At the Rise of AI Summit 2020, Dr. Jesse Lehrke, a researcher at Freie Universität Berlin who describes spending recent funding breaks learning Python and neural networks, argues that most current AI strategies are really just strategies for winning the AI race, not for what to do once AI, or a technological singularity, actually arrives. He frames strategy as the moves and countervailing moves opposing wills use to get more out of a situation than the underlying balance of power suggests, particularly useful for weaker parties, and argues academics should not fear "futurology" given the stakes. Because the singularity is by definition unknowable, Lehrke proposes thinking about the limits physics imposes on any actor, human or AI, including the speed of light and quantum-level constraints on secrecy, as tools that can be used to shape an AI's behavior through embedded uncertainty rather than eliminated uncertainty. Drawing on Clausewitz's trinity of passion, reason and chance, and on nuclear deterrence as a model of control through non-use, he argues design decisions made today, not decisions made after the AI race is won or lost, determine how controllable future AI actors will be.
Key takeaways
- Lehrke argues most published AI strategies are really only strategies for winning the AI race, not for what to do once it ends.
- He argues weaker parties can gain leverage by manipulating a stronger actor's, including an AI's, physical limits rather than trying to close a capability gap.
- He reframes Clausewitz's trinity of passion, reason and chance as decision, information and uncertainty for the AI era.
- He argues academics' fear of being called "futurologists" is itself a risk, since a topic's importance, not its ease of study, should determine whether it is studied.
- In Q&A he argues supervised learning on labeled data is a dead end compared with semi-supervised, unsupervised and reinforcement learning.
Detailed notes
Strategy, the singularity and why most AI strategies fail
Lehrke defines strategy as the moves and countervailing moves opposing wills use to extract more from a situation than the initial balance of power suggests, a logic especially useful to weaker parties. He frames the singularity, the point where technological capability exceeds human comprehension, as his second focal point, arguing it is more useful to discuss than incremental AI steps because it is an extreme, unqualified case, even though by definition no data exists about what lies beyond it. He argues that most published AI strategies, despite their titles, are only strategies for winning a zero-sum "AI race" up to that point, not strategies for how to actually use or control AI once the race is over, and that current security planning suffers from a "lack of imagination."
Why planning for the unknowable is still necessary
Anticipating the objection that planning for an unknowable singularity asks too much, Lehrke argues it is both possible and already happening: he cites Henry Kissinger writing on AI despite admitting he knows little about it, and historical military strategists who envisioned air power and naval power as dominant tools before the technology existed. He says academics' fear of being dismissed as "futurologists" is itself a danger, because the value of studying a topic lies in its importance, not its ease. He argues the singularity's dual religious-and-scientific character requires an interdisciplinary approach, and that even though pursuing AI is partly a belief-driven, self-fulfilling process, some form of advanced AI or singularity technology is effectively unavoidable, since any country's restraint only lets others advance along a different path toward it.
How the nature of conflict changes: interaction, violence and latent actors
Using Clausewitz's concept of the nature of war as interactive, violent and political, Lehrke argues all three elements will change as AI advances. On interaction, he says today's framing of an "AI race" between known human rivals will give way to facing AI-enabled adversaries whose capabilities are unknown, and eventually to AI itself as a "latent actor" that political science's usual models, built for human rivals, are not equipped to handle. On violence, he argues speed will matter not just as faster hardware but as faster thought, production, via real-time additive manufacturing updates, and faster outcomes, potentially producing conflicts resolved without visible fighting at all, which complicates the traditional strategic goal of defining "victory" and "political purpose."
Managing uncertainty through the limits of physics
Lehrke reframes Clausewitz's trinity of passion, reason and chance as, respectively, decision, information and uncertainty, and argues that in the singularity era, strategy should not always aim to reduce uncertainty, since much of it cannot be reduced, but instead to manage or redirect it, including by increasing an adversary's uncertainty in ways that expand rather than threaten their options. He proposes an axiom that any agent with limits behaves differently because of those limits, and that understanding an actor's limits, including physical limits such as the speed of light and quantum-level constraints on what can be observed or kept secret, lets a weaker party shape a stronger or AI actor's behavior. He illustrates this with a thought exercise: an AI with unlimited time but a processing speed capped at the speed of light faces so many possible futures to evaluate that maximizing its choices becomes a form of deterrence, an argument he ties to design decisions available today, such as deliberately embedding uncertainty into an AI system using existing tools like Bayesian statistics and chaos theory.
Q&A highlights
Asked for practical advice for startups and investors, Lehrke says giving AI systems a genuine conception of time and space, building on existing mechanisms like LSTMs and attention, is an important next step, and gives high-frequency trading's physical placement between exchanges as an example of space and time already mattering commercially. Asked whether investing in high-quality labeled data provides durable advantage, he says he believes, citing Google's own direction, that supervised learning on labeled data is a dead end compared with semi-supervised, unsupervised and reinforcement learning, and warns that Europe's strengths in physics and robotics research could be undermined if its AI initiatives are bought out by companies like Facebook or Huawei before their longer-term payoffs materialize.
Quotes
Strategy is actually most useful when you are weak. It's needed to counter weakness and again get more out of the situation than the balance of power suggests.
Jesse Lehrke
Everything has limits, and an agent with limits thinks and acts in a different way. They take those limits into account in their behavior.
Jesse Lehrke
Questions answered in this talk
Why does Jesse Lehrke think most current AI strategies are not real AI strategies?
According to Jesse Lehrke at the Rise of AI Summit 2020, most published AI strategies only aim at winning a zero-sum AI race and end the day that race ends, rather than planning for how AI will actually be used and controlled afterward.
How can a weaker actor gain leverage over a more powerful AI, according to Jesse Lehrke?
According to Jesse Lehrke at the Rise of AI Summit 2020, by understanding and manipulating the physical limits any actor faces, such as the speed of light or quantum-level constraints on information, since an agent's behavior can be shaped through those limits even without matching its overall capability.
Is labeled training data a lasting competitive advantage in AI, according to Jesse Lehrke?
According to Jesse Lehrke at the Rise of AI Summit 2020, no: he argues supervised learning on labeled data is a dead-end path, and that semi-supervised, unsupervised and reinforcement learning represent the more durable direction.