> For the complete documentation index, see [llms.txt](https://cafebedouin.gitbook.io/potm/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://cafebedouin.gitbook.io/potm/section-1-the-crisis-of-expertise/1-the-mullahs-map/002_simulation_and_territory.md). # Simulation and Territory: When Practice Transfers and When It Doesn't > "Question your maps and models of the universe, both inner and outer, and continually test them against the raw input of reality." ## Simulation and Territory: When Practice Actually Helps *"Question your maps and models of the universe, both inner and outer, and continually test them against the raw input of reality."* In 2016, AlphaGo beat Lee Sedol, one of the world's greatest Go players. AlphaGo learned by playing millions of games against itself—no physical board, no human opponents. It didn't just match human skill. It got better than any human, discovering moves that masters now study. In 2024, AI can write convincingly about grief and loss despite never experiencing either. It pieces together patterns from millions of human accounts to create text that sounds like someone who's been through it. These two cases show something important: **sometimes practice in simulation works better than human practice. Sometimes it just creates convincing fakery.** The difference isn't whether simulation is good or bad. The difference is in three basic questions about what you're practicing and how. ### Three Questions That Matter **1. How realistic is the practice?**\ Does it match the real thing closely enough that what you learn actually applies? **2. How quickly do you know if you got it right?**\ Can you tell immediately when you succeed or fail, or do you have to wait weeks or years to find out? **3. Does getting good at this help with that?**\ When you practice one thing, does it make you better at something else? And by how much? These three questions tell you whether your practice is building real skill, giving you useful ideas, or just making you confident about things you can't actually do. ### When Practice Is Basically the Real Thing AlphaGo works because Go has specific features: * Everyone sees the whole board—no hidden information * The rules never change * You know right away if a move worked * Win or lose is crystal clear * You can play millions of games to test ideas There's no meaningful difference between "practice Go" and "real Go." Everything that matters for getting better is captured in the practice. What you learn playing simulated games transfers completely to tournament play. This works for other things too: **Predicting how proteins fold**: The physics doesn't change. With enough computing power and verified examples, AI can predict shapes faster and more accurately than humans doing lab experiments. The practice (computational prediction) transfers directly to real results. **Spotting strokes in brain scans**: The AI needs to recognize "this pattern means stroke" faster than human doctors can look at the image. Speed saves lives. The practice is realistic enough that the skill transfers. **Practice transfers completely when:** * The rules or physics are consistent and clear * You know immediately if you succeeded * The task is recognizing patterns in large, well-structured data * Speed or precision matters more than judgment * You can practice way more than any human could This is real. In these specific cases, simulated practice produces genuinely superhuman performance. ### When Practice Helps But Isn't Enough Most things aren't like Go. Take surgical training. You can study anatomy, watch videos, practice on cadavers, use VR. All of this helps. But the first time you cut into a living person—when there's actual bleeding that responds to what you do, when the body reacts in ways the simulation didn't show, when you have to make judgment calls under pressure with real consequences—you're dealing with something the practice couldn't fully prepare you for. The practice is **less realistic** than the real thing in ways that matter: * Living tissue behaves differently than dead tissue * Real patients have complications the simulation doesn't include * The pressure of actual stakes affects how you perform * Getting it wrong means something completely different (simulation: try again; reality: someone's harmed) Some things transfer well (knowing anatomy, recognizing structures, basic technique). Other things—the judgment that comes from seeing hundreds of real cases, knowing how to adapt when something unexpected happens, the feel for when something's off—these need practice in the actual situation. **What transfers varies.** Book knowledge transfers well. Basic execution transfers okay. Judgment under real pressure transfers poorly until you've done it for real many times. ### When Different Things Help Each Other Now consider someone who downhill skis trying inline skating. Skiing isn't practice *for* skating—they're different activities with different gear and techniques. But skiing helps with skating because they share underlying structure: * How you control edges and shift weight * How you catch yourself when you start to fall * How you read terrain and manage speed * The physical sense of controlled falling Someone who skis will learn skating faster than someone who doesn't. The **cross-training effect is real** even though skiing isn't a simulation of skating. This is different from realism. It's recognizing that skill in one area can speed up learning in another area when they share deep similarities—even if they look completely different on the surface. **Cross-training doesn't require simulation.** It requires shared constraints, similar feedback, or overlapping pattern demands. Compare this to "fishers of men"—the religious metaphor for recruiting disciples. Fishing and evangelism both involve: * Patience and timing * Reading conditions * Knowing where to focus your effort * Landing what you're after The metaphor works. It's memorable. It makes fishing expertise seem relevant to a completely different domain. But does being good at fishing actually make you better at recruiting people to a religion? Almost certainly not. The surface similarity (both "catch" something) doesn't map to shared underlying skills. What a fisherman knows about water, fish behavior, and equipment doesn't transfer to reading social dynamics, timing persuasive appeals, or building community. **The cross-training effect is basically zero** despite the metaphor being clever. ### Four Types of Practice This gives us four cases: **1. Practice that's basically the real thing**\ What you learn transfers directly because the practice captures what matters. Go, protein folding, certain pattern recognition. Nearly perfect transfer. **2. Cross-training that helps**\ Different activities that share deep structure speed up learning. Skiing helps with skating, playing one instrument helps with another. Real transfer, but varies depending on what specific skill you're measuring. **3. Ideas that help you think but not do**\ Analogies or frameworks help you understand but don't build performance skill. "Fishers of men," business frameworks you haven't tested, book knowledge without practice. Low transfer for actual performance, though it might help you understand concepts. **4. Practice that makes you overconfident**\ Surface similarity creates the feeling of skill but practice doesn't improve—or actively hurts—performance in the real thing. Business case studies building confidence without judgment, therapy training that's all frameworks and no supervised practice, AI writing about experiences it can't have. Zero or negative transfer. ### When Practice Makes You Worse The most dangerous case is the fourth one: practice that feels like it's working but isn't. Consider business school case studies. They simulate decisions—you analyze a situation, make recommendations, see what actually happened. Feels like practice. But it's fundamentally different from real business leadership: * You don't live with the results of your decisions * You have way more information than real decision-makers had * You're graded on "good analysis" not "sustainable outcomes" * Everything's compressed (you analyze years in one class) * Getting it wrong means completely different things (bad grade vs. company fails) Someone can be excellent at case analysis and terrible in actual executive roles because what they practiced wasn't what they thought. They got good at case analysis—which doesn't transfer much to real leadership under constraints, ambiguous information, and actual stakes. **AI trained on its own output** shows this clearly. Each generation drifts further from the original human text that reflected actual world experience. Errors build up. The model starts optimizing for its own patterns rather than reality. Practice becomes less realistic over time, and with it, what transfers to useful real-world work. The pattern shows up everywhere: * Academic philosophy disconnected from actual contemplative practice * Strategic plans disconnected from execution constraints * Expertise claimed from consuming information rather than making decisions with consequences * Frameworks that explain everything but predict nothing **Practice degrades performance when:** * Feedback is too slow, unclear, or missing entirely * What happens when you're wrong doesn't match what happens when you're wrong for real * You're practicing for different goals than what the real thing requires * Context and stakes matter but can't be made realistic enough * The practice builds habits that don't transfer or actively interfere ### What This Means Economically This reveals something about economic value when sophisticated simulation is cheap. If your work can be done entirely with information—synthesizing, applying frameworks, producing explanations—you're competing with AI that can do this at massive scale for nearly free. The practice is realistic enough that AI performance often meets or beats humans. If your work requires checking simulation against reality—using judgment from consequences, recognizing when frameworks fail, adapting to constraints that weren't in the model—you have value that's hard to replace precisely because the practice can't be made realistic enough for simulation to work. The question for any domain: **Can the important feedback be simulated realistically enough that practice transfers to real performance?** * **Yes**: AI will likely match or beat humans * **Partially**: Simulation helps but you need real practice too * **No**: Simulation creates false confidence and might make you worse Value concentrates where practice can't be made realistic enough, or where the cost of getting it wrong is too high to accept simulated practice. ### What Actually Moves Between People When someone with real experience talks about their domain, they're not just sharing information. They're reporting from territory they've crossed—telling you where their models broke down, what their frameworks missed, how reality was different than expected. This has different properties than simulation: * It's grounded in specific situations with real limits * It includes what *doesn't* transfer (not just what does) * It includes negative knowledge (what doesn't work, what to avoid) * It's calibrated by consequences (they paid costs for being wrong) Simulation can be accurate about what's documented. It can't tell you about the gap between documentation and reality—because that gap is only visible from inside the experience. This matters more as simulation gets better. The better AI gets at producing convincing synthesis, the more valuable it becomes to say: "I've actually done this, here's where the model diverged from reality, here's what you can't learn from study." But only if you've actually crossed the territory. Only if you've paid the costs through real practice with genuine stakes. ### The Question to Ask For anything you're working on or trying to learn: **Does practice at this actually help with that?** * Learning Go by playing millions of games: very realistic, direct transfer * Learning surgery from VR: somewhat realistic, needs real practice too * Learning therapy from textbooks: helps with ideas, low performance transfer * Learning executive judgment from case studies: might build false confidence The answer tells you whether you're building real capability or just getting better at simulation. And when sophisticated simulation is free, that distinction is everything. *** ### Questions **Where might practice be making you overconfident?**\ Find one area where you've studied, practiced in simulation, or learned frameworks—and assumed you're capable. What would actually testing yourself reveal? Does practice at what you've been doing actually help with what you need to do? **When has cross-training actually worked for you?**\ Think of a time when skill from one area genuinely sped up learning something else. What did they share that made the transfer work? How did you know it was real versus imagined? What would've happened if you assumed more transfer than actually existed? **How quickly do you find out if you're right?**\ For your most important work, how immediate and clear is the feedback? How long between doing something and knowing if it worked? How clear is the signal? What does this tell you about whether simulation could build the capability you need? ### Practice **Practice > Understanding**\ Do the thing, even if you don't fully understand it. Let understanding trail behind like a shadow. *Use when*: You're tempted to wait until you're ready, clear, or in the right headspace.\ \&#xNAN;*Remember*: The body knows things the mind cannot teach in time. *** ### Footnotes \[^1]: