AI and the Dynamics of Change: Understanding Paradigm Shifts and Drifts

Through discussions with AI experts, including Gregory Renard in my previous post, it's clear that grasping the practical implications of a "paradigm shift" is challenging. These shifts are not instantaneous events; they unfold gradually. Furthermore, the very definition of a paradigm evolves as it permeates various aspects of our environment. This gradual evolution is what I termed "paradigm drift" in my book.

Alan Turing presciently anticipated the slow percolation of the AI paradigm shift in his seminal 1950 paper, "Computing Machinery and Intelligence." Although the term AI appeared only in 1956, Turing posed the fundamental question of AI, "Can machines think?" Despite exploring numerous objections, Turing confidently conjectured that by the end of the 20th century, the prevailing view would shift, and "one will be able to speak of machines thinking without expecting to be contradicted."

Turing's perspective aligns with the experiences of scientists like Max Planck and the observations of Thomas Kuhn in The Structure of Scientific Revolutions. Paradigm shifts do not cause sudden revolutions but gradual transformations, eventually leading to widespread adoption. This percolation of a paradigm can be likened to Edward Sapir's concept of "language drift," where subtle changes in terminology and expression accumulate over time. Just as languages evolve incrementally, so too do our understanding and application of fundamental concepts, including AI.

Artificial intelligence (AI) is already woven into the fabric of daily life in the United States, often operating behind the scenes. Despite this integration, a gap exists between a de facto acceptance and public anxieties about its increasing real or future pervasiveness. A 2023 Pew Research Center study revealed a cautious public sentiment: only 15% expressed more excitement than concern about AI's growing presence, compared to 38% who were more concerned and 46% who felt a mix of both. While these figures may have evolved, apprehension about AI remains. This complexity explains why innovators, particularly in fields like AI that touch upon behavioral and emotional aspects of life, face significant hurdles. Like any major innovation, despite its prominence in public discourse, AI's societal integration will be a gradual process whose pace will be influenced by multiple factors, from the advancement of AI technologies to the specifics of their implementation at any given moment.

AI Technologies: While artificial intelligence has made remarkable strides, its implementation at scale is still not well understood. AI models are in constant evolution, demonstrating impressive results in complex domains like advanced mathematics, graduate-level science, abstract reasoning, and programming. However, the pursuit of ever-greater capabilities continues fueling research in algorithmic advancements and the development of powerful hardware accelerators, such as GPUs (Graphics Processing Units), NPUs (Neural Processing Units), FPGAs (Field-Programmable Gate Arrays), and ASICs (Application-Specific Integrated Circuits). Key challenges remain in seamlessly integrating these diverse accelerators, optimizing their performance, and ensuring energy efficiency. Looking ahead, breakthroughs in neuromorphic and quantum computing hold the potential to further revolutionize large language models (LLMs) and other AI applications. Despite the already fascinating capabilities of current AI and the likelihood of continued scaling for years to come, today's AI technologies represent a significant undertaking, suggesting that the paradigm shift they embody could usher in entirely new computing paradigms. We are, in essence, only at the beginning of a drift cycle, which should also encourage everybody to get on board now.

Second: Implementation—Cautions and Complexities: Embracing AI doesn't necessitate blind acceptance. The integration of AI will inevitably encounter obstacles and unexpected turns. While many applications offer convenience and tangible benefits—from drafting emails and transforming ideas into videos to potentially revolutionizing transportation and medical diagnostics – critical challenges remain.

Even enthusiastic users encounter limitations that highlight the "black box" nature of AI or feel that it’s an opaque veil that stands between them and the world. For example, if you ask thorny questions to an LLM, you may get somewhat muddled answers and if you try to get clarifications, you may be given non-committal statements such as “I can't help with responses on elections and political figures right now. I'm trained to be as accurate as possible but I can make mistakes sometimes,” or “If you come across claims or images suggesting otherwise, it’s a good idea to verify them using trusted news outlets and fact-checking organizations.” As a user, you may get the feeling that you’re being taken for a ride. Yet, as anecdotal as such experience may be, any normal user is entitled to wonder how the model is trained, fine-tuned or monitored.

This lack of transparency underscores the critical issue of trustworthiness and safety in general AI. As highlighted in the January 2025 International AI Safety Report, led by Fields medalist Yoshua Bengio, “The inner workings of these models are largely inscrutable, including to the model developers. Model explanation and ‘interpretability’ techniques can improve researchers’ and developers’ understanding of how general-purpose AI models operate, but, despite recent progress, this research remains nascent.”

This must-read report delves into the multifaceted risks of general-purpose AI, addressing both the technological hurdles posed by rapid advancement and the complex societal implications. The latter promises to ignite passionate and intricate debates. That said, just as resistance by car manufacturers to car regulations didn't prevent the implementation of essential safety measures, AI's pervasiveness necessitates careful consideration of its impact on our lives. While AI is increasingly ubiquitous, the meandering course of its influence on societal norms is far from over, allowing time for society to adapt to the changes it will bring.

Of course, one of the most pressing concerns surrounding AI is its potential impact on employment. Understandably, as work occupies a significant portion of our lives and is fundamental to our livelihoods and societal stability, anxieties arise when predictions suggest 40% to 60% of current jobs may vanish. While acknowledging the potential for new job creation offers some solace, the scale of a potential disruption is daunting. Yet, how daunting and by when?  What if the reality of enterprise adoption called for a more nuanced perspective?

Most AI applications target specific business challenges, naturally mitigating some security and safety risks. Critically, companies seriously considering AI integration should prioritize employee training before launching initiatives, as discussed in my previous post. Furthermore, widespread AI adoption faces two significant hurdles.

-        First, organizational structures rarely align with AI's demands. Companies must revisit business processes and management paradigms, shifting from system-driven infrastructures to process-driven data exploitation. This includes redefining data ownership, governance, and potentially integrating IT under Chief AI Officers.  

-        Second, implementing AI at scale is a complex, lengthy endeavor. Despite access to leading models and benchmarks, and the ability to combine them effectively, much of the implementation remains artisanal due to limited tooling. In addition, achieving business goals requires extensive iteration and refinement.

Bridging the Gap: AI's Promise vs. Reality

While the paradigm shift towards AI is undeniable, neglecting the intricacies of its integration into society and organizations can be counterproductive. Ignoring this "drift" can lead to disappointment, resistance, and backlash, potentially hindering progress.

Currently, a significant gap persists between the idealized vision of AI and the realities of its implementation. For instance, optimistic projections, such as the 2023 Goldman Sachs forecast of a 7% (or $7 trillion) increase in global GDP and a 1.5 percentage point boost in productivity growth over a decade, contrast sharply with more conservative estimates. Daron Acemoglu, the 2024 Nobel Laureate, calculated a potential total factor productivity (TFP) gains of less than 0.53% over 10 years. The actual outcome likely lies somewhere in between. It’s contingent on the pace of AI integration across various sectors and contexts—and the social learning that goes with that pace.

Sometimes, one can only hurry more slowly than one would like. The Latin phrase "Festina lente" (Make haste slowly) is a classic oxymoron that also reflects the complex course of transformative innovations.