The Questions Were Free. Pursuing Them Wasn't.

How AI is changing the economics of curiosity, one experiment at a time.

For most of my forty years writing software, curiosity was free. Pursuing it was considerably more expensive. I could wonder about something, imagine a way to investigate it, and even possess the skills to build the necessary software. But before doing anything, I’d have to decide whether the question was worth the effort. Today, with AI-assisted development, that calculation has changed rather dramatically.

Last week, working with Claude, I built a sophisticated interactive visualization in approximately thirty minutes. A comparable project might previously have consumed 300 hours of development time, although I’ll never know precisely because I have no intention of building it again the hard way. The remarkable thing, however, wasn’t the time I saved. It’s that I would never have spent those 300 hours in the first place.

For most of my career, I’ve carried around a collection of questions, ideas, and passing curiosities that occasionally begged for software to help answer them. Some were professionally relevant, others were intellectually interesting, and a few probably deserved to remain passing thoughts. I generally knew how to build what I imagined, but knowing how and having the time were entirely different propositions. Every interesting idea came with an invisible invoice, and most were quietly abandoned before anyone had to pay it.

Recently, that invoice became considerably smaller. And I discovered that curiosity behaves rather differently when you can afford to indulge in it.

The Question That Started It

Something had been lingering in my mind about the data behind mass shootings. I wondered what the history might look like if I could see incidents unfolding geographically, over time, across countries and decades. I wasn’t looking to write a political argument or confirm a particular position. I wanted to see a pattern that was difficult to grasp through individual news stories and disconnected statistics.

Historically, satisfying that curiosity would have meant collecting data, cleaning records, reconciling locations, building geographic visualizations, creating a timeline, designing filters, and making the whole thing interactive. None of those tasks were beyond my abilities. Collectively, however, they represented enough work to make the project difficult to justify, particularly when there was no client, revenue opportunity, or practical obligation attached to it.

Nevertheless, I opened Claude and started working. About thirty minutes later, I had a functioning Mass Shooting Atlas, with geographic exploration, chronological playback, severity filters, classifications, and supporting data documentation. It was considerably more sophisticated than the sketch I might once have produced in that amount of time.

The visualization has limitations, some of them substantial. Historical coverage is inconsistent, international reporting is uneven, and differences in available records complicate comparisons across countries and periods. Those limitations matter, especially with a subject this serious. Building an instrument to examine a question doesn’t automatically make its readings authoritative, but it does give us something concrete to investigate, improve, and challenge.

I had gone from wondering about something to having an instrument for exploring it in about 30 minutes. That alone was remarkable, but the more interesting realization came afterward. I started wondering what else I could explore.

Apparently, Curiosity Is Contagious

My next question was considerably less disturbing. How quickly had the universe grown during my lifetime? More precisely, how many planets beyond our solar system had humanity discovered since I was born, and what would that expansion of knowledge look like if I could watch it happen?

The result was How Quickly Did the Universe Get Bigger?, an interactive visualization using NASA’s exoplanet data. Enter a birth year and watch confirmed discoveries accumulate over time, examine how planets were detected, and explore their sizes and distances. The underlying dataset contained 6,451 confirmed exoplanets when I built it. That number is a snapshot of scientific knowledge, not a claim that those worlds suddenly appeared when we found them.

There is something delightful about watching humanity’s known planetary inventory expand while realizing that the universe itself was getting along perfectly well without our accounting department. For most of human history, we had no confirmed planets orbiting other stars. Today we can explore thousands of them through a browser, including discoveries made within our own lifetimes. What began as a passing curiosity became a way to experience the history of discovery rather than merely read about it.

Then I turned my attention from the universe to something much closer to home. How much of a human life disappears into commuting? We talk about travel time in minutes, occasionally hours, but rarely in the accumulated currency of our finite existence. I wanted to see what those daily journeys amounted to over a working lifetime.

That became Going Somewhere Else, an interactive tool using US Census commuting data. It lets you explore geographic differences, adjust assumptions about working years and commuting schedules, and watch the accumulated time appear as days taken from a lifetime. A hypothetical 45-minute commute in each direction, five days a week for forty years, consumes 14,400 hours. That’s 600 complete days, or approximately 1.6 years spent traveling to and from work.

Naturally, the calculation makes assumptions. Few people maintain identical commutes throughout their careers, and remote work changes the equation considerably. Nevertheless, watching those days accumulate feels different from reading an annual average. The numbers haven’t necessarily changed, but the experience of understanding them has.

Three questions had produced three working applications. One concerned human violence, another the expanding boundaries of scientific knowledge, and the third the ordinary arrangements through which we spend our lives. They shared no product roadmap, market opportunity, or plausible venture capital pitch. Their common denominator was simply that I wanted to know something, and for once, pursuing that curiosity required remarkably little justification.

The Economics of Curiosity

We’ve spent much of the AI conversation discussing productivity, usually expressed as time saved doing work we already intended to perform. That makes sense because organizations understand labor costs, schedules, budgets, and throughput. If a developer finishes a forty-hour assignment in ten hours, we can construct a respectable spreadsheet explaining the improvement. Someone might even create a dashboard, because no organizational achievement is complete until somebody gives it a traffic light.

But that accounting misses an entire category of value. What happens when the original cost of investigating something was so high that the investigation would never have occurred? There is no existing task to accelerate, no approved budget to reduce, and no baseline project schedule against which to measure success. Traditional productivity calculations struggle to describe work that exists only because the economics of attempting it have changed.

That is why my hypothetical 300-hour saving deserves some skepticism. I didn’t have 300 hours allocated to building an atlas, and nobody was waiting for its delivery. The actual alternative wasn’t a slower implementation. It was probably no implementation whatsoever, followed by the question quietly disappearing beneath the obligations of everyday life.

The questions were always free. What AI has changed is the cost of doing something about them.

This is what I find fascinating about AI-assisted development. When the cost of exploration falls far enough, it changes which questions we consider worth pursuing. We no longer need to predict the value of every experiment before conducting it. Sometimes we can build enough of an answer to discover whether the question was interesting in the first place.

And, occasionally, an experiment that produces nothing useful is still a perfectly reasonable investment. That possibility ought to appeal to anyone who has ever pursued a line of inquiry simply because they couldn’t stop thinking about it.

When Software Becomes a Way of Thinking

Software has traditionally been something we build to accomplish a defined purpose. We establish requirements, design a solution, implement it, test it, and deliver it to someone who needs the resulting capability. Those disciplines remain important, especially when software carries financial, operational, or safety consequences. But they represent only one way software can serve us.

I increasingly see another possibility. Software can become a temporary instrument for thinking, something we create to examine a question and perhaps discard after it has served its purpose. A visualization need not become a commercial product to be valuable. It may justify its existence by changing how its creator understands a problem, revealing a surprising relationship, or producing a better question than the one that prompted it.

There is a useful distinction here between generating an answer and building an environment in which an answer can be explored. I could ask Claude how many exoplanets have been discovered, and it could return a number. I could ask how much time a person spends commuting, and it could calculate the hours. But watching those discoveries accumulate or changing assumptions about a working life creates a different intellectual experience. The user is no longer merely receiving information. The user can investigate it.

In that sense, I haven’t necessarily answered my three original questions. I’ve created three ways to explore them, and each allows for new questions that might not have occurred to me beforehand. That distinction matters because the intellectual value of an investigation isn’t always a definitive conclusion. Sometimes it’s a better understanding of what we don’t yet know.

Of course, the ease of creating these instruments introduces another problem. A beautiful visualization can make questionable data feel authoritative, and an interactive interface can lend false precision to assumptions that deserve scrutiny. The easier it becomes to produce convincing software, the more important it becomes to distinguish convincing presentation from reliable understanding.

AI can help build the instrument. It cannot absolve us of responsibility for what we claim the instrument demonstrates.

The Productivity Debate Is Asking a Different Question

It would be tempting to generalize from my experience and declare that AI has made software development hundreds of times faster. That would make an excellent headline and a fairly terrible engineering conclusion. Building an exploratory visualization from public data is not equivalent to maintaining a mature production platform with years of accumulated requirements, integrations, security obligations, and institutional knowledge.

Research has already demonstrated how misleading universal productivity claims can become. In a 2025 randomized study, METR found that experienced open-source developers working in familiar repositories actually took 19% longer when allowed to use the AI tools available at the time. A February 2026 update suggested newer tools might be producing better results, but the researchers acknowledged that selection effects made the size of any improvement difficult to establish.

Those findings don’t contradict my experience. They help define its boundaries. The value of AI depends heavily on the task, the environment, and what we’re trying to accomplish. Reducing the cost of building an exploratory instrument is a different achievement from accelerating modifications to a complex, established system.

But that distinction raises an uncomfortable question for organizations. How many ideas never reach a prototype because their expected value cannot justify the traditional cost of discovering whether they work? Our approval processes, investment committees, and product roadmaps were largely designed around a world where experimentation consumed substantial resources. When those costs change, some of the assumptions behind those processes deserve examination.

Organizations might discover that they can afford to explore far more possibilities before deciding which ones warrant serious investment. They can test an unconventional interface, examine an overlooked operational pattern, or make an abstract business problem visible without first establishing a project large enough to support its own governance apparatus. Naturally, security, data protection, and responsible evaluation still matter. A cheap prototype is not permission to make expensive mistakes in production.

What changes is the point at which we demand certainty. We may be able to investigate first and construct the business case afterward, rather than insisting that every idea arrive with a financial justification before anyone has seen whether it works.

The Questions We Never Asked Twice

There’s a broader implication beyond engineering and organizational productivity. For generations, we have filtered our curiosity through the practical cost of satisfying it. We learned to prioritize questions that justified research, funding, equipment, specialized expertise, or substantial amounts of our limited time. Many worthwhile curiosities never survived that filtering process.

I suspect that history contains an enormous number of abandoned investigations, not because their answers lacked value, but because the means of pursuing them were disproportionately expensive. We rarely notice these losses because they leave no evidence behind. There is no repository for the software nobody built, no archive of investigations abandoned before they began, and no productivity report accounting for the ideas that never escaped someone’s imagination.

My three little experiments haven’t transformed astronomy, commuting, or our understanding of violence. I wouldn’t pretend otherwise. What they have transformed is my willingness to follow an interesting question somewhere unexpected, even when I have no idea whether the destination will justify the journey.

After decades of programming, I find that rather extraordinary. I’m spending less time deciding whether I can justify building something and more time discovering what happens when I do. Some of those experiments will be worthwhile, others will be nonsense, and a few may lead somewhere I couldn’t have anticipated.

The questions were always free, but for most of my life, answering them came with a price I often couldn’t justify paying. AI hasn’t made curiosity more valuable or guaranteed that the answers will be correct. It has made more of that curiosity actionable, sometimes within the span of a coffee break.

We’ve become very good at asking how many hours AI can save us. I’m beginning to think the more interesting measure may be how many investigations it makes possible that would never have happened otherwise.

And I wonder how much of human curiosity has been waiting for this moment.

---

Further Reading

• The NASA Exoplanet Archive and Exoplanet Follow-up Observing Program (Christiansen et al., June 2025) Describes the infrastructure behind the scientific data used in the exoplanet visualization, including how the archive maintains and organizes confirmed planetary discoveries.
• 2020–2024 American Community Survey 5-Year Estimates (US Census Bureau, January 2026) The source of the geographic commuting statistics used in Going Somewhere Else. These are survey estimates, while the lifetime calculations in the visualization depend on adjustable assumptions.
• Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (Becker et al., July 2025) A randomized study demonstrating that AI assistance does not automatically accelerate every engineering task. Its findings concern experienced developers working on mature repositories, rather than rapid exploratory development.
• We Are Changing Our Developer Productivity Experiment Design (METR, February 2026) A useful follow-up explaining why researchers found it increasingly difficult to measure AI coding productivity reliably as tools and developer behavior changed.

---

Disclaimer: The perspectives shared in this article are our own and do not represent those of our employer or any affiliated organizations. All company names, product names, logos, and brands mentioned are the property of their respective owners and are used for identification and illustrative purposes only. No endorsement, sponsorship, or affiliation is intended or implied. References to specific companies or case studies are based on publicly available information and are used solely for educational and discussion purposes.