AI Revolution: More Than Just Models

Hillestad observes that while AI exhibits astonishing capabilities, it also frequently fails in many tasks. He cites Sanity's experience, where Claude successfully provided a fantastic pipeline analysis for their operations. However, an attempt to use AI agents for feature planning failed miserably, demonstrating the technology's inherent contradictions.

Both observations are true, highlighting the unpredictable nature of current AI applications. This paradox suggests that despite significant advancements, the technology still has considerable limitations and inconsistencies in real-world deployment.

Industry leaders hold conflicting views on the timeline for Artificial General Intelligence (AGI). Dario Amodei, for instance, suggests that AGI is merely one to three years away, indicating a rapid progression towards human-level AI.

Conversely, Alex Karp, while acknowledging the potential of AI, criticizes current Large Language Models (LLMs) as being "completely and irresponsibly oversold." Hillestad points out these contrasting perspectives underline the uncertainty and debate surrounding AI's immediate future.

Alchemists pursued the philosopher's stone for centuries, believing it could turn base metals into gold and grant eternal life. Despite their ultimate failure to achieve these goals, their efforts led to a more significant, albeit unintended, outcome.

Through their persistent experimentation, alchemists developed crucial skills in distillation, crystallization, heating, cooling, measuring, and combining substances. This practical, experimental knowledge laid the groundwork for the modern field of chemistry, demonstrating that the value often lies in the process of discovery rather than the initial, ambitious objective. Hillestad posits that the pursuit of AGI should be viewed similarly, as a path that will yield foundational knowledge even if the ultimate goal remains elusive.

This viewpoint distinguishes between the impressive performance of AI in specific tasks and a broader, more general form of intelligence.

He suggests that while AI can execute complex functions with high proficiency, it still lacks the comprehensive understanding and adaptive reasoning that even animals exhibit. This stance emphasizes the need to evaluate AI based on its functional utility rather than anthropomorphic comparisons.

Instead, he argues that the journey itself, marked by continuous learning, definition, and innovation, holds greater significance.

He suggests that the true value lies in defining the 'chemistry of intelligence' – the fundamental principles and mechanisms that govern intelligent systems. This ongoing process of discovery is expected to catalyze a broad economic and technological revolution, enabling widespread participation and benefit, regardless of whether AGI is ever fully realized.

It began with a proto-industrial foundation, setting the stage for subsequent developments.

This was followed by the period of "canal mania," characterized by extensive infrastructure building. The next phase involved a "panicking crash," a period of significant economic upheaval. Finally, the era of "railroads and revolution" emerged, solidifying the new industrial order and driving profound societal changes.

The AI revolution can be understood by mapping it against the Industrial Revolution's phases, starting with the eruption of technologies like transformers and the subsequent launch of ChatGPT. This period initiated a widespread frenzy of interest and investment in AI.

Despite widespread predictions, the anticipated crash has not yet occurred, indicating that the current boom continues. The critical deployment phase, where AI technologies are widely integrated into society and industry, still lies ahead.

During the early Industrial Revolution, Britain rapidly constructed canals to address a pressing economic problem. These waterways significantly reduced the cost and time associated with transporting heavy materials such as coal, iron, and stone, which were crucial for nascent industries.

Canals attracted substantial private capital investment and fostered the development of advanced engineering skills in logistics, surveying, and large-scale coordination. While foundational for industrial growth, canals were ultimately transitional technologies that paved the way for more advanced infrastructure like railroads.

Hillestad cautions against the assumption that current dominant AI technologies, such as models, chips, and data centers, will solely define the future of the AI revolution. While essential, these elements represent the present infrastructure, akin to the canals of the Industrial Revolution.

He stresses that true innovation requires looking beyond the immediate frenzy and learning from the trajectory of transitional technologies. Focusing exclusively on what dominates the current moment can obscure the path to more profound and enduring transformations.

The Industrial Revolution was not merely a progression of superior machines; it fundamentally altered the operating system of work. It transitioned labor from dispersed cottage industries to centralized factories, introducing new organizational technologies.

These innovations included the division of labor and standardization, which profoundly reshaped productivity and societal structures. Hillestad suggests that AI should similarly be viewed as an organizational change, offering new ways to structure work and processes rather than simply replacing human tasks.

Current perspectives often envision AI agents as replacements for existing human roles within organizations. This approach involves slotting AI into predefined positions, effectively substituting a person with a machine.

However, Hillestad argues this is likely a flawed abstraction because agents and humans function in fundamentally different ways. Agents do not require the same structural mimicry of a full human workday, suggesting that their integration should involve re-imagining work processes rather than merely replicating existing ones.

He emphasizes the need to define the 'chemistry of intelligence' and develop new content structuring frameworks. This approach would allow agents to augment and transform workflows in ways that leverage their unique capabilities, moving beyond simple automation of human tasks. Hillestad's upcoming presentation in New York promises a new framework for structuring content, aiming to integrate probabilistic outputs with deterministic knowledge systems for more effective AI agent design.

He advocates for a hybrid model that combines probabilistic outputs with deterministic knowledge systems.

This integration aims to leverage the strengths of both methodologies, enhancing AI's capacity for accurate decision-making and reliable performance. A new framework for structuring content, which will be unveiled at a New York conference in September, is expected to provide a practical methodology for achieving this balance.

Hillestad stresses the importance of defining the context of "how work is done" rather than solely focusing on the content of the answer.

He draws a parallel to human organizations, where the process and context of problem-solving are as crucial as the solution itself. This perspective aligns with Bill Campbell's philosophy in 'The Trillion Dollar Coach,' which prioritizes identifying the right person to solve a problem over merely finding an answer, emphasizing the interactive and contextual nature of work.

He indicates a lack of specific insights into this domain, suggesting it is an area that has not received significant attention or development from their side.

Machines demonstrate significant struggles with the concept of time, making it one of the least compatible elements for AI interaction compared to human cognition. Agents often attempt to process time in a human-like manner, yet their underlying mechanical operations function on entirely different principles, leading to discrepancies.

The core challenge lies in 'context engineering,' which is more critical than mere model training in addressing this incompatibility. Hillestad humorously notes that effective interaction with AI requires treating it as if it simultaneously possesses the intellect of a five-year-old and a PhD genius, highlighting the vast disparity between human intuition and machine logic in processing temporal concepts.