How can we achieve human-like driving?

Humans never start from zero

You say humans don’t start at zero performance. What do you mean by that?

Humans are rarely completely useless at a task, even the first time we try it. Take driving as an example. When you start driving at 18, you’re probably not a good driver yet, but you already have a rough idea of how driving works. You’ve spent years watching other people drive. You understand that turning the steering wheel changes the car’s direction, that pressing the pedals affects speed, and that driving off a cliff is a genuinely bad idea. You don’t need to do it ten times to learn this.

This happens because we humans have what you might call an internal simulation engine. We can mentally simulate actions and their consequences, using that to make reasonable plans, even for tasks we’ve never personally done before.

How is this different from how AI systems learn?

Current AI systems do not have that kind of built-in simulation capability, at least not in any general sense. When they face a task they have not seen during training, they often struggle because they cannot reason through the situation the way humans can.

Reinforcement learning is one example of this limitation. On its own, it often learns slowly through trial and error, requiring many attempts before performance improves. It tends to work better when combined with methods that first give the model a stronger understanding of patterns and structure and then use reinforcement learning to fine-tune behavior based on feedback.

It’s like studying first and then practicing with guidance. That combination makes learning faster and more effective than relying on trial and error alone. But these systems still need many examples to perform well. They do not have the same intuitive starting point humans do.

The brain is a powerful learning engine

You seem to suggest that the human brain has some notion of how it should modify its connections rather than operating by trial and error? Can you explain that?

If we speculate about what drives this difference, it helps to start with how humans build an internal model of the world. We are probably not born with a fully formed simulation engine. Instead, we seem to develop it gradually by observing the world around us. Long before we can act competently, as I said earlier, we watch, listen, and absorb patterns: how objects move, how people behave, and how actions lead to consequences.

In that sense, humans are constantly building a world model from experience, often before we can act effectively ourselves. A baby learns a great deal simply by looking around, noticing regularities, and forming expectations about what will happen next. That is a very different starting point from that of a system that must discover everything by acting and being rewarded.

So the key idea is not that the brain behaves exactly like a deep learning model, but that it may rely on a richer learning signal than trial and error alone. That may be part of what allows humans to learn so quickly and generalize so broadly.

That is one reason it seems unlikely that the brain learns primarily through pure reinforcement learning. Reinforcement learning typically depends on repeated interaction and extensive experience to discover what works. Humans, by contrast, appear to build useful internal models from relatively little data, often before much direct practice.

Do you think this simulation capability is something we are born with, or something we gradually develop through experience?

It seems likely that we are born with a powerful learning engine, while the capacity for simulation develops gradually through experience. The amount of information encoded in our DNA seems too limited to specify a rich internal model of the world in detail from the start.

This suggests that the brain may have a more structured way of learning: mechanisms that gradually adjust neural connections to improve its predictions and simulations of the world. In that sense, the brain may have something in common with deep learning systems that update themselves as they learn, even if biology almost certainly works very differently in practice.

At the same time, reward clearly matters too. Dopamine signaling is a classic example of how outcomes reinforce behavior. So the brain may not work like pure deep learning or pure reinforcement learning, but rather as a combination: a system that builds a world model efficiently while also using reward to shape behavior over time.

Even if we could build an AI system with a similar simulation engine – something that could reason through never-before-seen situations – that still would not be enough to fully match human performance. Humans also keep updating their internal models as the world changes, which is another major difference from most AI systems today

The problem of adaptation: Why AI “lives in the past.”

What else do humans do that current AI systems can’t?

Beyond simulating the world, humans are incredibly good at adapting to new environments. Put a human in an unfamiliar context, and they tend to get better and better at handling it over time. AI systems, by contrast, are more like the main character in the movie Memento. In the film, he’s trying to solve his wife’s murder, but every morning he wakes up having forgotten everything he did the day before. He tattoos clues on his body and leaves photos all around to compensate.

AI models work similarly. They’re trained on massive datasets. Once training is complete, we deploy them, and from then on, they stop learning. This applies not just to driving models, but also to large language models like ChatGPT. While you’re talking to such a model, it has no real notion of past conversations with you, or with anyone else around the world. It exists only in the context of the current interaction.

It’s as if you woke up on January 1^st, remembering everything up to that date, but were unable to form any new memories. Every conversation with a colleague would be forgotten immediately afterward. You wouldn’t remember what you had for breakfast. You would be permanently stuck in the past.

What’s the consequence of this for deployed AI systems?

Because these models can’t adapt once deployed, they get worse over time as their environment changes. They are trained to handle the world as it looks during training. But the real world is not static. As conditions shift, the models don’t track those changes. Their performance degrades because they’re effectively frozen snapshots of an earlier reality.

Humans, on the other hand, continuously maintain and update their internal models. We seamlessly switch between learning and using our predictions. For AI systems, training and deployment are distinct, separate phases.

Building products within today’s limits

Given these limitations, what does this mean in practice for companies like Zenseact?

We need to embrace the limitations of current AI systems because we’re building real products that must work in the real world. Tesla summed up the challenge well when they said it’s all about the “99.999%”. If you’re dealing with models that can’t handle situations they’ve never seen before, you have to try to show them everything in advance.

That means collecting vast amounts of data, training models on a wide range of driving scenarios, and using a fleet of customer vehicles to capture rare edge cases. The goal is to expose the model to as many situations as possible so that, at deployment time, very little feels truly new.

Data alone won’t deliver full autonomy

But is it enough to just keep scaling data and compute?

No. As you deploy a driving system across more and more scenarios and environments, it encounters increasingly challengingand complex situations. At some point, simply showing examples of everything stops working.

You cannot keep scaling data exponentially forever. At some point, we need models that can do more than recognize patterns quickly. They also need to reason through unfamiliar situations.

Humans can switch between two ways of solving problems. In familiar situations, we act intuitively and almost automatically, drawing on patterns we have seen before. But when we face something new, we can step through possible futures in our internal simulator and reason our way toward a solution.

That is similar to what is often called System 1 and System 2 thinking: fast, instinctive responses on one hand, and slower, more deliberate reasoning on the other. We need both – otherwise, we are just building faster pattern matchers, not systems that can pause, reflect, and work through unfamiliar situations. Current AI models do not really have this ability today, at least not in any general sense.

How does this affect the competitive landscape in AI and autonomous driving?

This could actually be considered good news. Many competitors are currently focused on scaling up: more data, more machines, more compute. With essentially infinite resources, they can keep pushing that path for a while. But they’ll eventually hit the same wall: the need for exponentially more data to cover all possible scenarios.

This suggests that the future of AI won’t just belong to the players with the largest datasets or most GPUs. It will also belong to those who find new ways forward. To those who can find ways to make AI more human-like in how it learns, reasons, and adapts.

So, where does this leave us regarding AI’s future?

There is still plenty of room for innovation in AI. We are far from the limits of what is possible. The key will be moving beyond static models trained once on massive datasets toward systems that can learn more efficiently, build richer internal models of the world, adapt as environments change, and combine fast pattern recognition with slower, more deliberate reasoning.

The future, in other words, will not belong only to those who can gather the most data. It will also belong to those who can make their AI think and learn more like us.