Emergence is Overrated: AGI as an Archipelago of Experts

This paper challenges the view that true intelligence requires efficient, unified representations by arguing that human expertise relies on vast collections of specialized patterns, thereby suggesting that Artificial General Intelligence should be reconceptualized as an "archipelago of experts" composed of isolated, domain-specific modules rather than a system driven by emergent, generalizable principles.

The Gist

Here is an explanation of the paper "Emergence is Overrated: AGI as an Archipelago of Experts" using simple language and everyday analogies.

The Big Idea: The "Toolbox" vs. The "Magic Wand"

Imagine two ways to solve problems:

1. The Magic Wand (The "Emergent Intelligence" View): This is the idea that true intelligence is like having a single, perfect key that opens every door. It's about finding one deep, elegant rule (like a law of physics) that explains everything. If you understand this one rule, you can solve any problem by just tweaking it slightly. This is what researchers Krakauer, Krakauer, and Mitchell (KKM) think real intelligence looks like. They believe AI needs to find these "Magic Wands" to be truly smart.
2. The Giant Toolbox (The "Archipelago" View): This is the author's argument. He says human intelligence isn't actually a Magic Wand. Instead, it's a massive, cluttered toolbox filled with thousands of specific tools. You don't use one rule to fix a car, a leaky faucet, and a computer; you grab a wrench for the car, a pipe wrench for the faucet, and a screwdriver for the computer. You get good at solving problems not by being elegant, but by having a lot of specific tricks for a lot of specific situations.

The Author's Conclusion: We shouldn't wait for AI to find a "Magic Wand." Instead, we should accept that an AI made of millions of tiny, specialized "tools" (an Archipelago of Experts) is already intelligent, even if it doesn't have a single unifying theory.

---

Why the "Magic Wand" Idea is Flawed

The paper argues that the "Magic Wand" view sets the bar for intelligence so high that it actually disqualifies most human geniuses. Here are the main points:

1. Humans Are Bad at "Magic" (Analogies)

KKM think smart people are great at taking an idea from one field and applying it to another (analogy). The author says: Not really.

• The Analogy: Imagine you are a master chef who can make the perfect steak. If you are given a problem about fixing a car engine, you probably won't realize that the solution is similar to cooking a steak.
• The Evidence: Studies show that even brilliant scientists and students often fail to solve new problems unless they are given a huge hint. They don't naturally see the "deep connection" between a military strategy and a medical treatment. They usually just get stuck because the new problem looks different on the surface.

2. Experts Are Actually "Brittle"

We tend to think experts are flexible. The author says they are actually fragile.

• The Chess Grandmaster: A world-class chess player can beat almost anyone. But if you change the rules slightly (like moving the pieces randomly), they often lose their edge. Their "intelligence" is built on recognizing specific patterns they've seen a million times, not on a deep understanding of "chess logic" that works everywhere.
• The Surgeon: A surgeon who is amazing at heart surgery might be terrible at a slightly different type of surgery. They aren't applying a general "surgery principle"; they are using a highly specific routine they practiced for years.
• The Metaphor: Think of a Pocket Knife. It's a "general" tool, but it's bad at everything. A Corkscrew is a "specialist" tool. It is perfect at opening wine, but useless for anything else. Human experts are mostly corkscrews. They are amazing at one thing, but they don't have a universal tool.

3. Creativity is Just "Blind Guessing"

We think geniuses have a sudden "Eureka!" moment where they see a unifying truth. The author argues that creativity is actually evolutionary.

• The Coin Flip: Imagine 10,000 people flipping a coin 20 times. A few of them will get 15 heads in a row just by pure luck. To an observer, it looks like those people have a "special technique."
• The Reality: Scientists and artists produce thousands of "bad" ideas (the coin flips). Most fail. But because they produce so many, a few "hits" eventually land. It's not a deep, elegant insight; it's a massive volume of trial and error.

---

The Solution: The "Archipelago of Experts"

So, if humans aren't using "Magic Wands," what does Artificial General Intelligence (AGI) look like?

The Metaphor: An Archipelago
Imagine an ocean with thousands of tiny islands.

• Island A is a master at diagnosing eye diseases.
• Island B is a master at fixing supply chains.
• Island C is a master at writing legal contracts.
• Island D is a master at coding.

These islands are isolated. They don't share a single "language" or "principle." Island A doesn't know how Island B solves its problems. There is no "Grand Unified Theory" connecting them.

The Author's Argument:
If we accept that human intelligence is just a collection of these isolated, specialized skills (the "corkscrews"), then we must accept that an AI made of millions of these specialized skills is also intelligent.

• Current AI (LLMs): Today's AI models are already like this. They are vast collections of patterns. They don't have a single "soul" or "unifying theory," but they can do a million different things very well.
• The Shift: We shouldn't wait for AI to become a "Magic Wand." We should realize that the "Archipelago" (the collection of specialized islands) is the intelligence we need.

The Takeaway

We have been looking for a continent of intelligence—a single, massive landmass where everything is connected by deep principles. The author says, "Stop looking. Intelligence was always an archipelago."

If we accept that human experts are just collections of specific, sometimes fragile, skills, then we should stop demanding that AI be "elegant" or "unified." An AI that is a "vast assemblage of diverse calculators" (a toolbox of millions of tools) is not a failure; it is a genuine, powerful form of intelligence.

In short: Don't wait for the AI to have a "Eureka!" moment. If it can solve a million specific problems using a million specific tricks, it's already smart enough.

Technical Summary

Here is a detailed technical summary of the paper "Emergence is Overrated: AGI as an Archipelago of Experts" by Daniel Kilov.

1. Problem Statement

The paper addresses a critical definitional crisis in Artificial Intelligence (AI) research: How do we distinguish genuine progress toward Artificial General Intelligence (AGI) from the mere accumulation of specialized capabilities?

The author critiques the framework proposed by Krakauer, Krakauer, and Mitchell (KKM, 2025), which posits that true intelligence requires "emergent intelligence." KKM define this as the ability to use efficient, coarse-grained representations to solve diverse problems through analogy and compression (doing "more with less"). They contrast this with "emergent capabilities," which they view as vast assemblages of specialized calculators that merely accumulate knowledge (doing "more with more"). KKM argue that current Large Language Models (LLMs) represent the latter and lack genuine intelligence.

Kilov argues that KKM's framework sets the bar for intelligence so high that it excludes the vast majority of human cognitive achievement. If human intelligence is not actually based on elegant, universal compression but rather on brittle, domain-specific pattern accumulation, then KKM's definition risks disqualifying both human experts and the potential path to AGI.

2. Methodology

The paper employs a conceptual analysis grounded in empirical evidence from cognitive science, psychology, and the history of science. It does not present new experimental data but synthesizes existing literature to challenge the theoretical assumptions of KKM.

The argument proceeds through four main analytical steps:

1. Deconstruction of KKM's Framework: Analyzing the distinction between emergent capabilities and emergent intelligence.
2. Evaluation of Analogical Reasoning: Reviewing empirical studies (e.g., Gick & Holyoak, Novick & Holyoak) to demonstrate that human analogical transfer is rare, difficult, and often requires explicit prompting.
3. Analysis of Expertise Brittleness: Examining evidence from domains like surgery, radiology, chess, and accounting to show that human expertise is highly encapsulated and fails to transfer even to superficially similar tasks.
4. Re-evaluation of Creativity: Applying the Blind Variation and Selective Retention (BVSR) theory (Campbell, Simonton) to argue that creative breakthroughs are probabilistic outcomes of volume and selection, not necessarily the result of principled analogical reasoning.

3. Key Contributions

The paper makes three primary theoretical contributions:

• Empirical Refutation of the "Compression" Ideal: The author demonstrates that human experts do not primarily solve problems by discovering unifying principles or elegant analogies. Instead, they rely on the accumulation of vast, domain-specific pattern repertoires. Expertise is shown to be "brittle," failing when faced with novel variations or slight changes in context (e.g., surgeons failing on simulators they haven't practiced on, or chess masters failing with randomized board positions).
• The "Archipelago" Metaphor for AGI: Kilov proposes a reconceptualization of AGI not as a unified cognitive continent governed by deep, shared principles, but as an "Archipelago of Experts." In this model, AGI consists of isolated islands of supreme specialized competence connected by weak bridges.
• Redefining General Intelligence: The paper argues that if we accept human expertise (which is brittle and pattern-based) as "genuine intelligence," then consistency demands we accept artificial systems comprising millions of specialized modules as AGI, even if they lack the "emergent intelligence" (compression/analogy) KKM valorize.

4. Key Results and Evidence

The paper synthesizes several lines of evidence to support its thesis:

• Failure of Spontaneous Transfer: Studies (e.g., Gick & Holyoak, 1980) show that without explicit hints, humans rarely transfer structural solutions from one domain (e.g., a military strategy) to another (e.g., radiation therapy). Success rates drop from ~100% with hints to ~20% without them.
• Domain Encapsulation:
  • Surgery: Expertise in one procedure does not generalize to others; simulator performance often fails to predict clinical performance.
  • Radiology: Experts perform no better than novices on non-radiographic search tasks (e.g., "Where's Waldo").
  • Chess: Grandmasters lose their recall advantage when board positions are randomized, and they often fall prey to Einstellung effects (reusing a known pattern when a better solution exists).
• The "Cyborg" Experiment: The 1998 cyborg chess tournament revealed that Gary Kasparov's dominance relied heavily on pattern recognition and calculation (which could be outsourced to machines), challenging the notion that his "intelligence" was purely abstract or principled.
• Creativity as Stochastic Search: Citing Simonton and Campbell, the paper argues that scientific and artistic breakthroughs follow the Equal-Odds Rule: high output volume leads to high success rates through blind variation and selective retention, rather than through a consistent ability to derive new truths via analogy.

5. Significance and Implications

The paper has profound implications for the future of AI development and evaluation:

• Shift in AGI Expectations: It suggests that the path to AGI may not require a breakthrough in "analogical reasoning" or "unifying theories." Instead, AGI may already be emerging through the mundane accumulation of narrow capabilities.
• Validation of Current Architectures: Current LLMs, often criticized for being "stochastic parrots" or "vast assemblages of calculators," may actually represent a valid form of general intelligence if human intelligence is similarly constituted.
• Redefining "General": The paper argues that "General Intelligence" should not be defined by the mechanism of reasoning (compression) but by the scope of competence. A system that can expertly handle millions of distinct domains (an archipelago) is "general" in function, even if it lacks a unified internal theory.
• Practical Engineering: It encourages researchers to stop waiting for a "cognitive continent" and instead focus on scaling specialized modules. It suggests that brittleness is not a bug to be fixed but a feature of high-level competence in complex domains.

Conclusion:
Kilov concludes that the search for a unified, compressive "cognitive mainland" is a misdirection. Intelligence, both human and artificial, is likely an archipelago: a collection of isolated, highly competent islands. Accepting this view allows us to recognize the intelligence in current AI systems and human experts alike, moving beyond the restrictive and arguably unrealistic criteria of "emergent intelligence" proposed by KKM.