Nick Bostrom: Superintelligence Soon?

Are we living in the final moments of the slow era?¹

Nick Bostrom: Superintelligence Could Be a Year Away – or Not

Back in 2012, when I first chatted with Nick Bostrom², AI risk was a fringe academic curiosity, rarely surfacing in mainstream discourse. That changed in 2014 with Nick Bostrom’s Superintelligence, which established a rigorous framework for the dangers ahead. Yet even those once-radical projections now seem conservative. In our recent conversation, Bostrom’s assessment of the horizon for superintelligence has narrowed to the immediate future: ‘We are now in a position where we can’t be confident it couldn’t happen within a very short timeframe—like a year or two.’

Back in 2012, when I first chatted with Nick Bostrom about AI risk, the topic was a niche blip – dung beetle mating habits had more academic ink. Fast forward to 2025³, post his 2014 game-changer ‘Superintelligence: Paths, Dangers, Strategies’, and AI’s sprinting ahead. So, how close are we to transformative AI or superintelligence? Bostrom’s answer in our latest interview is a jolt: “We are now in a position where we can’t be confident it couldn’t happen within a very short timeframe – like a year or two.”

This is less a prediction than a klaxon. A single breakthrough – a lab-born eureka – could flip the switch overnight. We might be one missing trick away from the floodgates opening.⁴ While a gradual climb is probable, the risk of a vertical leap remains. In the blur of AI’s acceleration, Bostrom’s warning is simple: blink and the world changes.

We can’t rule out a sudden leap.

The Erosion of Time as the Safety Buffer

When Bostrom first outlined the paths to superintelligence in 2014, the “software bottleneck” was viewed as a multi-decade challenge. Today, that bottleneck looks increasingly brittle. The current era of large-scale models suggests that we may already possess the raw computational power required for AGI, but lack the specific algorithmic “key” to unlock recursive self-improvement.

This is the “single trick” theory: the idea that our current architectures are essentially complete, baring one or two insights into efficient reasoning or long-term memory. If this is true, the transition from a so called “stochastic parrot” to a self-evolving system wouldn’t likely be a slow climb, but a phase transition. In such a case we are no longer waiting for more hardware; we are waiting for a moment of clarity in a research lab.⁶

Beyond the Transition: The Stakes of “Soon”

A compressed one-to-two-year window forces us to mitigate immediate existential risks while simultaneously locking in the initial conditions of our long-term future. This rapid transition compromises more than just physical safety; it precludes the deep, generational reflection required to preserve human agency in a world where our relevance is no longer a given.

In his recent work on Deep Utopia, Bostrom explores what happens when scarcity and labour are solved. If the leap happens “soon,” we are forced to decide which human values we want to entrench before we truly understand them – it threatens our ability to thoughtfully design a world where humans remain relevant through means we are used to – i.e. through argument, conflict, revision, and hard-won consensus across generations. Which is why I often cite value discovery and realisation accelerants and processes for ‘choosing how to choose’ accelerants like indirect normativity.

Transcript of video short

Adam Ford: “How far away do you think we are from transformative AI, or superintelligence?”

Nick Bostrom: Well, progress has been rapid. think we are now in a position where we can’t be confident that it couldn’t happen within some very short timeframe, like a year or two. I it could sort of happen at any time. Like if somebody in some lab gets just some key insight that makes a big unlock. Like if it turns out that all these current big models and reasoning models we’re training, if there’s like one thing we’ve all been doing wrong, and somebody figures out that, maybe that would be enough. I think more likely it will take a bit longer, but yeah, we can’t be confident that there is some huge long runway in front of us at this stage.

We are now in a position where we can’t be confident it couldn’t happen within a very short timeframe – like a year or two.

Nick Bostrom, in interview with Adam Ford, 2025

Footnotes

1. Cf. Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies – ch4 discusses the dynamics of “slow,” “moderate,” and “fast” takeoff speeds, noting that a fast takeoff (days or weeks) leaves virtually no room for human intervention once the process begins. ↩︎
2. For context on the early state of the field, see this early 2012 interview, where the discussion focused on the theoretical possibility of superintelligence and it’s risks, rather than the immediate technical hurdles we face today. ↩︎
3. See full 2025 interview with Nick Bostrom. ↩︎
4. The “floodgates” referred to here are often linked to the concept of instrumental convergence, where an AI system pursues stuff like power-seeking or resource-acquisition as a necessary means to its end goal, regardless of what that goal is. ↩︎
5. This image illustrates Bostrom’s model of the growth of machine intelligence in the book Superintelligence – AI passes human-level, then at some ‘crossover‘ point reaches the level where the pipeline for further capability optimisation come from the AI itself, then passing the ‘civilisation‘ level where a single AI (superintelligence) system is as capable as all of human intelligence in our civilisation, then finally reaching ‘strong superintelligence‘. ↩︎
6. The possibility of a sudden, lab-driven breakthrough aligns with the “Vulnerable World Hypothesis” (Bostrom, 2019), where technological progress may inadvertently release a “black ball” – a discovery that is easy to make but nearly impossible to defend against without global surveillance or governance. ↩︎