The Age of Imagineering will give us machines capable of creating almost anything. Our next challenge is proving which of those creations deserve to become real.
We are entering an era in which the cost of creating an idea is collapsing. Software that once required a team can increasingly be generated from a conversation. A molecule can be proposed by a model. A factory can be simulated before it is built. A clinical trial can be modeled before a patient is enrolled. An autonomous agent can design a process, write the code, execute the workflow, analyze the results and propose the next experiment.
This is part of what I have been calling the Age of Imagineering. For most of human history, imagination was abundant and execution was scarce. We could imagine far more than we could build. Artificial intelligence is beginning to reverse that equation. Increasingly, the distance between I can imagine this and I can create this is collapsing.
But that creates another problem. A much bigger one.
How do we know whether what the machine created is right?
That may become one of the defining questions of the next technological age. Because the faster our ability to create grows, the more important our ability to verify becomes.
Creation Is Becoming Cheap. Truth Is Not.
Generative AI has already demonstrated something extraordinary. Machines can produce plausible answers at enormous scale.
Code. Research. Images. Designs. Strategies. Hypotheses. Models. Simulations.
The problem is that plausibility and truth are not the same thing. That distinction is manageable when the output is a paragraph. It becomes much more consequential when the output controls something in the physical world.
A hallucinated sentence is inconvenient. A hallucinated financial model can destroy capital. A hallucinated engineering calculation can collapse a structure. A hallucinated medical conclusion can harm a patient.
As artificial intelligence moves from bits to atoms, verification stops being a quality-control feature. It becomes infrastructure.
The Age of Imagineering therefore cannot simply be built on increasingly intelligent models. It will also require increasingly sophisticated systems for determining when those models should be trusted.
We Need to Verify Two Different Things
There are actually two verification problems emerging. The first is capability verification. Can this model reliably perform the task we are asking it to perform? The second is output verification. Is this particular answer, prediction, design or action correct? Those are not the same question.
A model may be excellent at predicting drug metabolism and poor at predicting immune response. It may write secure Python extremely well while performing poorly on another language. It may accurately model one population while systematically underperforming on another.
The question cannot simply be:
Is this a good AI model?
We need a much more precise question:
For which problems, under which conditions, and within what limits has this system demonstrated that its outputs correspond with reality?
That is a fundamentally different way of evaluating artificial intelligence. And I believe we are going to need an entirely new verification infrastructure to support it.
Benchmarks Are Only the Beginning
Today we often evaluate AI systems using benchmarks. We give a model a collection of questions or tasks, measure its performance, and compare it against other models. That works reasonably well when we are evaluating mathematics, coding, reasoning or language. But the physical world is much more complicated.
Medicine is a good example. Imagine giving an AI everything humanity knew about a particular medicine before a major clinical trial occurred.
Its chemistry. Mechanism of action. Preclinical evidence. Dosage. Phase I data. Patient characteristics. Biomarkers. Known side effects.
But we deliberately hide what happened next. Then we ask the model:
What will happen when this treatment is given to 10,000 people?
The model makes a prediction. We lock the answer. Then we reveal history. We already know what happened.
Did the drug work? What adverse events occurred? Which populations responded? Which did not? What happened at different doses? Were there unexpected interactions?
Now compare the prediction against reality. And repeat the experiment. Not ten times. Thousands of times.
Eventually, something much more powerful than a conventional benchmark emerges. We have created a historical replay system for artificial intelligence. History itself becomes the test suite.
The Most Valuable Dataset May Be the Past
This is where I think one of the largest opportunities of the Age of Imagineering may exist. We already possess enormous amounts of ground truth. Human civilization has been running experiments for centuries.
Clinical trials. Engineering projects. Manufacturing systems. Financial crises. Infrastructure failures. Cybersecurity incidents. Weather events. Scientific experiments. Product launches. Supply-chain disruptions. Industrial accidents. Disease outbreaks.
We know the inputs. And because they already happened, we know the outcomes. That makes historical data uniquely valuable.
We can freeze the world at a moment in time, give an AI only the information that existed then, ask it to predict what happens next, and compare the prediction with reality. That is extraordinarily powerful. It gives us something synthetic benchmarks cannot provide.
Ground truth.
And I increasingly believe that carefully curated historical golden datasets may become one of the most important moats of the next technological era.
Models will improve rapidly. Today's leading model will eventually be replaced. Compute will become cheaper. Algorithms will diffuse. Capabilities that are proprietary today will eventually become commonplace.
But a high-quality corpus of verified historical reality, complete with provenance, context, timestamps, uncertainty, outcomes and repeatable evaluation methodology, becomes more valuable as more models appear.
The model may not be the moat. The truth against which every model must be measured might be.
Imagine Reality Becoming a Benchmark
The same idea extends far beyond medicine. Take structural engineering. Give an AI the design specifications and environmental conditions for thousands of historical bridges before they were constructed. Ask it to predict fatigue, maintenance requirements, structural failure and lifespan. Then compare its predictions with what happened.
Or cybersecurity. Give a model historical versions of software before vulnerabilities were discovered. Can it find the vulnerabilities before being shown the exploit?
Or finance. Give a model everything available before historical credit events, liquidity crises or market failures. Could it correctly estimate the risks?
Or manufacturing. Give the system telemetry from industrial equipment before a failure occurred. Can it identify the failure before history reveals it?
The principle is the same. Freeze the information. Predict the future. Reveal reality. Measure the gap.
Suddenly human history becomes an enormous verification environment.
Verification Will Need Its Own Infrastructure
Every technological age eventually develops infrastructure that did not exist when the age began. The industrial revolution required far more than steam engines. It needed electrical standards, building codes, standardized measurements, safety systems, testing laboratories and inspection regimes.
The information age required more than computers. It needed TCP/IP, encryption, digital identity, certificate authorities, cybersecurity standards, data governance and software-testing frameworks.
The Age of Imagineering will require its own infrastructure. We will need systems for:
- model evaluation
- historical replay
- provenance
- uncertainty measurement
- reproducibility
- simulation validation
- continuous monitoring
- population and demographic testing
- model drift detection
- output verification
- chain of custody
- machine-readable evidence
- auditability
And perhaps most importantly, we will need a way to express where a model is trustworthy and where it is not. Rather than giving an AI system one universal intelligence score, imagine something closer to a capability map:
- Drug metabolism: 97%
- Cardiac toxicity: 94%
- Immune response: 89%
- Rare genetic disorders: 71%
- Pregnancy outcomes: insufficient verified evidence
That would tell us something much more useful than whether one model beat another on a leaderboard. It tells us how much demonstrated reality lies behind a particular capability.
Verification Becomes Even More Important When AI Can Act
This problem grows exponentially as AI becomes agentic. Today an AI often recommends an action to a human. Tomorrow the AI may take the action itself.
It may write software and deploy it. Design a component and send it to a manufacturing system. Choose a chemical formulation and instruct robotic equipment to synthesize it. Adjust an industrial process. Control a warehouse. Coordinate a fleet of machines. Or participate in designing a medical intervention.
At that point, we are no longer simply asking whether a model can produce a convincing answer. We are asking whether society has built enough confidence in the model to permit it to change the physical world. That requires a verification layer between imagination and execution.
The emerging stack may eventually look something like this:
Imagination
Artificial intelligence
Simulation
Verification
Physical execution
Real-world evidence
Feedback
Better models
The verification layer is what keeps the system grounded in reality. Without it, the entire stack becomes fragile.
Medicine Shows Us What This Could Become
Consider how medicines are developed today. A candidate moves through preclinical research, Phase I, Phase II, Phase III and regulatory review. That process exists for good reason. Human biology is extraordinarily complicated, and historically the only reliable way to understand what happens in humans has ultimately been to test something in humans.
But imagine biological models becoming dramatically more capable. Imagine those models being continuously evaluated against decades of historical clinical data. A new treatment might first be evaluated across millions of virtual patient profiles.
Different ages. Different genetics. Different medications. Different medical histories. Different immune systems. Different diseases. Different combinations of risk factors.
The model would not replace human evidence. Instead, it could identify where uncertainty remains. Human experimentation could increasingly focus on the unknown rather than rediscovering everything we already know.
That could make trials smaller. More targeted. Safer. Faster. And potentially much more informative.
But none of that works unless regulators, scientists, physicians and the public can trust the models being used. Which brings us back to the same issue.
The limiting factor may not ultimately be our ability to simulate biology. It may be our ability to prove that the simulation deserves to influence reality.
Trust May Become One of the World's Most Valuable Commodities
The Age of Imagineering will produce an extraordinary abundance of possibilities. Millions of designs. Millions of hypotheses. Millions of simulated futures. Millions of potential solutions.
Creation itself may become nearly free. But every one of those possibilities introduces a question:
Which one should we believe?
That may fundamentally change where value sits in the technology stack. The most important companies of the next era may not only be the companies building increasingly powerful models. Some may own the systems that determine whether those models can be trusted.
The datasets. The benchmarks. The historical evidence. The verification methods. The provenance. The credibility frameworks. The interfaces between computational prediction and physical reality.
Because ultimately, intelligence without verification is only possibility. And possibility alone cannot build bridges, approve medicines, operate factories or make decisions that affect human lives.
The same shift in value changes what customers are paying for when AI speeds up professional work.
The Age of Imagineering will give us machines capable of imagining millions of possible futures. Our next great infrastructure challenge is building the systems that can tell us which of those futures are real.
AI engineer, builder, and author of The Age of Imagineering.