On January 2, 2001, Larry Sanger sat down to dinner at a Mexican restaurant in Pacific Beach, San Diego, with a problem he had failed to solve for months.
He ran Nupedia, a free online encyclopedia funded by the internet company Bomis, where Jimmy Wales was a partner. To be credible, Sanger believed, it needed experts, and its advisory board of mostly PhD professors had agreed on what he later called “an extremely rigorous seven-step system” for reviewing every article. By the fall of 2000, Sanger and Wales agreed it was moving too slowly. Sanger had written several proposals for something simpler. Wales, by Sanger’s account, turned each one down because it needed more programming than the company could pay for.
Across the table, an old internet friend, Ben Kovitz, described Ward Cunningham’s WikiWikiWeb: web pages anyone could start and edit. That night or the next morning, Sanger wrote a proposal. A wiki was very simple to set up and needed no new programmer, so, as he put it, “Jimmy could scarcely refuse.”
The plan was modest. The public would draft articles on the wiki, and that content could be fed into Nupedia’s review. It went live on January 10. A clear majority of the advisory board wanted nothing to do with it. So on January 15, 2001, the wiki relaunched under its own domain name, as Wikipedia.
“Who could reasonably guess that it would work?” Sanger wrote later.
Nobody could have planned this
By the end of January, Wikipedia had roughly 600 articles. By May it had 3,900, and the rate of growth was rising. Wales asked Sanger to spend more and more of his time on it. That summer Nupedia tried a managed wiki of its own, the Nupedia Chalkboard. It went largely unused. By the winter of 2001, Nupedia had published approved versions of only about 25 articles.
When cause and effect can’t be seen ahead, run small, cheap tries and grow what works.
In the Cynefin framework, this is the response to a complex situation. David Snowden and Mary Boone, writing in Harvard Business Review in 2007, describe a domain where right answers “can’t be ferreted out” in advance and you understand why things happened only looking back. Instructive patterns can still emerge, they write, if a leader runs “experiments that are safe to fail.” Leaders, in their words, “need to probe first, then sense, and then respond.” Responding means feeding what takes hold and letting the rest fade. Their own example is eBay, which began in 1995 as Auction Web on Pierre Omidyar’s personal website. The first item for sale was his probe.
The wiki fit that shape. It cost almost nothing. It sat on its own domain, away from Nupedia’s name. As it grew, Wales moved Sanger’s time toward it. Nupedia’s seven steps were the opposite: a careful design, fixed before anyone knew how volunteers would behave. Sanger wrote that a “clear mistake” of his was assuming many volunteers would patiently work through such a complicated system.
Two cautions. The story is told because the probe worked; the Chalkboard was a probe too, and it went nowhere. And Sanger doesn’t blame the process alone. Nupedia, he wrote, “withered due to neglect,” largely for lack of money for paid organizers once the internet bubble burst. Probing also isn’t the answer to everything. In Snowden and Boone’s scheme, a crisis calls for acting first, and a routine problem calls for the known fix.
Three weeks, two tries
Imagine your team is split between two offices that barely talk. You don’t know why, and a reorganization would take months to undo. So you try two small things for three weeks: a fifteen-minute call between the offices twice a week, and one pair, one person from each office, sharing a small task.
Before you start, you write down what you’ll watch, such as questions asked across offices in the team chat, and what would make you stop either one. By week two the call has gone quiet. The pair is still messaging after the task ends. You drop the call and start a second pair.
A model released in steps
In February 2019, OpenAI released the smallest version of its GPT-2 language model, 124 million parameters, and held back the larger ones over concerns about misuse such as fake news. It released a 355-million-parameter version in May, a 774-million version in August, and the full 1.5-billion version in November.
Between steps, it gave partners such as Cornell University and the Middlebury Institute’s Center on Terrorism, Extremism, and Counterterrorism access to study misuse, detection and bias, and monitored online forums. Its threat monitoring “did not find evidence of GPT-2 direct misuse in publicly-accessible forums,” though it saw discussion of misuse. The report, in November 2019, said it found minimal evidence of misuse. It also said this did not rule misuse out, because some actors, such as those at nation-state scale, are harder to monitor.
That last sentence is the part to borrow. A probe tells you only what your sensors can reach. When you roll out an AI tool to a team, start with a few people and one task. Write down what you will check, such as a sample of its outputs you review by hand each week, what would make you widen it or stop, and what your checks cannot see.
What a probe leaves behind
Early on, Sanger proposed a rule for the young wiki: “Ignore All Rules.” He meant it, he wrote, as “a temporary and humorous injunction” to add content and not fuss over formatting. He and Wales also agreed not to throw anyone out except in extreme cases, a hands-off policy he saw as provisional. It hardened, he wrote, into “a difficult-to-change tradition.”
Sanger later came to reject the rule altogether. One participant commented that “this rule is the essence of Wikipedia.”