The Five Senses of Artificial Intelligence

An AI agent can write a program, execute it, discover that something went wrong, and try again. But what happens when the program works perfectly and the person using it still finds it useless? There’s an important difference between completing a task and understanding whether the outcome was valuable. As AI becomes increasingly capable of acting in the world, I think this distinction will become more important, and it may change the role humans play in the development of artificial intelligence. For most of the history of computing, software has been something we build, package, and sell. Creating it required specialized knowledge, time, and significant resources. Entire companies were built around applications, and the software itself became the product. But I think we’re approaching a fundamental change in how we create and value software. Over the past several decades, humanity has built an extraordinary collection of open-source libraries, frameworks, operating systems, and tools. Almost every application we use today relies on work that thousands of people have contributed to over time. We’ve collectively created an enormous foundation of reusable components, and AI is becoming increasingly capable of assembling those components for us. We’re not yet at a point where AI can reliably build and maintain any complex software system without human involvement, but the direction is becoming clearer. If this continues, software could gradually become a commodity, not because it will stop being valuable, but because the ability to produce it will become increasingly accessible. And if building software is no longer the hardest part, I think we need to reconsider where the real value will come from. The World Beyond the Model AI models have learned an extraordinary amount from the information humanity has already produced. They can generate code, reason about complex problems, and propose solutions to situations they have never encountered in exactly the same form. But knowing about the world and learning from direct interaction with it are two different things, and this distinction becomes increasingly important as AI moves beyond generating answers and begins taking actions. An AI agent can interact with tools, observe the results of its actions, and adjust its approach based on what happens. In that sense, AI is beginning to develop ways of experiencing its environment, even if those experiences are very different from our own. The important part is not simply that it can act, but that the difference between what is expected and what actually happened can become an opportunity to learn. Humans develop much of their understanding through a similar process. Our five senses allow us to perceive the world, but intelligence doesn’t come from simply collecting information. We form expectations, take actions, observe the consequences, and gradually change what we believe. A child learning to walk doesn’t become better by reading instructions about balance. They learn through repeated interaction with the physical world, discovering what works and what doesn’t. Something similar could happen as AI systems become more autonomous. But there is an important difference between observing that a task succeeded and understanding whether that task was worth doing in the first place. A model might generate an application that works according to its specifications, yet the people using it discover that it is confusing. An automated system might optimize a process while overlooking consequences that matter to the people affected by it. These are situations where human experience becomes essential, because we bring more than observations. We bring intentions, judgment, preferences, and an understanding of the world in which these systems operate. In a sense, humans could become the five senses of an evolving AI ecosystem. Not because machines cannot observe the world, and not as passive sensors feeding information into them, but as active participants connecting artificial intelligence to the reality it is supposed to serve. Our interactions could help these systems understand not only what is technically possible, but what is useful, meaningful, and beneficial. Who Owns What We Learn Together? This brings me to something I find difficult to reconcile with how the AI industry currently operates. Today, much of our interaction with AI happens through platforms controlled by private companies. We ask questions, correct mistakes, test ideas, develop workflows, and increasingly allow agents to act on our behalf. Not every interaction becomes training data, and companies handle this information differently, but these interactions can generate valuable experience about what works, what fails, and what needs to improve. If millions of people are contributing to that process, should the resulting experience necessarily become an asset controlled by the platform through which it was generated? I’m not arguing that companies shouldn’t own the models they develop or benefit from the enormous investments required to build them. There is real value in that work, and commercial incentives have played an important role in technological progress. But I think we need to distinguish between the technology that enables an interaction and the experience generated through that interaction. If I use an AI agent to solve a problem, discover that its first approach fails, and help it find a better solution, something valuable has emerged from our collaboration that goes beyond the model itself. The same is true when an autonomous agent discovers something without direct human intervention. If it spends hours attempting to solve a problem, testing different approaches, and eventually finding a solution, the experience generated through that process could be valuable to other systems. Should that learning disappear when the agent restarts or when we switch to another model? Should another agent have to repeat the same mistakes simply because the experience was never preserved in a reusable form? As AI becomes more capable of acting independently, I think these questions will become increasingly important. The value of an intelligent system may depend not only on what its model already knows, but also on what it has learned through its interactions with the world. Preserving that learning doesn’t necessarily mean sharing private information. It could mean sharing useful structures, evidence, and the circumstances under which something worked. We Already Know How to Build Together There’s something interesting about the history of open-source software. Much of the technology we use today exists because generations of developers built on each other’s work. Someone created a library, someone else improved it, another person discovered a bug, and eventually thousands of people benefited from something none of them could have built alone. What made this possible wasn’t just access to code, but the ability to share discoveries, examine mistakes, challenge decisions, and continuously improve what already existed. In many ways, open source became a form of collective learning. But AI could change this dynamic in an unexpected way. Imagine a future where anyone can ask an AI agent to build an application. Instead of developers collaborating on shared projects, millions of agents could independently generate different solutions to similar problems. One agent encounters a failure and finds a solution, another encounters the same problem using a different model and starts from scratch, and a third discovers a better approach that remains isolated in its own environment. We might end up producing more software than ever before while repeatedly rediscovering the same lessons. This is why I think the next important opportunity for open source may not simply be another library or framework, but a way to preserve and connect what we learn from interacting with the world. Software allows us to reuse an implementation, but experience can tell us why it worked, what assumptions it depended on, and when it might fail. That understanding carries context, uncertainty, and history that we need to preserve if we want it to be useful elsewhere. An Intelligence That Learns With Us Imagine an AI system that spends its day building and improving software. It receives a goal, tries an approach, encounters an unexpected failure, and eventually discovers a solution. The important thing isn’t just that it completed the task. It’s that the entire process generated experience. The failed attempts, the decisions that led to them, and the evidence that eventually revealed a better approach could all become useful when the system encounters a similar problem again. Now imagine that this experience doesn’t disappear when the system restarts or remains useful only to the model that generated it. Other agents could build on what was discovered, while still being able to question whether the original lesson applies to their own circumstances. Over time, this could create a form of collective learning where progress is not limited to improving individual models, but also comes from accumulating and sharing what different systems discover through experience. Of course, learning from experience is not as simple as remembering what happened. A successful result doesn’t always explain why something worked, and an unsuccessful attempt can be just as valuable if we understand what went wrong. Experience can also be misleading, outdated, or dependent on circumstances that no longer exist. Any system designed to share learning would need ways to preserve this context, validate conclusions, and allow others to challenge what has been learned. There is also a broader question about what should guide the evolution of these systems. Today, the companies developing the most capable AI models are naturally influenced by commercial incentives. They invest enormous resources, compete for customers, and need sustainable businesses. There’s nothing inherently wrong with that, but the problems that matter to humanity don’t always align with the problems that generate the most revenue. An open learning ecosystem could allow useful experiences to contribute to a broader understanding, even when no individual company has an immediate financial reason to prioritize them. This doesn’t mean every experience should be public or that openness automatically guarantees neutrality. People and organizations must retain control over their private information, and shared learning will inevitably contain biases, disagreements, and mistakes. But openness gives us the opportunity to inspect what has been learned, question conclusions, test alternatives, and correct mistakes rather than depending entirely on decisions made behind closed doors. What Will Be Valuable Tomorrow? For a long time, the ability to build software was itself a scarce skill. And while engineering will continue to matter, I think AI will increasingly reduce the effort required to turn ideas into working systems. What won’t become so easy is discovering what those systems should do, how they behave in the real world, and what we can learn when our expectations turn out to be wrong. That requires experience, and experience is something humanity generates continuously through our interactions with the world. Perhaps the future of AI isn’t simply about building models that know more or agents that can create software faster. Perhaps it’s about building an intelligence that develops through its relationship with reality, where humans are not merely users but active participants in the process of learning. As AI becomes increasingly capable of acting independently, our role may shift from telling machines exactly what to build toward helping shape what they learn, what they value, and how they understand the consequences of their actions. If that experience becomes one of the most important resources for the next stage of AI, I think we should be careful about allowing it to become permanently locked inside individual companies. The learning generated through our interactions should remain under our control, and when we choose to share what we’ve learned, that knowledge should have the opportunity to benefit others. If humanity is going to help AI experience the world, humanity should have a say in how that intelligence evolves. This is one of the ideas behind Project 507, an open-source effort to help AI systems learn from experience and build on what they discover. If you’re interested in helping build it join the community through the Community tab on our website.