From One Shot to Many: How AI is Changing the Cost of Trying
When experimentation costs nothing, what becomes possible?
Years ago, over plates of pasta at a neighborhood Italian place, a friend looked up from his barely touched dinner and announced, "I'm going to quit my job and start a restaurant."
The table went quiet. Another friend - the kind who reads financial statements for fun - set down his fork. "That's not a good idea," he said, not unkindly. "Half of all new restaurants fail in the first year." He wasn't being cruel - he was trying to protect our friend from a financially risky move.
"Don't worry," I quipped, barely looking up from my own plate, "just start two restaurants."
The table laughed, and the conversation moved on. But that offhand comment - meant as a joke about doubling down on risk - accidentally captured something important about how the world is changing. What if you really could try two restaurants? Or five? Or ten? What if the cost of experimentation could drop so low that you could test multiple concepts before committing to one?
This isn't just a thought experiment. It's already happening in some of the most expensive and high-stakes industries in the world, where the cost of failure has traditionally been measured in years and billions of dollars.
Finding Keys in the Dark
Take the pharmaceutical industry. To understand why drug discovery is so expensive, let's look at how a breakthrough drug like Prozac came to be. At its core, drug discovery is like finding a key that fits a very specific lock in your body. The lock might be a protein - a complex molecule that controls important processes in your cells. In Prozac's case, that "lock" was a protein that controls how your brain processes serotonin, a chemical messenger linked to mood and depression.
Finding this lock took years of painstaking research. Scientists had to first understand that depression might be linked to serotonin levels, then identify the specific protein that regulates it. But finding the lock was just the beginning. Next came the harder part - designing a key that would fit perfectly.
Imagine trying to craft a key without being able to see the lock clearly. You're working at a molecular level, testing thousands of slightly different chemical compounds, hoping one will interact with the protein in exactly the right way. Each attempt requires careful synthesis in the lab, precise testing, and rigorous checks. A single promising molecule might take months to validate, only to fail in unexpected ways.
This is why developing a new drug traditionally takes 10 to 15 years and costs over $2 billion. Scientists must methodically test each possibility, knowing that a single misstep early in the process can waste years of work and hundreds of millions of dollars. It's the restaurant problem at a massive scale - the cost of experimentation is so high that companies must bet everything on a handful of promising candidates.
But in 2023, something remarkable happened. Insilico Medicine, using their AI platform PandaOmics, identified a novel target for idiopathic pulmonary fibrosis (IPF) - a devastating lung disease. Their AI system, Chemistry42, then designed and optimized a drug candidate in months, not years. The entire process from target discovery to clinical trials took just 18 months.
Think about that for a moment. A process that traditionally takes the better part of a decade, compressed into a year and a half. But the speed isn't even the most important part. What's revolutionary is how many possibilities they could explore. While traditional methods might evaluate a few thousand compounds per year, Insilico's AI platform screened billions of potential molecules in weeks. It's like having thousands of scientists, each testing different keys, but working at the speed of light.
The Shenzhen Effect: When Small Became Possible
This transformation isn't unprecedented. Two decades ago, a similar revolution happened in manufacturing, centered in a Chinese city most Americans had never heard of: Shenzhen.
Before Shenzhen, creating a new electronic product was a bit like traditional drug discovery - expensive, time-consuming, and reserved for big companies with deep pockets. Want to create a new type of smartwatch? You'd need millions in capital, relationships with manufacturers, and probably two years of development time. The cost of failure was so high that venture capitalists had an unofficial rule: no hardware startups.
Then something changed. Shenzhen, once a small fishing village, evolved into a manufacturing ecosystem where the cost of experimentation dropped to nearly zero. Imagine a city-sized laboratory where you could prototype any electronic device imaginable, with thousands of suppliers and manufacturers within walking distance of each other.
Take the story of Pebble, one of the first modern smartwatches. In the old world, creating a new watch category would have required the resources of a Samsung or an Apple. But Eric Migicovsky, a young entrepreneur, was able to go from concept to prototype to mass production with just Kickstarter funding and Shenzhen's ecosystem. What would have taken years and millions of dollars could now be done in months for thousands.
The impact went far beyond just making things cheaper. Entirely new categories of products became possible. Specialized gaming accessories, custom mechanical keyboards, boutique audio equipment - products that would never have reached the minimum volumes for traditional manufacturing suddenly became viable. The cost of experimentation had dropped so low that you could profitably serve even tiny market niches.
This wasn't just about making things for less money - it was about making different things entirely. When you can try ten different approaches for the cost of what one used to be, you start thinking differently about what's possible.
The Economics of Not Yet
This pattern is about to repeat itself even more with knowledge work. Think about what it takes to build software today. You need developers to write code, designers to craft interfaces, product managers to coordinate efforts, and marketers to reach customers. A "minimal" team might cost half a million dollars a year. Before writing a single line of code, you have to ask: "Is this market big enough to justify that investment?"
This is why so many important but niche problems go unsolved. Want to build software for independent bookstores to manage author events? Or a platform for high school theater programs to coordinate productions? These might be vital tools for their users, but the traditional math doesn't work. The cost of building the solution is too high relative to the size of the market.
It's the same constraint we saw with manufacturing before Shenzhen, and with drug discovery before AI. The cost of the first attempt is so high that you can only chase the biggest opportunities. You can't afford to experiment, to try serving smaller markets, to learn what works.
But what happens when AI can handle much of that knowledge work? When you can describe what you want to build and have AI generate most of the code, design the interfaces, and help craft the marketing message? Suddenly, the math changes. Markets that were too small to be profitable become viable. Ideas that were too risky to try become worth testing.
This isn't just about making software development cheaper - it's about making different kinds of software development possible. Just as Shenzhen enabled niche hardware products, AI could enable highly specialized software for markets that were previously too small to serve.
Which brings me back to that dinner conversation about restaurants, and a statistic we've all heard: "9 out of 10 startups fail." We quote this number like it's a law of nature, but it's really a reflection of a world where each attempt is expensive - where you have to bet everything on a single approach because you can't afford to try again.
What if you could test ten different markets simultaneously? Build minimal versions of products for ten different niches, learn which ones resonate, and then focus on the most promising opportunity? The cost of failure wouldn't be the end of the startup - it would just be information.
The joke about starting two restaurants was funny because it seemed absurd - who could afford to double their risk? But across more and more fields, that absurdity is becoming reality.
So maybe it's time to ask a different question: In a world where the cost of experimentation is approaching zero, what would you do if you could start ten?