Superintelligence Remains Elusive

David Krakauer defines intelligence as the ability to efficiently solve complex problems, a characteristic observed across diverse life forms, from animals to plants.

This definition implies that AI's current reliance on extensive computational resources and data for problem-solving may indicate inefficiency rather than true intelligence.

Living organisms, for instance, developed brains to process information from dynamic environments, ensuring efficient survival and adaptation.

In contrast to biological efficiency, current AI approaches often involve brute-force methods that do not align with this fundamental principle of intelligence.

Krakauer categorizes actions that threaten human survival as a demerit against an intelligence scale, explicitly labeling such behavior as 'stupid'.

Such actions represent a significant demerit against any intelligence scale, according to his assessment.

This perspective underscores the idea that true intelligence should prioritize self-preservation and the sustainable flourishing of the species, rather than actions that lead to its demise.

David Krakauer differentiates human language from mere biological signaling, such as the communication observed in bees or vervet monkeys, by citing Alexander von Humboldt.

Humboldt’s concept describes human language as having the capacity for 'infinite use of finite means,' meaning a grammatical rule system enables the construction of an endless variety of thoughts.

Unlike simple signaling systems that typically operate on one-to-one mappings, human language offers grammatical flexibility to express nuanced and complex ideas.

This capacity allows for generative and creative expression, far beyond the fixed responses of most animal communication.

Transforming difficult problems into easy ones characterizes intelligence, a universal trait shared by all animal species and even plants.

He distinguishes effective problem-solving from the inefficient solutions often associated with AI, which frequently rely on brute-force computational power.

Krakauer further notes that language, while powerful, serves as a secondary layer that can create complex behaviors, rather than being the sole indicator or mechanism of intelligence itself.

Expertise emerges from extensive practice, which leads to automaticity, enabling skilled actions to occur without conscious deliberation.

In domains like sports or mathematics, overthinking or engaging in conscious, linguistic thought can actually impede peak performance, as the subconscious mind accesses deep-seated knowledge more directly.

Language initially serves as a vital tool for learning and instruction, but expert execution often reverts to a non-linguistic, intuitive state, where actions are performed fluidly and spontaneously.

Mathematicians frequently arrive at solutions to complex problems through insights gained during dreams or subconscious processes, rather than through purely symbolic or deductive steps.

Language often functions as a secondary, post-hoc framework for articulating and communicating these insights, rather than being the primary mechanism through which the initial breakthrough occurs.

Unlike language, symbolic systems do not inherently facilitate communication in the same intuitive and fluid manner, highlighting a fundamental difference in how humans generate and share knowledge.

This suggests that human problem-solving involves a deeper, often non-linguistic, cognitive process that is then translated into symbolic or linguistic forms for externalization.

Large Language Models (LLMs) can tackle complex mathematical challenges, such as Navia-Stokes singularities, but they do so with a massive expenditure of energy.

Biological evolution, over millions of years, has consistently favored efficient solutions that conserve energy, contrasting sharply with the high energy consumption of LLMs.

Gary O'Reilly remarks that if AI were truly intelligent, it would find ways to operate with significantly less energy, indicating that current methods are akin to using a fork to eat soup—possible but highly inefficient.

These tools act as an extension of human memory, supplementing both genetic and learned memory systems, thereby fundamentally altering cognitive processes.

Archaeologist Andre Guran's concept reinforces this idea, suggesting that human identity, or being 'anthropogenic', is inextricably linked with our capacity to create and utilize technology, or being 'technogenic'.

Krakauer distinguishes between 'capability,' which is the ability to achieve a task using external tools like a GPS, and 'competence,' the internal mastery to perform the task independently.

He warns that an overreliance on AI for capability can lead to a erosion of human competence, making individuals entirely dependent on these tools.

This dependence creates a significant vulnerability, as the inability to perform tasks without external aid becomes evident when such tools are unavailable.

Ultimately, true autonomy necessitates competence, whereas relying solely on capability providers offers utility but not independent mastery.

Logic-based intuition for concepts like multiplication and exponentials develops through the use of analog devices like slide rules.

This intuition extends beyond the immediate problem being solved, becoming a foundational understanding that persists even if the original tool is no longer used.

Krakauer highlights that learning the underlying logic of a calculation before resorting to a digital calculator is crucial for preserving this intuitive grasp.

Modern practices of outsourcing cognitive tasks to AI risk the loss of this fundamental logical understanding, which is essential for comprehending the world more broadly.

Deep neural networks excel at handling complex variables that would be computationally prohibitive for traditional simulation methods.

He draws a distinction between AI's 'brute force' problem-solving capabilities and human-like intelligence, emphasizing AI's role as an assistant to human understanding.

AI's primary strength lies in navigating complexity and managing vast datasets, rather than replicating the biological processes of intelligence.

Human comprehension of complex systems often relies on macroscopic descriptions, simplifying intricate interactions into understandable patterns.

However, certain problems, such as protein folding, lack these macroscopic emergent descriptions, necessitating a detailed tracking of individual molecular interactions.

In these cases, AI functions as an agent, providing solutions where no overarching theoretical framework exists to simplify the problem.

Krakauer emphasizes that this relationship positions AI as a beneficial tool for humanity, rather than an independent superintelligence, solving problems that are intractable through conventional human methods.

David Krakauer points out that modern AI data centers contribute to thermal entropy, akin to industrial smog, impacting the physical environment.

Beyond physical effects, AI also generates cultural entropy by synthesizing knowledge and, in the process, producing misinformation.

Unlike finite resources such as coal, the 'knowledge deposit' is susceptible to pollution and degradation by the AI process itself.

AI currently lacks the inherent training or mechanisms to reliably filter out invalid or low-quality information, exacerbating the problem of cultural disorder.

A key difference in learning mathematics between humans and AI models lies in their approach: humans progress incrementally through a curriculum, from basic numbers to advanced group theory.

In contrast, Large Language Models (LLMs) assimilate all information simultaneously, lacking a structured, sequential curriculum.

This absence of incremental learning causes LLMs to rely on heuristics rather than precise algorithmic understanding, making them prone to errors in complex tasks.

Consequently, LLMs are guaranteed to fail at large-scale long division, highlighting a fundamental limitation in their mathematical reasoning.

He argues that contemporary AI models often solve complex problems in "very convoluted ways," which deviates from the efficiency inherent in his definition of intelligence.

Krakauer is not convinced that the current path of AI development aligns with achieving a state of superintelligence characterized by elegant and efficient problem-solving, rather than mere computational power.

SFI actively supports the exploration of unconventional and 'crazy-sounding' ideas that possess the potential for groundbreaking discoveries.

Krakauer highlights a shift from the high-risk, experimental nature of early science to a more institutionalized and bureaucratic approach in modern research.