Does Neo know kung fu in the real world?
The short answer is no, at least not in the way he does inside the Matrix. When Neo says “I know kung fu”, he has downloaded a combat module into his brain. In the real world, his biological body has not undergone the years of physical conditioning, stretching, and muscle tearing required to perform high-level martial arts.
This leads to an uncomfortable conclusion: Neo only knows kung fu in the Matrix. Neo needs the Matrix to express this skill. You can even say that Neo doesn’t really know kung fu, he instead does kung fu inside the Matrix. He is a fragile expert.
A fragile expert is someone whose output stays strong only while the tool and context stay stable - and degrades sharply under outage, audit, or distribution shift. That brittleness comes from a gap between output and retained understanding.
How does one become a fragile expert? By following the white rabbit (1), deeper and deeper down the rabbit hole. You might remember your first white rabbit moment: when you first asked an LLM to write a complex Python script or draft a strategic roadmap, and it delivered something 90% perfect in seconds. That dopamine hit was intoxicating. Soon enough, in the time it took you to perfect a PRD you produced a working prototype. Or you replaced a week of glue code with an afternoon of tool use.
Now contrast this to the old way of becoming an expert: 10,000 hours of deliberate practice (2). Practice your way to mastery, be at the edge of your ability, be continuously committed to it, grind it out, etc. Practice, practice, practice.
Yet (expert) knowledge is not performance. If they were correlated in the past, the correlation is weakening in the age of AI. LLMs can let a competent, tool-savvy person perform surprisingly close to an expert’s output. If you can deliver materially the same output without building mastery, how you got there won’t matter to the consumer of that output. If the sausage tastes the same, nobody asks how it was ground.
Expertise has always been tool-mediated. The difference now is the speed and opacity of the capability download, and the difficulty of knowing when you’re off-distribution. Here are some of the ways in which the gap between output and retained understanding (aka fragility) can hurt you:
- If you don’t know the fundamentals, you cannot validate the output. If Neo downloads a corrupted combat file where he leaves his guard down every time he kicks, things will not end up well for him in the Matrix.
- AI works on patterns and probabilities (the 80% use case). Mastery is defined by how you handle the 20% (the novel, the weird, the broken). In Neo’s case, when he fights Bane (3) in the real world, he doesn’t know kung fu anymore; he has to rely on instinct and adaptation and the result is a messy, desperate, and uncoordinated brawl.
- Deep knowledge is owned; AI knowledge is rented. Relying entirely on the tool creates a dependency that can degrade skills. Outside the Matrix, Neo’s muscles have atrophied from not being used.
So what does this all mean? AI - “tool AI” and not “agent AI” (4), enables high output without mastery, creating ‘fragile experts’ who depend on tools. How this looks in practice runs the gamut. There are various degrees of how much of sensing / analysis / decision / action you hand over to the LLM (5). It will also look different depending on whether you retain ownership of the task decomposition vs. delegating it to the LLM (6). Regardless of the degree or mode, fragility always exists and is the cost of speed.
Even as a fragile expert, you still have to be an expert, not a mere operator. The expertise is in the quality of the output, including fitness for purpose, resilience and reliability. The fragility comes from the knowledge part. Do you remember all the decisions that you made (even if you understood them at the time and directed the LLM appropriately)? Can you speak fluently and authoritatively about the design / execution choices or do you need the LLM post hoc to ‘remind’ you about your shared thinking until you have enough repetitions to internalize the discourse?
Coming full circle, does it matter that Neo doesn’t know kung fu in the real world? Not really. Neo eventually transcends the “kung fu” module not by following its moves perfectly, but by understanding the code behind the moves (seeing the Matrix). The lesson for us? Embrace your newfound expertise but actively manage the fragility that comes with it. The speed is real. The output is real. But so are the moments when the gap between output and understanding will be exposed. When do those moments come? That’s what I’ll explore next.
This is the first article in a series about fragile expertise. In future articles I will explore:
- How do you minimize fragility by directing your practice / learning / understanding to map to the gaps in the “jagged frontier” (7)?
- What does it mean for hiring? Today’s hiring managers are still equating knowledge with performance. How can you, the fragile expert, adapt?
- How does fragility affect LLM adoption inside domains where tail-risk and accountability are paramount (law, government)?
NOTES (1) “If you want to know, follow the white rabbit.” The Matrix (Wachowski siblings, screenplay dated April 8, 1996; film released 1999, Warner Bros.).
(2) Malcolm Gladwell. Outliers: The Story of Success. Little, Brown and Company, 2008. and K. Anders Ericsson, Ralf Th. Krampe, Clemens Tesch-Römer. “The Role of Deliberate Practice in the Acquisition of Expert Performance.” Psychological Review 100, no. 3 (1993): 363–406.
(3) The Matrix Revolutions (Wachowski siblings, 2003; Warner Bros.).
(4) Richard S. Sutton, Upper Bound 2023 (YouTube keynote), timestamp ~08:12, “tool AI … agent AI,” https://www.youtube.com/watch?v=n4aev-6Z6U4
(5) Raja Parasuraman, Thomas B. Sheridan, Christopher D. Wickens. “A Model for Types and Levels of Human Interaction with Automation.” IEEE Transactions on Systems, Man, and Cybernetics—Part A 30, no. 3 (2000): 286–297.
(6) Steven Randazzo, Hila Lifshitz, Katherine C. Kellogg, Fabrizio Dell’Acqua, Ethan Mollick, François Candelon, Karim R. Lakhani. Cyborgs, Centaurs and Self-Automators: The Three Modes of Human-GenAI Knowledge Work and Their Implications for Skilling and the Future of Expertise. Harvard Business School Working Paper 26-036, 2025.
(7) Dell’Acqua, F., McFowland III, E., Mollick, E., Lifshitz-Assaf, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2023). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality (HBS Working Paper No. 24-013).