Even a Perfect World Model Can't Beat Carlsen: The Gap Between AI Prediction and Decision-Making

Why a Perfect World Model Still Isn't Enough

AI researcher Alexander Mattick has put forward a thought-provoking argument: even if you had a perfect world model of chess, you still couldn't solve it — and you still couldn't beat Magnus Carlsen. Behind this claim lies a distinction that's far too often underestimated: being able to predict the future and being able to infer the correct action are two fundamentally different things.

World models have become one of the hottest directions in AI research. The core idea is to train a model to predict how an environment evolves given a sequence of actions. But Mattick argues that having a model that accurately predicts outcomes is not the same as having the ability to make optimal decisions in complex situations. The rules of chess are fully known and deterministic — yet actually winning against an opponent requires high-quality inference across a vast search space, not merely the ability to simulate board states.

This distinction matters because it cuts to the heart of current AI capability boundaries: prediction is about "what will happen," while decision-making is about "what should I do." The gap between the two is far wider than it appears on the surface.

The concept of a world model traces back to the 2018 paper of the same name by David Ha and Jürgen Schmidhuber. The core idea is for an agent to maintain an internal compressed representation of the external environment, using it to "imagine" the consequences of different actions and plan behavior without interacting with the real world. Yann LeCun's heavily promoted Joint Embedding Predictive Architecture (JEPA) is an extension of this same line of thinking. World models have attracted enormous enthusiasm partly because they seem to offer a path toward "understanding" — if a model can accurately predict how the physical world evolves, does that mean it has some deeper grasp of the world? Mattick's argument is precisely a corrective to this intuition: prediction accuracy and policy quality are theoretically independent objectives. The former is a supervised learning problem; the latter is fundamentally a search and optimization problem. Even under the assumption of a perfect model, efficiently finding the optimal path through an exponentially exploding decision tree remains a separate and extraordinarily difficult computational challenge.

Why Current World Model Tasks "Look" So Impressive

Mattick argues that the capabilities world models currently demonstrate are overestimated, largely because the tasks they handle today aren't actually that difficult from a control perspective.

In other words, the decision complexity of many demo scenarios is quite limited. Models perform impressively in these relatively structured, predictable environments, which can create the illusion that world models are already highly capable. But once tasks involve real-world problems with high complexity and contact-rich control, the gaps quickly become apparent.

This is a reminder to be more careful when evaluating AI progress: a system that can complete flashy tasks in controlled environments hasn't necessarily mastered generalizable decision-making and control capabilities. Prediction accuracy and action capability cannot simply be equated.

Why Robots Can Do Backflips But Can't Get Milk from the Fridge

Mattick uses a particularly compelling example to illustrate this point. You'll see humanoid robots from China, the US, Germany, and elsewhere constantly demonstrating backflips and sprinting — but you almost never see them complete a simple task like: just stand up, walk into the next room, open the fridge, look toward the back, grab a carton of milk, and hand it to me.

The reason isn't that robots "don't want" to do everyday tasks — it's that tasks like these are actually far harder. Backflips and sprinting look spectacular, but they are fundamentally structured, pre-plannable motion sequences. Fetching milk, by contrast, involves a contact-rich environment — fine-grained finger-to-object contact, force control, and manipulation of soft or irregularly shaped objects.

Mattick specifically points out that we still lack truly reliable high-resolution tactile sensing. Contact-rich manipulation is both difficult to model and difficult to collect data for — you can't gather large volumes of high-quality tactile interaction data the way you can simply record video. This is the core bottleneck behind why everyday manipulation tasks remain so stubbornly unsolved.

Contact-rich manipulation is a foundational challenge in robotics because contact itself introduces physical discontinuities. When a finger grips an object, the system's dynamics equations change abruptly at the contact point — force direction, friction coefficients, and object deformation all shift dramatically within milliseconds, making traditional differential-equation-based control methods poorly suited to handling these "hybrid systems." By contrast, a backflip involves ballistic motion with no contact or a single contact point, where the physics equations remain relatively continuous throughout, making the problem far more tractable via reinforcement learning or trajectory optimization. The immaturity of tactile sensors compounds these difficulties: current commercial tactile sensors (such as vision-based tactile solutions like GelSight and DIGIT) have made progress, but still lag far behind human skin in durability, spatial resolution, and real-time signal processing. The data scarcity problem is equally serious — the internet contains vast quantities of video for visual models to learn from, but there is no comparable large-scale public dataset of high-quality tactile interaction data, which fundamentally limits the data-driven machine learning approaches that dominate modern AI in this domain.

Implications for AI Development

This line of argument reveals an important industry reality: there is an enormous technological gap between flashy robot demonstrations and genuinely useful everyday manipulation. A robot that can do backflips isn't necessarily able to grab something from your fridge, just as a system with a perfect chess model isn't necessarily able to beat a top grandmaster.

The core tensions can be summarized as follows: prediction capability is not decision-making capability; success in controlled environments does not imply real-world ability; and modeling contact-rich tasks while collecting the necessary data remains one of the hardest unsolved problems standing in the way of embodied intelligence.

For those following developments in AI and robotics, this offers a valuable "cut through the hype" lens. Rather than being drawn in by impressive-looking motions, it's worth focusing on how a system actually performs in complex, unstructured, contact-rich scenarios. Real breakthroughs, more often than not, hide inside tasks that look entirely unremarkable.