The Architecture of Innovation
Joseph's Berliner's Analysis of Innovation in the Soviet Industries
In recent months, I have been working toward developing a broader vantage point for studying innovation and technological advancement. Living, studying, and working in the United States, I noticed that the mainstream discourse is highly focused on domestic developments, often creating a form of informational isolation from broader global dynamics. I began examining other countries more systematically. One obvious choice was China because of its rapid economic growth, which suggests that, whatever criticisms may exist, there must have been technological and industrial strategies that succeeded. However, much of the U.S. narrative around China tends to focus on risk containment, making objective analysis difficult.
In parallel, I explored the historical evolution of a few industries more deeply, especially the oil and gas sector which I know the most about. This led to a piece I wrote a few weeks ago about Nikolai Baibakov, a key figure in Soviet energy management. From there, my curiosity expanded to other industries within the Soviet system, such as the steel industry. Looking for more unbiased assessments of the Soviet industries brought me to the work of Joseph Berliner, particularly his 1976 study, The Innovation Decision in Soviet Industry. Berliner provided a disciplined structural analysis of how technological innovation occurs, or fails to occur, within large, centralized, complex systems.
Inspired by Berliner's approach, I have loosely applied his framework, focused on prices, decision rules, incentives, and organizational structure, to examine the current state of innovation in artificial intelligence (AI) in the United States. The goal is to apply a structured perspective to a field undergoing rapid and consequential change. I understand that a main theme in Berliner’s book is to contrast the capitalist vs. communist ecosystem for driving innovation, and I understand that we’re five decades past the publication of his book, but when I was reading the book I noticed that many of the elements in his comparative framework are useful for a diagnostics approach to any industry anywhere in the world.
Defining AI Technology and Innovation
A technology is the collection of artifacts and methods by which goods and services are produced. In the case of artificial intelligence (AI), the technology consists of computational models, algorithms, data infrastructures, and human-machine procedures that together perform tasks previously requiring human cognitive effort. Examples include natural language processing, visual recognition, strategic planning, optimization, automation, and generative content creation.
Technological innovation, however, must be distinguished from pure technical discovery. Innovation occurs only when a new capability is successfully incorporated into actual production systems. An algorithm or model that remains in a research lab or internal demonstration, without reshaping processes, services, or outputs, does not constitute innovation. Only the effective adoption of new capabilities into working systems marks genuine progress.
For the purpose of this diagnostic note, AI technology will be considered as the set of artifacts (models, tools, platforms) deployed in practical applications. Innovation will be evaluated not by the technical sophistication of the tools themselves but by their integration into economic, social, and organizational life.
What Problem(s) Are We Trying to Solve With AI?
AI based on my observation is mainly deployed to address the persistent economic challenge of information processing such as the need to make faster, better, and more scalable decisions in increasingly complex environments.
In economic systems, three structural obstacles historically limit performance:
Decision complexity, as systems grow larger, making optimal or even satisfactory decisions becomes increasingly difficult.
Resource misallocation, as poor decision-making leads to wasteful use of labor, capital, and materials.
Incentive misalignment, as actors often have private incentives that diverge from societal optimal outcomes.
AI promises, in theory, to mitigate these frictions by automating information synthesis, providing predictive insights at scale, and enabling new forms of adaptive control over complex systems. Whether AI delivers on this promise depends not just on technical capability but on the structural coherence of the broader environment in which it is deployed.
Innovation outcomes are shaped by four primary structural variables. To formulate this, I create a simplified way of asking a questions in each variable and assess the risk associated with each of these answer.
1. Pricing Signals
Do AI applications operate in environments where prices or equivalent signals accurately reflect true costs and societal value?
Risk: if market valuations emphasize speed and model size but ignore trustworthiness, security, or energy cost, AI systems will optimize for short-term metrics rather than sustainable value.
2. Decision Rules
What formal or informal rules guide choices among competing AI initiatives?
Risk: if decision-making frameworks prioritize "first to market" or "scale at all costs," misallocations and societal harm may follow
3. Incentive Alignment
Are innovators, deployers, and operators of AI systems incentivized to pursue outcomes that align with broad societal interests?
Risk: incentives that reward speed and growth over safety or sustainability will produce cracks in the system.
4. Organizational Architecture
Are organizations structured to facilitate the diffusion, adaptation, and iterative improvement of AI innovations?
Risk: if innovation becomes concentrated in isolated "islands" within large firms or government agencies, broader economic dynamism will suffer.
Innovation Spillover and Structural Barriers
In studying the history of technological innovation, one recurring challenge is making sure that advances in one sector or organizational island translate into broader economic benefits. Within the Soviet Union, despite extraordinary achievements in fields such as space exploration and military technology, spillover into the civilian industrial economy was extremely limited. According to Berliner, several barriers, observed in historical cases, impeded effective knowledge diffusion:
Organizational Secrecy: military and space sectors operated under strict information controls, preventing easy transfer of techniques and insights to civilian industries.
Departmental Jealousy: ministries and organizations prioritized their own prestige and resources, discouraging open sharing.
Rigid Incentive Structures: civilian managers had little incentive to adopt externally developed innovations unless directly ordered, and even then faced career risks for failure.
Lack of Mobility: engineers and scientists were often institutionally siloed, limiting the organic spread of expertise.
In the context of AI today, the analogy to military technology is not perfect, however, the broader structural lesson remains relevant, because as AI capabilities concentrate within a few dominant technology firms, there is a risk that much of the broader economy becomes reliant on a small set of frontier innovations without developing internal capacities for adaptation and experimentation.
If the creation of new AI tools remains the domain of a limited group, over time, a gap could grow and a few companies will keep advancing AI technology, while many other sectors may be left behind, using basic tools without fully understanding or benefiting from the latest developments. Organizations outside the core AI developers may either become basic users of commoditized products or, worse, struggle to understand how to integrate and customize advanced AI capabilities into their specific operational needs. Effective socialization of knowledge through open standards, collaborative frameworks, workforce mobility, and incentives for internal innovation is important. Without structural attention to spillover mechanisms, the potential for AI to help broad sectors of the economy will diminish.
After reading Berliner's work on innovation and structural barriers in the Soviet Union, I became curious whether similar patterns might exist elsewhere. One natural case to examine was Iran, because I grew up there, and hear the news related to Iran more frequently given the widespread reporting on its significant progress in nuclear technology and long-range missile development over the past decade. Iran seems to have achieved real technical advancements under difficult conditions, including sanctions and international isolation. The nuclear sector, particularly uranium enrichment and reactor technology, and the missile sector, especially solid-fuel and guidance systems, demonstrate genuine scientific and engineering capabilities (regardless of their intent or destructive regional policy). What I wonder is if these advances spilled over into broader sectors of the Iranian economy? The evidence suggests that, much like the Soviet Union, spillover has been extremely limited. Several structural barriers explain this outcome. Organizational separation, because Iran’s military-industrial sectors, including its nuclear program, are heavily siloed under specialized agencies like the IRGC and the Atomic Energy Organization of Iran. Civilian industries operate in parallel but largely disconnected systems. Secrecy and security constraints, because the national security concerns enforce strict information compartmentalization, minimizing the diffusion of technical knowledge outside defense and strategic sectors. Incentive structures within strategic sectors revolve around national prestige and regime security. And finally, engineers and scientists trained in nuclear and missile programs are rarely transferred to civilian R&D roles, limiting talent mobility.
The consequences are visible across multiple civilian industries:
Manufacturing remains reliant on outdated technologies, with no major leap forward despite advanced aerospace capabilities.
Civil aviation continues to depend on retrofitting aging fleets, without a viable domestic aircraft industry.
Energy technologies have not seen a significant indigenous innovation boom, despite clear expertise in complex engineering domains.
Biotechnology and healthcare show some isolated successes, but these emerged separately rather than as direct spillovers from nuclear or missile programs.
Iran today exhibits a pattern of "innovation islands" similar to what Berliner identified in the Soviet Union where certain strategic sectors achieve technical excellence, but the broader economy remains technologically stagnant, unable to capitalize fully on those achievements.
Key Risks Facing the Current AI Innovation Landscape
Over-centralization: dominance by a few firms risks bottlenecking innovation and misallocating talent.
Incentive distortions: commercial incentives currently favor rapid scaling of models over long-term reliability and societal integration.
Knowledge siloing: proprietary control over data and models threatens to constrain the ecosystem's adaptive capacity.
Resource exhaustion: energy and human cognitive costs of frontier AI development may not be properly priced or managed.
AI represents an extraordinary technological potential to address critical economic problems, but technical advancement alone does not guarantee beneficial outcomes. The structural architecture in which AI systems are created, evaluated, and deployed will determine whether innovation enhances or erodes societal resilience.
Must read sources:
Berliner, Joseph S. The Innovation Decision in Soviet Industry. Cambridge, MA: MIT Press, 1976.
Chertok, Boris. Rockets and People, Volume II: Creating a Rocket Industry. Edited by Asif A. Siddiqi, translated by Deborah Weston. Washington, D.C.: NASA History Division, 2006.