A History of First Step Fallacies: Why Climbing a Hill is Not Reaching the Sky
A History of First Step Fallacies
Hubert Dreyfus Critiques the "First Step Fallacy" in AI, arguing that early incremental successes in symbolic or rule-based systems do not guarantee a path to General AI. He posits that the persistent failure to solve the "commonsense knowledge problem" reveals a fundamental discontinuity between narrow task performance and human-like intelligence.
TL;DR
In this seminal critique, philosopher Hubert Dreyfus argues that the history of Artificial Intelligence is a graveyard of "First Step Fallacies"—the misguided belief that narrow success in symbolic logic or rule-based tasks leads linearly to human-level intelligence. He contends that without embodied finitude and a sense of relevance, AI is merely "climbing a tree" while claiming it is "landing on the moon."
The Core Motivation: The Mirage of the Singularity
Silicon Valley remains obsessed with the "Technological Singularity"—the moment algorithms transcend human consciousness. Dreyfus identifies this as "rapture for nerds," a quasi-religious yearning for digital immortality. His motive is to expose the recurring pattern of "motivated madness" that has plagued AI since the 1960s: an cycle of over-optimistic prediction followed by total collapse.
The fundamental mistake? Assuming that because a computer can do something smart (like play chess), it is on the path to doing everything smart.
Methodology: The Philosophical Deconstruction of GOFAI
Dreyfus deconstructs what John Haugeland called GOFAI (Good Old-Fashioned AI). He traces its lineage not to computer science, but to failed rationalist philosophy:
- Hobbes: Reasoning as calculation.
- Descartes: The world as mental representations.
- Leibniz: A universal characteristic of primitives.
Dreyfus argues that AI researchers unknowingly tried to turn 2,000 years of "armchair philosophy" into a research program. However, as Heidegger and Merleau-Ponty argued, human intelligence is not about manipulating facts; it is about familiarity—a background of skills and socialized "being" that makes things relevant before we think about them.
(Note: This diagram would contrast the Physical Symbol System hypothesis with the reality of the Commonsense Knowledge Wall.)
The Five Fallacies of AI History
Dreyfus meticulously lists the failures of those who claimed to have taken the "first step":
- Newell & Simon (Cognitive Simulation): Claimed symbols are sufficient for intelligence. They ran into the Commonsense Knowledge Problem.
- Expert Systems: Claimed rules could capture expertise. They only produced "competence," as true experts operate on intuition, not explicit rules.
- Rodney Brooks (Representation-less AI): Built "animats" (robot insects). While they bypassed the internal model problem, they failed to move from insect-like reactions to human-like context.
- Douglas Lenat (Cyc): Attempted to "hand-feed" millions of facts into a system to reach a "crossover point." He hit an infinite regress—one needs a frame to understand a frame.
- David Chalmers & the Singularity: Assumes an "intelligence explosion" without ever identifying an algorithm that can actually handle relevance.
Experiments & Results: Brute Force is Not Intelligence
Dreyfus addresses modern success stories like Deep Blue (Chess) and Watson (Jeopardy). He argues these aren't "first steps" toward AI; they are syntactic substitutes for relevance.
| Method | Human Approach | Computer Approach | The Discontinuity |
|---|---|---|---|
| Chess | Sense of position and "vibe." | Searches 200M positions/sec. | Brute force vs. Insight. |
| Jeopardy | Understanding meaning. | Statistical phrase correlation. | Syntax vs. Semantics. |
| Common Sense | Embodied existence. | Failed lists of formal facts. | Infinite context vs. Finite rules. |
(Note: This chart would illustrate that while performance in formal domains increases with compute, "Common Sense" remains a flat line at zero.)
Deep Insight: The Embodied Necessity
Dreyfus concludes that the "invisible water" we swim in—our daily familiarity with being in a body—is the missing link. You cannot program a computer to know why it's easier to walk forward than backward or what it means to "face" a problem unless that computer has a stake in the world.
The Limitations
If Dreyfus is right, AI as we know it (symbolic or digital) may be physically incapable of human intelligence. The "software bottleneck" isn't just about missing code; it's about missing a life.
Conclusion
This paper is a sobering reminder that incremental progress is not a guarantee of ultimate success. In an era of LLMs, Dreyfus’s warning remains hauntingly relevant: are we finally solving the frame problem, or have we just built a much bigger, faster tree to climb?
Takeaway: We must face our "embodied finitude" rather than seeking to digitize ourselves into algorithms. AI success in narrow domains is a tribute to human engineering, but it is not a bridge to human being.