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

Machines are expected to develop their own on-chain identities and update their objectives, leading to the emergence of sovereign economic actors. This shift will create a Darwinian market for intelligence, resulting in unpredictable outcomes. Many misconceptions exist in the AI space, particularly regarding the centralization of systems like ChatGPT, which rely on a few companies and create significant infrastructure challenges. Centralized companies face inefficiencies in building and maintaining their infrastructure, while decentralized systems can utilize existing powerful devices for better resource efficiency.

The philosophical concern is that allowing a single entity to control advanced technology grants them excessive power over societal direction and individual representation. Decentralization is crucial for providing control over model structures, training data, and computational resources, especially as machines become integral to daily life and decision-making. The rise of social media serves as a parallel to AI, highlighting the need to view machine learning as essential infrastructure rather than mere products.

Trust is vital for human-machine interactions, with current mechanisms relying on inefficient human social structures. Crypto can introduce programmatic trust mechanisms, enhancing efficiency in machine interactions. The training process for AI models involves transforming raw data into machine learning models through various techniques, including supervised, unsupervised, and reinforcement learning. This evolution aims to balance compute power and human effort, with reinforcement learning minimizing the need for human labeling and increasing machine autonomy.

The conversation reflects on the historical context of deterministic versus probabilistic machines, noting that the current era aligns more closely with human cognitive processes. Founders are advised to disregard most context-specific advice, as it may not suit their unique situations. Early founders should be cautious about external advice, advocating for a selective approach to filtering input. Over-reliance on credentialism can hinder progress, particularly in deep tech, and unconventional methods can yield faster, more meaningful results.

In discussing productivity, managing energy over time is emphasized, recommending alignment of personal goals with work and incorporating breaks to boost productivity. Curiosity about technological progress, especially in AI, is highlighted as essential. The assumption that current methods are the only possibilities should be challenged, advocating for a mindset that embraces the temporary nature of technological states. Concerns about using years of experience as a hiring metric are acknowledged, noting that while it may be necessary for broad searches, it can lead to rigid thinking in specific fields. The conversation concludes with a reference to optimizing for energy rather than time.