For those tracking the advancements in AI, the concept of specialization is likely a familiar one. At Dharma AI, we see it as a foundational principle for building effective AI systems, influencing everything from cost and performance to reliability and even system sovereignty. A recent paper by Goldfeder, Wyder, LeCun, and Shwartz-Ziv (2026) offers a particularly rigorous and compelling case for this perspective.
This article delves into the insights from their 2026 work, titled “AI Must Embrace Specialization via Superhuman Adaptable Intelligence.” The paper’s strength lies in its “convergence case,” which draws parallels across diverse fields like optimization theory, evolutionary biology, organizational economics, and machine learning. While the original paper provides the intellectual backbone, the framing, organization, and synthesis you’ll read here are Dharma AI’s interpretation and elaboration.
It’s natural to assume that as AI systems become more capable, they should also become more general-purpose. We often expect that increased resources, refined methods, and expanded training would lead to systems confidently tackling a wider array of tasks. This intuitive idea suggests that greater capability and broader applicability should go hand-in-hand.
However, the real-world pattern tells a different story. The AI systems that achieve truly groundbreaking results in specific domains are almost always those with a narrow, focused design. Think about the revolution in protein structure prediction; it came from a system meticulously engineered for that single scientific challenge. A closer look at AI’s historical milestones reveals intense domain targeting, not expansive generality, as the key to success.
This recurring pattern is remarkable. It appears consistently across different domains, over many decades, and even within vastly different architectural choices. Such a consistent trend strongly suggests a common underlying cause—one that likely originates beyond the confines of AI research itself.
The Universal Rule: Fit Trumps Generality
In 1997, Wolpert and Macready delivered a mathematical proof that, surprisingly, often goes overlooked in AI discussions. They demonstrated that no single, general-purpose optimization algorithm can outperform all others across all possible problems. Mathematically, if you average performance across every conceivable problem a learner might encounter, every algorithm performs equally well—and equally poorly.
What this means is that an algorithm gaining an advantage on one set of problems must necessarily concede ground on others. Performance is redistributed, not magically multiplied. The practical takeaway is profound: as Goldfeder et al. (2026) put it, “an algorithm wins by being a good fit for the target problem.” This theorem doesn’t rule out generality, but it emphatically states that generality isn’t a performance advantage.
The imperative for concentration—trading breadth for a perfect fit—becomes even sharper when we introduce finite resources. Any real-world system operates under limits: finite compute power, finite data, and finite development time. Given these constraints, an approach that dedicates resources to mastering a finite set of tasks will inevitably outperform one that scatters those same resources across an unlimited range. It’s simple arithmetic: as the number of tasks expands endlessly, the resources available per task dwindle to almost nothing. In the context of finite resources, universal coverage and meaningful performance are in direct conflict.
The theorem doesn’t suggest that generality is inherently bad. Instead, it offers a more practical insight: as the paper states, “universal generality is a theoretical concept, but in practical terms it is a myth.” What ultimately thrives under real-world constraints isn’t the system that attempts to do everything, but rather the system that perfectly fits its specific target problem.
Lessons from Biology and Markets
Intriguingly, two other complex domains independently arrived at the same fundamental prediction long before optimization theory articulated it. Evolutionary biology, for instance, perfectly illustrates this principle: every performance gain in one ecological niche comes at a cost elsewhere. A biological generalist possesses traits suited to many environments but isn’t optimal for any single one; its competence is spread too thin to dominate under specific conditions.
Evolutionary selection strongly favors designs perfectly matched to local conditions over those optimized for uniform coverage. Organisms that successfully reproduce are not necessarily the most generally capable, but the most specifically adapted. Over vast evolutionary timescales, this leads not to the dominance of generalists, but to specialists meticulously filling countless ecological niches. As the paper eloquently puts it: “Specialization is not an accident of biology; it is a predictable consequence of limited resources, competing objectives, and environments that reward performance on a small subset of evolutionarily relevant challenges.”
Competitive markets, though operating through entirely different mechanisms, mirror this dynamic. Organizations and strategies that fail to meet performance thresholds are weeded out, not through extinction, but through market exit, defunding, or replacement by more effective alternatives. Competition acts as a powerful selection mechanism, amplifying successful strategies and eliminating ineffective ones. While there’s no biological inheritance or mutation, the structural pressure remains the same: finite resources, stringent performance requirements, and the systematic removal of entities too broadly distributed to excel where it truly matters. When performance standards are clear and consistent, concentrated capacity consistently outcompetes diffuse capacity.
It’s striking that evolution and markets, with their distinct mechanisms, timescales, and units of selection, both yield the same outcome under resource pressure: a premium on fit over breadth. When a third domain, like optimization theory, independently arrives at the same conclusion, the pattern transcends a mere theorem. It begins to look like a universal truth about how constrained systems behave.
Specialization’s Recurrence in Machine Learning
Within machine learning itself, the same pattern has emerged—not through theoretical derivation, but from the hard-won experience of building and refining systems. A clear manifestation is negative transfer, a measurable performance degradation that occurs when a system trained on multiple tasks suffers because those tasks compete rather than cooperate (Ruder, 2017).
While sharing structure between tasks can be beneficial, tasks that compete for representational capacity or impose conflicting gradients during training will see individual performance fall below what a dedicated system could achieve. The perceived gain from breadth ironically becomes a cost to depth. This is a well-documented consequence of distributing finite capacity across tasks that pull in different directions. A specialist system, facing no such internal competition, avoids this performance penalty.
The architecture of cutting-edge models, particularly Mixture-of-Experts (MoE) systems, offers another form of evidence. These systems achieve their impressive breadth not through uniform generality across all parameters, but by routing each input to a specialized subset of the network, activating different “experts” for different tasks. The paper’s authors interpret this as a structural concession: a system designed for generality ultimately achieves its results by internally recovering specialization. While MoE architectures were primarily designed for computational efficiency, their success hints at the inherent limits of true generality, suggesting that even the most general-purpose systems reach peak performance by doing internally what specialist systems achieve by design.
Historically, perhaps the clearest example is AlphaFold. It achieved a monumental leap in protein structure prediction by meticulously targeting that specific task with a bespoke architecture and tailored training choices (Jumper et al., 2021). Its gains stemmed from a narrower, sharper focus, not from broader coverage. AlphaFold serves as an archetypal case, illustrating how intense domain targeting—rather than expansive general competence—has frequently characterized AI’s most significant milestones.
Three distinct domains. Three different underlying mechanisms. Yet, they all arrive at the same profound finding.
Why Scaling Doesn’t Undermine Specialization
Any discussion about AI must address Sutton’s “Bitter Lesson,” which observes that methods relying on domain knowledge are consistently outperformed by methods that simply scale computation (Sutton, 2019). At first glance, this might seem to contradict the case for specialization: if scale leads to generality, perhaps specialization is just a temporary heuristic for resource-constrained environments.
However, this objection rests on a critical conflation between two distinct concepts: domain knowledge and domain specialization. Domain knowledge refers to hand-coded features, engineered priors, or explicit rules designed to imbue a system with insight into a particular area. The Bitter Lesson rightly targets and refutes this; systems embedding explicit domain knowledge have indeed been consistently surpassed as computational scale increases.
Domain specialization, conversely, is a strategic decision about scope. It’s the choice to direct a system’s resources, architecture, and training toward a bounded set of tasks, rather than distributing them broadly. This isn’t about encoding explicit knowledge about a domain; it’s about defining the system’s operational focus. The paper draws this crucial distinction with precision:
- “The diminishing usefulness of domain knowledge is distinct from the usefulness of domain specialization. As scaling progresses, we will need to know less about proteins to build a system that does protein folding; however, such a system still benefits from focusing specifically on proteins.” (Goldfeder et al., 2026)
Scaling undoubtedly changes what systems can learn effectively from data. But it doesn’t alter the fundamental principle that concentrating resources on a finite task set will outperform dispersing them across an unlimited range. The Bitter Lesson and the specialization argument operate on different dimensions: one dictates how knowledge should be acquired, while the other defines what a system should be aimed at. Both can be simultaneously true.
Scaling refines the mechanisms by which systems learn; it does not dissolve the fundamental constraint that makes fit more valuable than sheer breadth. Across four distinct analytical traditions—optimization theory, evolutionary biology, competitive markets, and machine learning—the same powerful pattern emerges through divergent paths. This isn’t a mere coincidence demanding explanation; it is compelling evidence.
Ultimately, when finite resources encounter selection pressure—whether in an optimization problem, an ecosystem, a market, or a machine learning training run—fit consistently beats breadth. The specific mechanisms, timescales, and units of selection may vary significantly. Yet, the underlying structural dynamic remains identical, consistently yielding the same result. The theorem doesn’t dictate this pattern in biology, nor does biology impose it on markets, and neither causes it in machine learning. Instead, they all independently confront the same foundational constraint: high performance under scarcity invariably demands concentration.
Source: Hugging Face Blog
