Calls to "pace the frontier" aren't new. We've been here before—for example, with environmental modification tech. During the Cold War, the superpowers worried about the dangers of unleashing "a Pandora's box" of foreseeable risks. Getting to yes still took 10+ years. Why? 2/

The problem is an informational catch-22 that I call the 𝙤𝙥𝙖𝙘𝙞𝙩𝙮 𝙩𝙧𝙖𝙥. Nobody wants to leak technical secrets that could benefit rivals. So they conceal details about what they truly know. Yet when everybody conceals, nobody can tell who's likeliest to pull ahead. 3/

The opacity trap massively complicates anticipatory cooperation. Why? Information about who will win a race usually only arises near the finish line, as efforts scale. By the time states learn whether competition is dangerous or futile, somebody is ready to deploy. 4/

In the article, I examine how the opacity trap has panned out in past cases like ENMOD. Fear of the trap itself is sometimes enough to motivate cooperation. But this is rare (I estimate 4-10%). Absent contrary evidence, US officials assumed that if they wanted to win a tech competition, they could. Deals only happened once new intel could prove that assumption wrong. 5/

Do we want to "run faster," "trip the competition," or simply pump the brakes? The US has just said it will prioritize tech competitions it can win—and consider cooperation where it cannot. But telling them apart can be fundamentally hard without the benefit of hindsight! 6/

The opacity trap is playing out for AI right now. Cooperation means trading innovation for insurance that nobody is sure they need. But the frontier model gap is apparently narrowing. The negotiating table is starting to seem more attractive. Ultimately, getting to yes may depend on whether competition seems intense enough. 7/