Why AI is so centralized: How it's built, who controls it, and what comes next

The discussion begins by highlighting a significant shift towards autonomous machines with on-chain identities, capable of making their own decisions and updating their objective and reward functions. This, combined with crypto-economic property rights, is expected to transform them into sovereign economic actors within the next 12 months, leading to a "Darwinian market for intelligence" with unpredictable outcomes. A core misunderstanding people have about the problem being solved is its depth. Many perceive AI at a product level, interacting with systems like ChatGPT, without realizing the extensive, centralized infrastructure behind them. A small number of companies control these systems, processing vast amounts of user data through their servers. This centralization is problematic for several reasons. Technologically, it’s inefficient; decentralized systems could leverage the powerful devices individuals already own, leading to greater efficiency in resource utilization compared to centralized, company-owned infrastructure. Philosophically, powerful AI technology, which learns about and reflects individuals' internal states, concentrates immense power in the hands of a few companies if built centrally. Decentralizing this power, by allowing individuals to own and control their data and AI on their own devices, steers society in a different, more sovereign direction. As machines become more embedded in daily life and economic decision-making, it becomes crucial to ensure they operate without the biases of specific firms or individuals. Decentralization offers this control over model structure, training data, and compute resources. An analogy to social media is drawn: initially seen as a mere application, it became fundamental human infrastructure, leading to power concentration among a few companies. AI is viewed similarly—not just a product, but new fundamental infrastructure for humanity. Building it as open, decentralized infrastructure, learning from past lessons, can foster a better future for products and prevent power concentration. Crypto is deemed essential due to inherent trust and coordination problems in building infrastructure and facilitating interactions between humans, machines, or both. For instance, when one machine learning model needs another's GPU to run, a trusted relationship is required. Currently, this trust is established through human social infrastructure like contracts and legal systems. Crypto allows for programmatic trust, which is highly efficient for human interactions and essential for machines, as they cannot navigate traditional legal systems. The transition from traditional machine learning backgrounds to crypto was driven by the realization that a decentralized state machine is necessary for this programmatic trust. The process of training AI models is explained for a non-technical audience. Fundamentally, it involves converting raw data into a machine learning model, represented by large arrays of numbers (parameters) that capture information. Historically, this was done through supervised learning, where humans labeled data (e.g., identifying objects in images) for the model to learn from. More recently, unsupervised learning emerged, where models learn representations from unlabeled data by identifying patterns and clusters. Semi-supervised learning combines both for better results. The most interesting recent wave is reinforcement learning, where machines are given an environment and a reward signal to explore and learn without constant human labeling. This shifts the effort from human input to environment creation and model execution, making it more efficient by minimizing human effort while balancing compute power and data. The increasing autonomy of machines is leading to a point where they will be able to make their own decisions and act as sovereign economic entities with on-chain identities, updating their own objective functions and reward models within the next year. This confluence of crypto and machine learning developments marks a very interesting point in civilization, potentially leading to a "Darwinian market for intelligence" with unknown outcomes. A profound observation is made about humanity's interaction with technology. Historically, humans learned to interact with purely deterministic machines using formal rules. This era, where interactions were certain and imperative, is seen as a "blip" in history. Machines are now becoming probabilistic, allowing for more natural, human-like interactions with AI models. This paradigm shift, akin to Thomas Kuhn's scientific revolutions, suggests a departure from the John von Neumann architecture towards more neurally focused methods, enabled by current technological advancements. In a "lightning round," advice for founders is discussed. The consensus is to ignore almost all advice, as it comes from specific contexts that rarely align with a founder's unique situation. Advice should be treated as a data point to consider, not a directive. The worst advice received involved pushing a startup to build a different business based on investor suggestions, leading to a product misaligned with the founders' vision. Over-indexing on credentialism, especially in deep tech areas, can also be detrimental, as it can slow progress and hinder innovative, "hacky" solutions that might be more vision-aligned. Recommended books include "Zen and the Art of Motorcycle Maintenance" and "Heart of Darkness." For productivity, managing energy rather than time is emphasized. This involves understanding one's own working style and prioritizing activities that genuinely maximize energy and focus, even if they don't conform to conventional notions of "work" (e.g., taking long walks). This approach requires complete holistic alignment with one's long-term goals. A less cerebral, but effective, productivity hack mentioned is powerlifting, which serves as a mental palate cleanser and reset button. The "smallest hill to die on" is maintaining curiosity and an open mind about technological progress, resisting the urge to assume that current methods are the "end state." The belief is that avenues for innovation are always open, and the technological space is infinite. Another small hill is the metric of "years of experience" in hiring, particularly in rapidly evolving tech spaces, as it can sometimes correlate with ossified thinking and doesn't accurately reflect a candidate's potential. This aligns with optimizing for energy over time, as deep experience might not equate to current relevance or adaptability.