Is AI the Alchemy of Our Era?

The term "alchemy" carries a significant weight, stemming from those who sought to distinguish it from chemistry. We often equate any comparison to alchemy with a dismissal of seriousness, lumping it with pseudosciences like astrology or the aether theories. However, I contend that comparing AI to alchemy can be seen as a critique of the dubious methodologies rather than a rejection of its potential to evolve into a legitimate science.

It's undeniable that today's deep learning techniques emerged with minimal theoretical support, relying heavily on trial and error, abundant data, and evolving heuristics. The lack of rigor, limited peer review, and frequent unstructured experimentation starkly contrast with the methodologies of experimental physics, and indeed resemble ancient alchemy.

However, AI does produce results. Its utility has become increasingly apparent, even if we lack a comprehensive understanding of its optimal applications. The absence of robust theories does not invalidate AI's accomplishments; rather, it’s impressive that we’ve achieved so much without a clear understanding of the underlying processes.

In this regard, AI is not like alchemy, which historically offered little practical benefit and remains so today.

Engineering Science vs. Alchemy

Thomas Krendl Gilbert, a machine ethicist, believes we should not hastily dismiss the alchemical analogy. He argues that AI creators often view their work as magical, relying on arbitrarily constructed neural networks in increasingly powerful computers, with the hope of achieving superintelligence.

Gilbert criticizes AI developers for their lack of attention to the underlying mechanisms of their systems, labeling their work as mystical exploration devoid of genuine insight.

Conversely, not everyone shares this viewpoint. Meta's Yann LeCun recently defended AI against the alchemical label, asserting that dismissing empirical science and engineering as mere alchemy is misguided. He emphasized that both theoretical and practical approaches are valuable and insightful.

LeCun elaborated on his perspective in a talk titled "The Epistemology of Deep Learning," asserting that modern AI, particularly deep learning, is closer to engineering science than alchemy. He referenced a previous NIPS talk that conflated the two and clarified that engineering science involves creating new artifacts through methods that include intuition and experimentation, distinguishing it from the fruitless pursuits of alchemy.

He noted that theoretical understanding typically follows the creation of an artifact, driven by curiosity. We often discover the "what" before unraveling the "how" and "why." Sometimes, experimentation and understanding interplay, generating valuable insights that rely on both aspects.

LeCun's argument is compelling: attempting various methods and yielding significant results is fundamentally different from the alchemical quest for transmuting lead into gold. AI has produced tangible advancements, from machine translation to generative models like ChatGPT.

The capabilities of modern AI are far removed from the failed endeavors of alchemists. While alchemy resulted in no progress, AI continues to advance in diverse domains.

Reevaluating Alchemy's Legacy

Following LeCun's comments, Gilbert offered a nuanced response, acknowledging LeCun's points but suggesting a generational divide regarding the perception of science. He argued that many current AI researchers, particularly those in organizations focused on large language models, are less concerned with theoretical foundations and more focused on practical applications.

Gilbert asserted that the control over the design and discussion of AI has shifted, stating, "The alchemists are now in charge."

In essence, AI resembles alchemy not only in methodology but also in the lack of concern for foundational principles. There exists a parallel hope that the exploration of AI will lead to transformative breakthroughs, akin to finding a philosopher's stone.

Even if Gilbert's observations about the disinterest in scientific rigor among younger AI researchers hold true, this doesn't necessarily undermine AI's potential. While alchemy may have been viewed as pseudoscientific, it ultimately contributed to the emergence of chemistry.

Alchemy and chemistry share a historical relationship; early chemistry emerged from alchemical practices, establishing itself as a legitimate science. In this context, equating AI with alchemy may not be a criticism but rather an indication of its potential to give rise to a new scientific discipline.

Today's AI researchers may be laying the groundwork for a future field that garners respect alongside established disciplines. Even if they seem indifferent to strict scientific protocols, it may not be detrimental. Alchemists lacked a scientific framework, yet chemistry arose from their explorations.

Modern AI researchers might consciously reject a scientific approach, which would be concerning, but they could also be navigating the complexities of experimentation without theoretical constraints. Their endeavors don't prevent others from pursuing theoretical avenues, though the substantial resources required for cutting-edge systems present challenges.

Ilya Sutskever's comments about AI as alchemy reflect a hopeful perspective on experimentation rather than a dismissal of scientific inquiry. He sees their work as creating processes that enable the extraction of insights from data, akin to an alchemical transformation.

All Pursuits Deserve Our Interest

However, I want to take this further. Even if alchemy never evolved into chemistry through rigorous methods, it still held value!

Consider Isaac Newton, widely regarded as one of the greatest physicists. His interests extended to alchemy, a discipline often relegated to pseudoscience. Newton, credited with significant contributions to physics and mathematics, also engaged with alchemical pursuits.

This juxtaposition may seem perplexing. Why would such an eminent figure invest time in a field dismissed as unscientific?

The answer is straightforward: from his perspective, the boundaries between alchemy and physics were not as clear-cut. Driven by curiosity and intellect, he explored various avenues, some yielding profound insights while others, now seen as misguided, contributed to his legacy.

This retrospective classification is only meaningful with the benefit of hindsight. We recognize the utility of physics and mathematics today, while alchemy is viewed as irrelevant. Yet, for Newton, both paths were worth exploring.

His experience underscores a critical lesson: no pursuit is inherently unworthy of interest. This contrasts sharply with the perception of AI researchers, who are often labeled arrogant for anthropomorphizing AI. While this criticism is valid, many AI researchers remain open to exploration, acknowledging the potential for future insights.

This article is a selection from The Algorithmic Bridge, an educational newsletter that connects AI, algorithms, and people. It aims to clarify AI's impact on our lives and equip you with the tools to navigate the future.

You can also support my work and become a Medium member here.