The future of AI planning is pre-execution consequence simulation

Why next-generation AI systems need world models that emulate possible outcomes before committing to action.

I’ve been working and experimenting with AI for a long time, including building a company in the space before the current wave made it mainstream. I am extremely grateful for the opportunities LLMs have presented, yet this has raised the bar and opened my mind to what else is possible.

I’ve been on the Ai wave early. Since 2016 being a huge believer in building decision-making systems that take in data, are able to distill it, and make decisions about patterns they recognize. That’s when I started experimenting with Machine Learning and Computer Vision.

In Apr 15, 2018, I wrote an article on Medium — The future of your business depends on Ai. Evolve or die! (Very dramatic, I know…)

The same year, Me and my 2 co-founders we building a SaaS for the fashion model talent management industry. The concept was simple: A lead generation tool for talent scouts to find talent on social media. In the fashion industry, there is a specific criterion for the type of faces that have unique talent to be scouted into an fashion modeling agency.

Face geometry can be translated into numbers. So we used machine learning algorithms to train an AI model on “suitable for fashion modelling” looking faces and ”not suitable for fashion modelling.”

Astonishingly, we were able to achieve 89% accuracy rate on predictions on a cluster of images grouped by face matching of the person. That alone helped us secure a $312'000 ARR contract with the world's largest modeling agencies and put quite a few models on the cover of the magazines.

But at what cost?

Analyzing 200'000 profiles per day with 89% accuracy rate, you were left with at least 22'000 profiles of false positives that you can’t show to the clients. Top agency scouts don’t have time to review thousands of profiles of false positives. So we had to build an internal team of 7 people of “Account Managers” who were qualified to review the results picked by the AI and remove the false positives.

And that wasn’t even the greatest operating expense. The hardware, especially the GPUs, ate up half of the monthly revenue generated by the company. So we simply ran into a hardware challenge, and being totally honest with you, solving a problem for such a niche industry limited us in where we could scale, so there was not much room for growth, nor expertise for a pivot or proper expansion.

But today, the landscape is completely different. Regardless of the industry you are in, you now have access to general-purpose intelligence at a fraction of the cost of what specialized AI systems used to require.

That does not mean an LLM could simply replace what we built for the modeling agency scouting tasks. Our scouting system depended on computer vision, face matching, domain-specific training data, and operational review loops. You would not be able to open a chat ChatGPT interface and reproduce that product out of the box. It has no perception or sense of taste that suits the fashion industry.

But LLMs changed something else. They made reasoning, synthesis, decision support, and productivity broadly accessible. The bottleneck is no longer only whether a model can recognize a pattern. The new question is whether a model can understand a situation, reason through uncertainty, and help decide what should happen next.

With this, you run into a completely different set of problems. These problems define the boundaries we face today with the greatest AI models on the market. The more I build with these models, the more I keep getting frustrated by how seemingly smart they are, yet how quickly I run into a wall.

You can get it to one-shot a lot of things. That part is real. But the true test of intelligence is different. Given the right information, can the model actually understand it? And if it doesn’t have the information, can it recognize that, go find it, and then recognize that what it found is the right answer?

That last chain is where things consistently break down. The model can search. It can read what comes back. What it usually can’t do is identify its own knowledge gap, retrieve the specific thing that fills it, and verify that the retrieved thing actually fills it.

I want to lay out what I actually believe about how these systems should work. This is where I find myself agreeing with Yann LeCun, even though his position is unfashionable in a market where every company is racing to scale the same architecture.

The current generation of LLMs is, at its core, a next-token predictor. It’s a very good one. Trained on enough text, it produces output that looks like reasoning, and in many narrow cases, it’s genuinely useful. But the underlying operation is statistical pattern completion over language. The model isn’t modeling the world. It’s modeling what people have written about the world, which is a thin and lossy projection of the real thing.

That distinction matters because language is downstream of understanding, not upstream of it. A child understands gravity before they have the word for it. They’ve watched things fall, tried to catch them, and learned that heavy things hit harder. The word “gravity” gets attached later to a model they already have. LLMs work in the opposite direction. LLMs demonstrate functional understanding in language space, but they lack grounded understanding in world space. That’s why they can describe physics fluently and still fail at problems a six-year-old solves by intuition. Autocomplete can mimic thought, but it cannot replace grounded consequence.

What I think real intelligence requires, and what I think AI industry needs to build towards, is a system that learns the world the way a child does: by interacting with it, building an internal model of how it works, and using that model to predict what will happen and plan what to do. Not predicting the next word about the world. Predicting the world itself, in its own representation, and using language as an interface on top of that capability rather than as the substrate underneath it.

This is where I think the current conversation around AI gets too narrow. People keep asking whether the next model will have a larger context window, better tool use, or more training data. Those things matter, but they do not answer the deeper question. The issue is not whether the model can access more information. The issue is whether it has any grounded way to know what that information means, whether it applies, and what consequences follow from it.

This is the core of what LeCun calls world models, and what his new lab AMI is built around. The agent learns a compressed internal representation of how the environment behaves, makes predictions in that representation, and uses those predictions to reason about consequences. Language becomes a way to query and explain the model, not the model itself.

I think this is the right direction, and I think the current LLM-only approach is going to keep hitting walls until we add something underneath it.

This matters because goal achievement is not just about reaching an endpoint. A system can brute-force its way toward a destination and still fail in every meaningful sense. Imagine walking through a forest. One person might sprint blindly toward the exit and technically make it through. But if they ignored the terrain, stepped on a snake, lost blood along the way, and arrived barely alive, did they actually succeed? The destination was reached, but the goal was not achieved in a complete sense. A more intelligent person moves differently. They read the environment, notice patterns, infer danger, preserve their body, adjust their route, and constantly evaluate not only whether they are getting closer to the exit, but what the path is doing to them as they move through it.

This is the missing layer I keep coming back to. Intelligence is not only the ability to pursue an objective. It is the ability to understand the cost of pursuing it. To know whether the path that gets you there destroys the thing you were trying to preserve. Probably the developers will relate to this statement. It has nothing to do with development, you’ve experienced your AI rewrite your project to just fix a bug but then mess up or add redundancy somewhere else. To recognize that every action changes both the world and the actor. A model that cannot represent the state of the environment, the state of itself, and the downstream consequences of its actions is not truly planning. It is optimizing against a shallow definition of success.

This is where the idea of survival instinct becomes useful, but only as an analogy. A machine does not need fear, pain, or a biological desire to stay alive. What it needs is a representation of consequence. When a human walks through a forest, the body is constantly reporting state: fatigue, injury, hunger, danger, orientation, confidence, fear. Those signals are not separate from intelligence. They are part of how intelligence knows whether the path is working. If you get to the other side of the forest poisoned, bleeding, and unable to continue, you did not truly achieve the goal. You reached the location while destroying the system that was supposed to benefit from reaching it.

For an AI system, that consequence layer has to be digital. Before acting in the real environment, it needs an internal place where the action can be played out. Not merely described in language, but simulated as a state transition: if I do this, what changes? What breaks? What becomes harder later? What new risks appear? What resources are consumed? This is closer to a imagination within a brain that emulates a contained environment where possible futures can be tested before they are executed.

Blockchains already give us a simple version of this idea. Before signing a transaction, you can simulate what the transaction would do. If the simulation shows that a smart contract will drain your wallet, you do not execute it. The simulation creates a consequence before the consequence becomes real. That is the missing layer I want in AI systems. Not just a model that can explain a plan, but a system that can mentally execute the plan, inspect the damage, and refuse the path if the goal is achieved by destroying the thing the goal was meant to preserve.

Theoretically you can use version control, you can use virtual environments, and a lot of tools are already available to do something close to what was described. It is very advanced-level configuration. Thus far I haven’t seen any agents’ ability to spawn parallel sub-agents to run containerized experiments to find the best outcome. Out-of-the-box development is very synchronous.

I do not know what the right architecture or algorithm for this looks like, nor do I claim to be smarter than the people who invented the systems we use today. I am a power user and an avid fan of what exists. But the more I use LLM models, and the more I play out thought experiments about our own consciousness, the more I drift toward a different question: how do we take this collective, compressed intelligence we have created and move it closer to something that can actually understand and work with us?

The question is becoming clearer to me: can we build AI systems that do not simply describe possible actions, but, like us, can visualize and play out scenarios inside emulated environments before even making suggestions and taking real action?

And if so, how would those environments need to work? How could they be constrained by the same laws of physics, causality, incentives, and consequences that shape our own existence?

Can a system represent not just the goal, but the state of the world, the state of itself, and the cost of the path it is about to take? Can it recognize when a plan technically reaches the destination while damaging the thing the goal was meant to preserve?

That is what I want to explore next.