Genius on Demand: The Value of Transformative Artificial Intelligence

Source: Ajay K. Agrawal, Joshua S. Gans and Avi Goldfarb; NBER Working Paper 34316 · 2025-10

This paper examines how the emergence of transformative AI systems providing ``genius on demand" would affect knowledge worker allocation and labour market outcomes. We develop a simple model distinguishing between routine knowledge workers, who can only apply existing knowledge with some uncertainty, and genius workers, who create new knowledge at a cost increasing with distance from a known point. When genius capacity is scarce, we find it should be allocated primarily to questions at domain boundaries rather than at midpoints between known answers. The introduction of AI geniuses fundamentally transforms this allocation. In the short run, human geniuses specialise in questions that are furthest from existing knowledge, where their comparative advantage over AI is greatest. In the long run, routine workers may be completely displaced if AI efficiency approaches human genius efficiency.

Introduction. Recent advances in artificial intelligence (AI) have raised the prospect of machines with extraordinary cognitive capabilities—systems that could fundamentally reshape labour markets by providing what Amodei (2024) describes as a “country of geniuses in a data center.” Such systems, capable of applying and potentially creating knowledge across various domains, raise a fundamental economic question: How would labour markets absorb such an influx of cognitive capacity, and what would be the consequences for different types of knowledge workers? In order to analyse this question, it is important to distinguish geniuses from other knowledge workers. In this regard, we are inspired by this story concerning the approach taken by John von Neumann.

Then there is the famous fly puzzle. Two bicyclists start twenty miles apart and head toward each other, each going at a steady rate of 10 m.p.h. At the same time a fly that travels at a steady 15 m.p.h. starts from the front wheel of the southbound bicycle and flies to the front wheel of the northbound one, then turns around and flies to the front wheel of the southbound one again, and continues in this manner till he is crushed between the two front wheels. Question: what total distance did the fly cover? The slow way to find the answer is to calculate what distance the fly covers on the first, northbound, leg of the trip, then on the second, southbound, leg, then on the third, etc., etc., and, finally, to sum the infinite series so obtained. The quick way is to observe that the bicycles meet exactly one hour after their start, so that the fly had just an hour for his travels; the answer must therefore be 15 miles. When the question was put to von Neumann, he solved it in an instant, and thereby disappointed the questioner: “Oh, you must have heard the trick before!” “What trick?” asked von Neumann; “all I did was sum the infinite series.” Halmos (1973), 386-287.

While normal mathematicians would use tricks to simplify a problem, von Neumann could work from first principles. In effect, he reproduced the knowledge required to answer the question. Our notion of a genius in this paper is similar. We focus on how knowledge is both created and utilised by knowledge workers (including AIs) to impact decision-making. Routine workers rely on existing knowledge and draw inferences from that to answer questions. Geniuses (human or AI) generate new but ephemeral knowledge to answer questions put to them (working from first principles). Thus, routine workers and geniuses differ fundamentally in how they approach problems. The former can easily guess answers that are imperfectly accurate, while the latter can deliver accurate answers at a cost. Moreover, to align with Amodei’s conception, we assume that AI geniuses approach problems in a similar manner to their human counterparts. Using this distinction, we model how organisations allocate different knowledge worker types to questions that arise. The key idea is that when questions arise, it is important that someone knows what type of knowledge worker would best address that question. As knowledge workers use two distinctive approaches and also differ in terms of their availability, we view the task of “knowing who to ask” as a key learned task – from experience – and hence, as an equilibrium outcome.1 In our model, this function is represented by a manager who routes questions to the appropriate type of worker, taking into account both their absolute advantage and, because specifically human geniuses are scarce, their comparative advantage.2 We characterise the optimal routing rule that maximises the total value of answers provided. Importantly, our framework shows how routine and genius workers, approaching questions differently, have value generated with distinct curvature implications across the space of possible questions. This drives their allocation to question types. Following this, using this same setup, we examine the short and long-run impacts of an influx of AI geniuses. Recall that the task of knowing who to allocate a question to is a learned one, and so when AI geniuses become abundant, it takes a while for the manager to reoptimise routing. Thus, the short run is characterised by the manager not altering their pre-AI allocation rule and allocating AI geniuses to genius-type questions along with human geniuses. We show that this pushes human geniuses to tackle more novel questions. In the long run, the AI geniuses are shown to displace routine work as the manager adjusts their routing rule accordingly. The short and long-run framing allows us to consider the path by which AI geniuses might impact knowledge work. The paper builds on the model of Carnehl and Schneider (2025) that provides a microfoundation for how knowledge is used in decision-making.

Method. To analyse how transformative AI might reshape knowledge worker allocation, we require a framework that precisely captures the relationship between knowledge, uncertainty, and value creation. Our approach builds on a microfounded model of knowledge that explicitly connects the structure of what is known to the value of what can be produced. This microfoundation is essential for understanding knowledge work, where value derives not from physical output but from the quality of decisions made under uncertainty.

This formulation captures a fundamental characteristic of knowledge: uncertainty (variance) is lowest at the known point and increases linearly with distance from the known point. This pattern of uncertainty mirrors real-world scenarios, where extrapolating from established knowledge becomes increasingly difficult as one moves further from what is already known.

This value function has important properties: it approaches 1 (maximum value) as uncertainty approaches 0 (perfect knowledge) and decreases linearly toward 0 as uncertainty approaches the threshold q.

This regional structure (see Figure 1) creates interesting patterns of comparative advantage between different worker types, as we will explore in the next section.

Our economy features three distinct types of economic agents in the knowledge production and application process. First, routine workers possess specialised skills to apply existing knowledge to questions that fall within established frameworks. Second, genius workers have the unique ability to create new knowledge, enabling them to tackle questions that lie beyond the frontier of what is currently understood. Finally, managers serve as allocators who determine which questions should be routed to routine versus genius workers, maximising the overall value of knowledge work subject to capacity constraints. In this section, we elaborate on the capabilities, value functions, and real-world interpretations of each worker type.

Routine workers represent intellectual labour that applies existing knowledge to form conjectures about questions. These workers can be thought of as skilled practitioners who operate within established paradigms—they extrapolate from what is known but do not create fundamentally new knowledge. Examples include residents in medicine applying standard differential diagnoses, paralegals applying precedent to routine cases, or software developers implementing known algorithms for familiar problem types.

Discussion. The shape of the absolute-advantage function implies two regimes; see Figure 2. Thus, when η ≤1/(2q2), scarce genius capacity is directed first to the edges of the routine domain, x0 ± q, where the incremental value over routine is greatest. When η > 1/(2q2), the peaks are interior: initial genius allocations go to narrow bands centred at x0 ± d⋆, where the advantage over routine is locally maximised. In both cases, the remainder of Region 2 is served by routine workers, while Regions 1 and 3 are unanswered until genius capacity expands. As LG rises, these initial bands widen in the manner illustrated in Figure 2. In the edge-first regime, the expansion proceeds outward into Regions 1 and 3 and inward toward x0, with the centre left last because geniuses’ relative advantage is smallest at d = 0. In the interior-first regime, the bands around x0 ± d⋆broaden symmetrically; once their marginal advantage falls to the edge level 1 −ηq2, additional capacity is placed at x0 ± q and then extended further outward. At intermediate LG, the allocation takes the form of two mirrorimage strips flanking x0. With further increases, these strips grow toward one another, shrinking the central routine region; in the limit, abundant genius capacity covers the entire domain.

The emergence of transformative artificial intelligence (AI) systems capable of generating creative solutions and novel insights fundamentally alters the landscape of knowledge work allocation. As described by Amodei (2024), these systems can potentially provide “a country of geniuses in a data center,” dramatically expanding the supply of genius-level cognitive capacity. In this section, we extend our model to incorporate AI geniuses alongside human geniuses and routine workers, analysing how optimal allocation patterns and income distribution change in both the short and long run.

We model AI geniuses as having fundamentally similar capabilities to human geniuses—they can create new knowledge for any question in the domain, including those where uncertainty exceeds the threshold for routine workers. However, we allow for potential differences in efficiency by introducing a distinct cost parameter.

This value is depicted in Figure 3. The assumption that ηAI ≥η reflects the possibility that AI systems, while capable of genius-level work, may require more resources or exhibit lower efficiency than top human geniuses when creating novel knowledge far from established frameworks. This formulation creates a natural hierarchy: human geniuses have a comparative advantage over AI geniuses, who in turn have a comparative advantage over routine workers in regions where uncertainty is high.5 Crucially, we assume that human geniuses remain scarce (LG < 1) but AI geniuses are available in effectively unlimited supply. This assumption reflects the scenario described by Amodei (2024), where millions of instances of AI models can be deployed once developed.

In the short run, we assume that routine workers continue to operate within the thresholds established prior to the introduction of AI geniuses. Specifically, if routine workers were previously allocated to a set of questions A∗ R, this allocation remains unchanged immediately following the introduction of AI.

Conclusion. This paper has examined how transformative AI systems providing “genius on demand” would affect knowledge worker allocation and labour market outcomes. Our analysis reveals that when genius capacity is scarce, optimal allocation directs genius workers to questions where their absolute advantage over routine workers is highest — notably at domain boundaries where uncertainty is too high for routine workers. The introduction of AI geniuses transforms this allocation pattern, with human geniuses specialising in questions furthest from existing knowledge, where their comparative advantage over AI is greatest. These findings contribute to our understanding of knowledge work allocation, but have modelled AI geniuses as independent agents that replace human workers. This approach, while analytically tractable, may not capture the most promising potential of AI technologies. There may be categories of geniuses, such as artisans, artists or persuasive geniuses. It is not clear if the Amodei conception might cover those categories as well or if AI-geniuses might also be more efficient at doing what we have modelled as routine work in this paper (see, for example, Gans (2025b) for an example of this type of AI). If there were AI agents that undertook knowledge inference, then it could easily be the case that in the short run, routine workers would be impacted, with geniuses perhaps being moved to more novel tasks in the long run. The point here is that being specific regarding how human and AI workers engage with knowledge will impact predictions regarding transformative AI. Future research should examine how AI could function as tools in conjunction with knowledge workers rather than as automated independent agents (Agrawal et al., 2025). Such collaborative human-AI systems might leverage complementary strengths: human contextual understanding, ethical judgment, and creative insights alongside AI computational power and pattern recognition capabilities. Models of such complementary human-AI workflows would enrich our understanding of knowledge work evolution beyond displacement narratives.