Evaluating Decisive Strategic Advantage: Part 1

Much of AI policy thought is predicated on the assumption that if one actor has an AI sufficiently more advanced than the other, it would become possible to use it to disempower any opponents, and achieve total world domination. This idea is often referred to as “Decisive Strategic Advantage” (DSA).

This post is the first of a multi-part series that aims to discuss what concrete pathways could enable an actor to gain a decisive strategic advantage. In this article, we’ll discuss the idea’s history, and establish the analytical framework that we’ll apply to each pathway in future articles.

For each of these pathways, we’ll attempt to find:

1. How plausible is the pathway? Considering all factors, is it likely that an actor could exercise it to achieve world domination?

2. What is the technology readiness level? Does it rely on technology widely considered plausible, like drones, or does it posit fundamental breakthroughs like nanotechnology?

3. Which actors is this pathway available to? Is it available only to a country, or would it be possible for any individual with access to the AI?

4. What material precursors does the pathway need? For example, nuclear weapons require enriched uranium or plutonium, bioweapons need a wet lab, but cyberweapons need only a computer.

5. What would the breakout time be? How long would the actor need to control the superintelligence for? Could a rogue employee who had control of the AGI for only a few hours achieve DSA? What about a military coup that held it for a few days?

6. What are potential counters for the pathway? How can we reduce or mitigate the threat from a rogue actor wielding this weapon?

Background

History

The exact term “Decisive Strategic Advantage” can be attributed to Nick Bostrom, in Superintelligence (2014), but the idea has had precursors.

On the AI theory side, statistician I. J. Good was the first to discuss the concept of the “Intelligence Explosion” in 1965, noting that an ultraintelligent machine would be better suited to the development of even more intelligent machines than humans, and the cycle could continue until all humans are left far behind. His ideas were popularized in the 1990s by Vernor Vinge, a science fiction author. Vinge predicted the Singularity, a point in time after which recursive self improvement by superintelligences would lead to a total upheaval of everything we know. Vinge’s writing was influential for both Eliezer Yudkowsky and Nick Bostrom. Yudkowsky, an early AI safety advocate, started to popularize the risk from artificial superintelligence as early as 2001, predicting that it would lead to human extinction if left unchecked. Yudkowsky noted that an AI takeoff caused by recursive self improvement could happen very fast, perhaps in a matter of days, or even hours. He also coined the term “pivotal act” to describe using an aligned AI to take unilateral action to halt all competing (potentially unsafe) AI projects and thereby lock in a good future. Nick Bostrom was also influenced. In 2005, he wrote an article “What is a Singleton?” where he considers the possibility of a permanent world government enabled by AI.

Ideas similar to DSA can be found in military strategy as well. A close analogue to the situation of being the first one to develop superintelligence can be found during the early stages of the Cold War, around the late 1940s to the mid 1950s. (Trachtenberg) During this time period, the usual logic of mutually assured destruction didn’t yet apply: only the US, and not the USSR, had practical atomic weapons and the means to deliver them. Of course, both sides knew that the window to act was short, as the Soviets had their own nuclear weapons program on the way, guided by leaked data from their spies. Several hawks, including senior members of the military, advocated for a pre-emptive strike, or “preventive war”, although the idea never made it past a strategy exercise. Eisenhower balked at the expected death toll of such a war, stating “You can’t have this kind of war. There just aren’t enough bulldozers to scrape the bodies off the streets.” After the USSR developed bombers and eventually ICBMs, the possibility of launching a pre-emptive strike without massive retaliation dwindled.

Recent Developments

AI technology has progressed immensely since 2014, driven largely by the advent of Large Language Models (LLMs). AI is in common use powering chatbots, coding assistants, and automating some repetitive light white-collar work, and all indications point to further progress. Benchmarks like METR’s time horizons show that AI is making exponential progress year after year. It’s unknown whether this trend will continue, but large companies like Google, Amazon, and OpenAI are investing billions into new datacenters to train the next generations of AI. Some forecasters predict that Artificial General Intelligence (AGI) is possible by 2027, although a rigorous analysis of whether AGI is possible, and if so, when it will arrive are matters for a different article.

The looming backdrop of AGI has brought the framework of decisive strategic advantage into the limelight. While not explicitly named, it often seems to be a background assumption in a lot of AI policy work. Much of this is attributable to Leopold Aschenbrenner’s influential paper “Situational Awareness” (2024). Situational Awareness was widely read across much of the tech and government spheres. Notably, Ivanka Trump publicly recommended it, calling it “an excellent and important read”, indicating some extent of its reach into the US policy world.

Situational Awareness hypothesizes that world governments will eventually “wake up to AGI”. Recognizing that exclusive control of AGI will grant their nation unprecedented influence over the future, both the US and China will begin a race to build AGI. The paper claims that this will result in a few key effects. First, the growing power and data needs of AI would result in a massive amount of datacenter and energy buildout. Secondly, the US government is likely to nationalize AGI development to some extent, recognizing the risks if the labs are hacked or otherwise compromised. Additionally, they would centralize research, wanting to avoid duplicating work between labs. The US AI effort would be similar to how they behaved in the 1940s during the Manhattan Project.

While the government hasn’t yet shown any plans to centralize AI development or intervene within the AI industry, President Trump’s recent AI Action plan aligns with Situational Awareness in framing the development of AI as a “race to achieve global dominance in artificial intelligence”.

Recent developments have also helped to constrain our models of DSA. When the concept of DSA was developed in the early 2010s, the current LLM paradigm had not yet been solidified. Many of Bostrom’s contemporaries considering the possibility of AGI, such as Eliezer Yudkowsky, were initially skeptical of deep learning, and considered an approach based on explicit programming more feasible. This hypothetical form of AI would be closer to an old fashioned chess engine, using pure logic and search to achieve its goals. Since it doesn’t require a lengthy training process, its architecture would have much shorter iteration times. As such, an intelligence explosion based on this architecture could result in a “fast takeoff”, with the AI going from human level to strongly superintelligent within a timespan of hours to weeks. However, if we extrapolate out the current LLM paradigm, it would be a “slow takeoff” at best, taking months to years to become strongly superintelligent.

This has important implications for DSA: a fast takeoff regime means that the first actor to “take off” would be far more advanced than any of its rivals, and be able to gain access to technologies otherwise decades or centuries off, like nanotechnology. However, in a slow takeoff regime, multiple projects might be taking off at the same time. In such a regime, even though each project’s capabilities are improving rapidly, no one project has a massive lead over the other. Thus, achieving a DSA tends to be less likely in slow takeoffs.

Criticism

The theory of decisive strategic advantage is not without criticism. (overview) First, critics note that building a DSA may not be as easy as proponents think. The raw intelligence enabled by scaling up AI may not translate over to military advantage as fast or as well as assumed. One reason is simply that building a “wonderweapon” may take more than just pure intelligence, requiring significant amounts of power, physical experiments or other costly and time-intensive resources. Another is that integrating the weapon into the existing military infrastructure may be difficult, and that a novel weapon may not function correctly when tested in the wild. Secondly, even a strong weapon enabled by AI may not prevent retaliation from the victim. Even if the AI could launch a devastating attack against another country, it’s possible the country may still be able to launch a few nukes or take other actions that would make this costly for the aggressor. Additionally, the AI advantage of the leading actor may easily diffuse as secrets leak.

Superintelligence Strategy accepts that DSA is possible in theory, but argues against it by stating that if other countries believe that DSA is possible, then they will be strongly incentivized to launch a pre-emptive cyber or even kinetic attack to prevent the leading nation from reaching a level of intelligence that would enable them to create a DSA. It’s important to note that currently, frontier AI development is quite vulnerable to both kinetic and cyber attack, and this enables other countries to easily sabotage them.

Victory Conditions

Before we discuss the mechanisms for DSA, we need to carefully evaluate what it means for an advantage to be truly “decisive”, and what victory even looks like. In Superintelligence, Bostrom introduces the concept of a singleton, a permanent world government that has only one decision-making agency. Bostrom defines a “decisive strategic advantage” as the ability to create a singleton, and use it to “dictate the future”. This is fairly broad, and we’ll need to refine his definition somewhat in order to make it easier to judge whether a given intervention meets this threshold.

First, Bostrom was vague on what constituted a singleton. I argue that given the backdrop of slow takeoff, an actor that can permanently halt all competing attacks and the technological development of its rivals, while continuing its own technological development will eventually come to dominate the world. Second, Bostrom takes care to disambiguate “decisive strategic advantage” from “willingness to exercise said DSA”. However, I argue that this is a factor that should be taken into account when evaluating whether a pathway is a DSA. Many actors are risk averse, or have deontological or other political constraints. A pathway that an agent would never use, even as a form of retaliation, can hardly count as a DSA.

Thus, for the purpose of this article, we will operationalize an actor having “decisive strategic advantage” as that actor having the unilateral ability and plausible political will to halt all other actors from developing stronger technology than itself, while at the same time continuing its own technological development.

This leads to two important notes: First, what may be a DSA for one actor may not be for another actor. For example, consider a bioweapon designed to kill 99% of humans except for a select few who were immune. While a rogue terrorist group might consider this acceptable, for a democratic nation, it’s not an option on the table.

Second, achieving the DSA does not necessarily have to result in actual warfare, although in practice it will usually be implicitly backed up by that. For example, an enforceable international treaty creating a “CERN for AI” with strong punishments for defecting countries doesn’t involve explicit violence, but would grant a DSA to the collective leadership of the countries leading the project.

In general, the pathways we’ll consider tend to fall into one of two buckets:

• Public capabilities: These are capabilities that can be publicly revealed to other actors, and still be used. For example, nuclear weapons are a public capability of the US. If a pathway falls in this bucket, the mere threat of use can force other actors to back down without the pathway ever being exercised. Usually, these capabilities can only be held by nation-states.

• Covert capabilities: These are capabilities that must be held secret in order to exercise them, since if other actors knew about it, the pathway would be neutralized. For example, a secret drone swarm held by an AI lab is only valuable if the state doesn’t know about it. These capabilities can be held by both nation-states and non-state actors. These capabilities tend to result in greater harm, since in order to be useful, they have to be exercised.

Pathways can also be categorized in the vocabulary of nuclear strikes. There are three primary kinds of strategic nuclear strikes:

Countervalue: A countervalue strike aims to destroy a country’s sources of value. Bombs would be aimed at population centers, food sources, economic hubs, and infrastructure. This would typically be the target of a retaliatory strike. The more punishment you can inflict, the greater the deterrent value.

Counterforce: A counterforce strike aims to target an adversary’s military forces and war-supporting capabilities. This includes their nuclear forces, like ICBM launch sites and bomber bases, command and control nodes, and even their conventional forces. The purpose is to disarm the opponent and ensure they cannot strike you first.

Decapitation: This is a subset of the above, where the strike would be targeted solely on the leadership of a country, especially its nuclear command, control, and communications.

Mutually Assured Destruction vs Decisive Strategic Advantage

The concept of DSA contradicts the now prevailing military doctrine of mutually assured destruction (MAD). MAD is built on the idea that an opponent will survive a first strike, no matter how damaging, and be able to launch a second strike that will inflict unacceptable damage back onto the first country. Thus, it is never in a country’s interest to launch a first strike. In contrast, DSA postulates that with sufficiently advanced technology, one can eliminate an opponent’s ability to meaningfully retaliate.

However, MAD can be defeated if we undermine any of its assumptions:

Leaders must be able and willing to carry out a second strike

• A pacifist leader who others expected would not retaliate makes their country more vulnerable to first strikes.

• A decapitation strike, leading to total elimination of the opponent’s leadership, could remove their ability to order a retaliatory strike.

Ability to launch a second strike

• A perfect counterforce strike could eliminate a country’s ability to launch a second strike, if all ICBMs and SLBMs (submarine launched ballistic missiles) were incapacitated.

Ability for second strike to punish opponent in way that matters

• MAD would be undermined if an opponent nation constructed a system of nuclear shelters that enabled their population to ride out a retaliatory strike.

• MAD would also be undermined if the oppositional country had an anti-ballistic missile defense system.

• Terrorists and dictators may care neither for their own life nor for the lives of their country’s citizens. Thus, they cannot be deterred by a countervalue strike.

A DSA must undermine at least one of the properties listed above, otherwise it can be deterred by a retaliatory strike the same way pre-emptive nuclear weapon use would be.

Assumptions

The concept of decisive strategic advantage itself doesn’t assume a particular mechanism. In theory, a variety of mechanisms could grant a sufficiently superintelligent AI the ability to dominate the world, many of which are “unknown unknowns”. This makes our attempt to enumerate the pathways by which an actor could gain a decisive strategic advantage quite difficult.

However, there are a few heuristics which we can leverage to constrain our search:

1. Early ASI will likely reach technologies we know about first before moving on to things we don’t know about, as these technologies are “easier”. Thus, the first dangerous technologies to be developed by AI will likely be already on our radar as “near-future” or “next-gen” technology, things that we estimate would be developed anyway sometime in the next 20 to 50 years barring AGI.

2. An actor willing to exercise a DSA would have a strong incentive to take the “first viable DSA”, especially in a multipolar situation with fog of war. If the actor knew that they could achieve world domination with near-certainty, then waiting poses risks due to competitors catching up, and the uncertainty of gaining a future DSA with similar properties.

Taken together, these heuristics suggest that the mechanism of DSA will probably be a technology that we already know could conceivably be created in the future. However, we’ll also take the time to discuss more speculative options.

Note that these heuristics have held up in the past. The atom bomb’s mechanism was famously predicted 30 years in advance by H. G. Wells, in his 1914 novel “The World Set Free”. Wells knew, from the contemporary physics of his day, that radioactive decay of radium contained tremendous amounts of energy. Although he didn’t know the exact mechanism, he correctly identified, based on the laws of physics, that if atomic energy were unleashed at a much higher rate than natural radioactive decay, it would become an extremely dangerous weapon. This points to our ability to identify potential pathways before they are instantiated, although we may not get the details entirely correct.

What’s next in the series

Accepting the assumption that it’s possible to identify potential pathways for an actor to gain a DSA, we now need to enumerate the most likely of these pathways, and then evaluate their plausibility.

Many of these pathways have been considered by others before, but not in a uniform way. Part of the utility of this series is to serve as a sort of literature review over potential DSA mechanisms brought up by others.

This is not an ordered list, and we note that some of these pathways will be combined in a particular attack, but in the interest of space, we only list those pathways which might be “independently sufficient” to mount an attack:

• Enabling Counterforce Strike

• Autonomous Weapons: Drone Swarm

• Chemical and Biological Weapons

• Industrial Capture

• Social Engineering

• Cyberweapons

• Nanotechnology

• Unknown Unknowns

We’ll investigate them in order. The next post will cover the possibility of achieving DSA by having the ability to launch a pre-emptive counterforce strike.