abstractly.

the part i find less satisfying is that the path into that equilibrium depends on two governments recognizing the danger, negotiating an extraordinary agreement, verifying compliance, and sustaining it while the strategic value of advanced ai keeps rising.

so the question i keep coming back to is this: what if the technological structure of the race itself made that bilateral winner-take-all dynamic less dominant before a treaty became the only remaining escape route?

a genuine third source of frontier capability would change the negotiating table. the architecture could create more room for interpretability and control to advance alongside planning and autonomy. multipolarity could also make everything worse, obviously, if it just means more independent superintelligence programs with less oversight each. the safety case depends entirely on what kind of specialization and interdependence emerges.

the optimistic version is a world where no single actor controls every critical layer, advanced systems depend on technologies developed across multiple regions, and safety properties become commercially necessary for deploying intelligence inside complicated physical environments. that kind of interdependence makes cooperation easier, because access stays valuable and unilateral acceleration gets harder when the full stack cannot be reproduced inside one national ecosystem.

it also changes what people mean by accelerating ai safety. the goal stops requiring us to suppress capability progress across the board. we could aggressively accelerate world modeling, interpretability, control, verification, uncertainty estimation, robust planning, and the infrastructure needed to evaluate increasingly autonomous systems before they are trusted with irreversible decisions.

if those technologies improve capability at the same time, safety research stops looking like an external tax on the race. that is a more stable incentive structure than hoping every major company and government independently chooses restraint while suspecting somebody else is approaching a decisive advantage.

maybe the optimistic scenario is a different race

i have no idea whether this world happens, and there are several obvious ways the thesis collapses before any of the geopolitical consequences become relevant.

autoregressive agents might solve long-horizon planning through better memory, recurrent state, rl, search, and scaffolding, without requiring anything close to lecun's preferred architecture. world models might become extremely capable while remaining completely inscrutable, which would give advanced systems better strategic understanding with no corresponding increase in human control. ami could fail technically. europe could fail to retain the resulting companies and talent, which is the historical base rate. american and chinese labs could absorb every useful world model idea before europe gains any leverage from originating it.

still, the possibility sticks with me, because it is a plausible safety story that does not depend on technological stagnation, universal altruism, or one perfectly timed international agreement.

we could end up with much more capable ai, a more geographically distributed frontier, and architectures whose economic value depends on reliable state estimation, explicit planning, controllability, and interpretable internal structure.

maybe the best version of the ai race is the one where everyone eventually discovers that understanding and controlling intelligent systems is part of what makes those systems powerful enough to win.

---

tldr

• ai 2027 feels plausible because its story follows current trends, but the takeoff path still depends on assumptions about agent horizons, automated research, recursive improvement, and architectural continuity. ai 2040 uses fully automated ai r&d in 2030 as the default for plan a, while explicitly framing that date as one scenario in a wider range of uncertainty.
• world models could break that continuity by making persistent state, action-conditioned prediction, and hierarchical planning into a separate bottleneck that autoregressive agents may never smoothly cross.
• the strongest objection is that research state is symbolic and can be written down, so scaffolding solves it. my reply is that scaffolding gives you storage and not calibration, and the causal structure an autonomous researcher needs has to be learned rather than logged.
• the early evidence is real but narrow. v-jepa 2 plans robot actions through latent predictions, with human-specified subgoals for longer horizons, and hierarchical world models are starting to generate those subgoals themselves.
• world models seem friendlier to interpretability because they create explicit objects to inspect, with the caveat that labelled interfaces are a hypothesis about the computation rather than a guarantee. verifying them is itself unfinished interp work.
• the europe angle is a conditional, not a prediction. world models do not automatically favor europe, and if the paradigm rewards robotics volume then china is the obvious beneficiary. what europe plausibly has is the assurance stack: certification culture, safety-critical engineering, and industrial depth, which stop being a tax once autonomous systems have to be trusted around people.
• if that conditional holds, the ai race becomes less concentrated, safety becomes a source of capability, and the bilateral dynamic assumed by ai 2027 and ai 2040 becomes less inevitable. that is food for thought rather than a forecast.