In 2019, Jos Benschop stood up to speak at a workshop and took out his Samsung phone. ASML, the Dutch maker of chipmaking machines, had brought him in back in 1997 to start its extreme ultraviolet program, EUV for short. By his own account, he told the room: “I once said naively that EUV would be in volume production in 2006. I shamefully admit that it’s 13 years late.”
The machine he had promised was brutal to build. EUV light is 13.5 nanometers, and almost everything absorbs it, air and glass included, so the machine has to run in a vacuum and steer the light with mirrors. To make the light, a laser hits droplets of molten tin 50,000 times a second, twice each.
Thirteen years is a long time to fund a promise. Through all of it, ASML had a business to run.
Two shipments in one year
In 2006 ASML shipped its first EUV prototypes, to imec in Belgium and to a university lab in Albany, New York. That same year it shipped its first volume-production machine for immersion lithography, a refinement of the 193-nanometer technology it already sold.
One shipment was a bet on a new kind of light. The other was the old light, made better. The new light source kept slipping; “they practically crucified us at SPIE,” Benschop said of customers quoting ASML’s old roadmap back at it. ASML kept improving the old line, and by its own account two later immersion models, “the work horses of the chip industry,” went into factories around the world.
Getting better at what already works can crowd out the search for what works next. James March, of Stanford’s business school, set out the tension in a 1991 paper, “Exploration and Exploitation in Organizational Learning.” Exploration, he wrote, covers things like “search, variation, risk taking, experimentation, play, flexibility, discovery, innovation.” Exploitation covers “refinement, choice, production, efficiency, selection, implementation, execution.” A company that only explores pays for experiments without getting much back. One that only exploits is likely to find itself “trapped in suboptimal stable equilibria.”
Why the old line wins the argument
The two don’t compete on equal terms. The returns from exploration, March wrote, are “systematically less certain, more remote in time, and organizationally more distant” than those from exploitation. For ASML, a better immersion machine could sell this year. A brighter EUV source was a promise that kept moving.
So organizations learn, sensibly, to favor the sure thing, and each success makes the next one more likely. “Reason inhibits foolishness; learning and imitation inhibit experimentation,” March wrote. He warned that competence in an inferior activity can grow great enough to shut out a better one the organization hasn’t tried.
Order your usual dish and dinner is good tonight. Try the new place and dinner may be worse, but you learn something the usual can never tell you.
Kodak split it differently. It let its digital camera inventor keep working on digital cameras, compression and memory cards, yet declined to sell the 1.2 megapixel camera he built in 1989 because it would eat away at film sales.
ASML had help balancing. In 2012 Intel, TSMC and Samsung agreed to put €1.38 billion into its research over five years, aimed at next-generation lithography, EUV included. Not every exploring company has its biggest customers paying for the trip.
March’s own evidence was two simple simulation models, not company data, and he offered no formula for the right balance. Some of his findings run against intuition: in one model, an organization learned more in the long run when some of its members were slow to adopt its received wisdom, because their deviations kept new ideas alive long enough for the organization to learn from them.
The account that always pays
Imagine you manage six sales reps. Every quarter, their renewals come in on schedule, and your targets are built on them. A new kind of customer, small clinics, keeps asking questions none of your reps has time to answer.
You give one rep every Friday for clinics. For two quarters her numbers are the worst on the team, and in the reviews her Fridays look like waste. In the third quarter, the clinics account for three of her new deals and most of what you’ve learned about what they need. Your renewal numbers dipped a little to pay for it.
The button nobody pressed again
Machines that learn by trial and error face the same choice. Researchers call exploration a core capability for decision making. In a study published at NeurIPS 2024, a team at Microsoft Research tested whether GPT-3.5, GPT-4 and Llama 2 would explore. In the main version of a simple game, there were five buttons, and one paid off more often than the others.
With a basic prompt, GPT-4 played 500 rounds, repeated 20 times. In more than 60 percent of those runs, it pressed the best button only a handful of times, and after some point never pressed it again. It behaved much like a greedy rule, which always picks whatever has the best average so far.
Of all the setups tried, only one explored well: GPT-4 with a hint to explore, a chain of thought, and the history handed to it already summarized, as the number of presses and the average payoff for each button. The authors caution that such summaries may not be possible in messier settings. And these were the models available when the study was run, in a toy game.
If you let an AI agent choose again and again among options, such as which supplier to call first, keep your own table of how often each option was tried and how it did. Give it the table. Then check whether any option stopped being tried early, across several runs, since one run can differ from the next.
In 2019, Samsung released the first consumer product with EUV-made chips, a phone. At the workshop that year, after admitting he was 13 years late, Benschop waved his own Samsung phone at the room and added: “but some of you thought this thing would never exist.”