🔴🟩🦄ABC: action bundle complexity

[hyunjimoon]

Q. conditions AI-assisted decision-making (core: Minds that build coherent models and use them rationally (using incomplete modular world, uncertain abstract state, uncertain-probabilistic rules of game with meaning and inference function) are most helpful for business. To be specific, how does factors like company lifecycle stage, estimation and precision error (bias?), uncertainty, founder orientation (compete vs collaborate) and investment (execute vs control) affect the choice of value creation hypothesis (as combination of customer and technology) and value capture hypothesis (as combination of organization and competitor)?

I'm applying lakatosian defense for

ABC substantive theory

Category	Decision Variable	Details
🔴Action probabilistic reasoning can drive	multi-objective: weight	1. Nail vs. Scale optimization differs in nature (temporal (nail it has higher decision dimension so decisions interact less) vs spatial tradeoff): Adjust weights based on company stage, e.g., higher weight on talent attraction and motivation during Nail It, shifting towards long-term commitment and expertise during Scale It 2. Overall Total Utility = α * (Grow faster VS give up more ownership Utility) + β * (Flexibility VS efficiency Utility) + γ * (Incentive VS alignment Utility) Where α + β + γ = 1, and their values shift based on the company's stage (Nail It vs. Scale It) and specific circumstances - Grow faster VS give up more ownership: Balance talent_attraction_factor(pool_size) against founder_ownership_factor(pool_size) Total Utility = w1 * company_growth_utility(pool_size, is_competitive) + w2 * founder_utility(pool_size) Where w1 + w2 = 1, and w1 > w2 during Nail It, w1 < w2 during Scale It - Flexibility VS efficiency: Trade-off between decision_speed_factor(structure) and expertise_factor(structure) Total Utility = w3 * decision_speed_factor(structure) + w4 * expertise_factor(structure) Where w3 + w4 = 1, and w3 > w4 during Nail It, w3 < w4 during Scale It - Incentive VS alignment: Balance motivation_factor(vesting, upfront) against long_term_commitment_factor(vesting, cliff) Total Utility = w5 * motivation_factor(vesting, upfront) + w6 * long_term_commitment_factor(vesting, cliff) Where w5 + w6 = 1, and w5 > w6 during Nail It, w5 < w6 during Scale It
	multi-level: degree of relaxation	- relax A - relax TA - relax AD - relax TAD
🟩Bundle of choice probabilistic reasoning can inform	value create	- Technology choice (e.g., Tesla's lithium-ion batteries) - Customer segment (e.g., high-end sports car market)
	value deliver	- Supply chain decisions (e.g., outsourcing vs. vertical integration) - Product features (e.g., performance, range)
	value capture	- Organizational structure (e.g., agile vs. integrated) - Competitive strategy (e.g., first-mover advantage, IP strategy)
🦄Complexity probabilistic reasoning can lower	by stochasticity and utility	- EDMNO (Entrepreneurial Decision Making with Non-additive & stochastic utility) - decompose stochastic to uncertainty from Technology, Customer preference, Corporate strategy, Competition, Capital market, Regulatory policy, Industry structure, (business cycle??) - Reliability, Commitment, and Irreversibility factors - closer to matroid, doer has advantage (greedy)
	by character of world model	- Causal structure is modellable (non-modellable - use evolutionary learning rather than probabilistic as world model cannot be constructed) - Dynamic (model development speed is slower than evaluation??) - Estimating rare event probability is crucial - Form of information is diverse (text, numeric)

---

NEED UPDATE FROM ABCD to ABC.

Theory1. axiom architecture #244 (comment)

Theory2. bundle of choices #244 (comment)

Theory3. complexity-based learning #244 (comment)

Theory4.dynamic dilemma of how dilemma evolve across nail (modularity) to sail (integrated), which synthesize theory 1,2,3 bundle of choices with increasing complexity

1 architecture of axiom analyze three: "learning 📦 modular worlds: 🧭investment and orientation compass across 🌎NSS world", "learning 🎲 uncertain worlds: uncertainty across value create, delivery, capture", "making decision: freedom, constrain across 🌎NSS world", 2 bundle of choices introduces three algorithms for forming choice triplet and using that to form feasibility space with desirability vector, 3 complexity analyze algorithm complexity for entrepreneurial decision making problem, 4 decision dynamics review how complex choices can be organized into bundles on architecture of axiom and explain how to choose most reversible path given deterministic goals.

three hypothesis

H1. pivoting
H2. some combination of below uncertainty situation:

• scale, collaborate, control need more simulation (related to meaning function which translates natural language into probabilistic programming language (PPL) statements that represent linguistic meaning with respect to a symbolic world model and inference function computes probabilities over the space of possible worlds consistent with and conditioned on information in the language) i.e. complexity of meaning and inference function would change during NSS which is related to expressed world in language and actual world?? anyhow, world model alignment happens whenever scaling-need-motivated-monetization-need-motivated-collaboration.
• when agent has long enough time to invest in capabilities than resources
• when usefulness, usability (target) has low uncertainty so that fitness has less uncertainty

H3. outcome of 1, 2 can be applied to ensemble evolutionary leaning and bayesian learning

[hyunjimoon]

figure and table design
bayes-sd-valuesim-ent
decision partner
pivot game

[hyunjimoon]

1. 🔴Action probabilistic reasoning can drive

architecture for learning modular worlds and making decision across nail, scale, sail stage

learning 📦 modular worlds: 🧭investment and orientation compass across 🌎NSS world

• value create and capture strategy of ip, disruptor, value chain, architectural (🐆🐯🐬🐘) strategy across jungle, mountain, sea (🌳⛰️🌊)

learning 🎲 uncertain worlds: uncertainty across value create, delivery, capture

• uncertainty: entrepreneurs are uncertain about the value of particular strategies but also the range of value that might result from their idea
• noisy learning: learning is noisy and ongoing so finding out more about more route often allows a reassement of other alternatives

Error Type	Value Domain	Uncertainty Components
Statistical Error	Value Creation	Customer, Technology
Approximation Error	Value Capture	Organization, Competition, Regulatory, Industry
Optimization Error	Value Delivery	Product, Market

Category	Subcategory	Description
Value Create 🛠️	Technological Uncertainty 🔬	- Arises from inherent architectural differences across technologies - Leads to different abilities in quality levels, innovation levels, and cost structures - Technology: tools, techniques, designs, and knowledge used to create practical value for consumers
	Customer Uncertainty 🧑‍🤝‍🧑	- Customer: any person, group, or organization who will pay the startup for goods or services
Value Capture 💼	Agent: Organizational Uncertainty 🏢	- Arises from firm's agility or sluggishness in allocating resources to quality vs. innovation - Modular industry structure makes firms more agile than integrated structure - Agility varies across industries depending on product/service nature - Organization: initial choice of capabilities and resources determined by entrepreneur's team and culture - Enables delivering value creation and capture; turning vision into reality
	Agent: Competition Uncertainty 🥊	- Competitor: a firm providing similar product/service or solving the same/similar customer need
	Environment: Regulatory Uncertainty ⚖️	- Arises from uncertainty about regulatory action - Includes scope and timing related uncertainty - Impact of antitrust regulation on resulting industry structure
	Environment: Industry Uncertainty 🏭	- Arises from actions of firms, regulators, and nature of technology - Influenced by technological architecture and antitrust regulation - Introduces variation in rate of industry structure becoming integral or modular
	Environment: Industry Clockspeed ⏱️	#100
Value Delivery 📦	Market Uncertainty 📊	- Arises from actions of firms and consumers - Supply side: firm's decisions on switching costs, network effect, integration/modularization - Demand side: consumers' preferences for quality, innovation, price, and compatibility
	Product Uncertainty 🎁	- Set of potential product or service designs, features, and attributes

making decision: freedom, constrain across 🌎NSS world

Category	Ent Axioms	Nail It	Scale It	Sail It
Integer Optimization	Freedom	Large set of potential paths; Many possible actions	Focused set of growth strategies; Balanced set of actions	Limited strategic options; Few established processes
	Constraint	Minimal resource sharing; Rapid experimentation and pivoting	Rationed allocated resources; Structured growth and replication	Complex resource management; Bureaucracy and ossification
Decision Space Complexity		High: Complex due to high uncertainty and many unknowns, requiring rapid experimentation and flexibility	Medium: More structured but still complex, balancing known factors with new challenges as the business grows	Low: Highly constrained and potentially over-complicated due to established processes, bureaucracy, and high opportunity costs for changes
Matrix Analogy		"Fatty" matrix (many actions, fewer constraints)	"Square" matrix	"Skinny" matrix (fewer actions, more constraints)

• freedom: there is more than one potential path to create and capture value from an idea
• constraint: an entrepreneur cannot pursue all these paths at the same time
• tool: integer optimization theory and programming

$\underset{x \in X, \pi(x; \psi) >0}{max} u(\pi, x; \theta)$
connection with
Hakansson71-OptEntrepDec.pdf which says non-additive utility with opportunity-depedence makes entrepreneurial decision difficult
claude,

Variable	Meaning	Description	Tesla Roadster Example
x	Decision variables / Strategies	Represents the choices available to the entrepreneur, such as investment decisions, consumption choices, or resource allocation strategies. It's selected from the set X of all possible strategies.	Tesla's decision to use lithium-ion batteries instead of lead-acid batteries, despite higher costs.
X	Set of all possible strategies	The complete set of options available to the entrepreneur.	Tesla's range of possible strategies included partnering with established automakers, developing their own technology, or licensing technology from others.
π(x; ψ)	Profitability/Feasibility function	A function that determines whether a strategy x is profitable or feasible given parameters ψ. The constraint π(x; ψ) > 0 ensures only viable strategies are considered.	Tesla's assessment of the feasibility of producing a high-performance electric sports car given the state of battery technology and available manufacturing partnerships.
ψ	External parameters	Represents factors that affect the profitability or feasibility of strategies, such as market conditions, economic states, or other external factors.	The state of lithium-ion battery technology, availability of manufacturing partners like Lotus, and the market demand for high-performance electric vehicles.
u(pi, x; θ)	Utility function	The function the entrepreneur aims to maximize. It depends on the outcomes (pi), chosen strategy (x), and the entrepreneur's preferences (θ).	Tesla's evaluation of the utility of producing a high-end sports car first, considering factors like brand building, technological learning, and future market potential.
pi	Outcomes / Returns	Represents the results or returns from implementing strategy x.	The performance achievements of the Roadster (0-60 mph in 3.9 seconds, 245-mile range) and its impact on Tesla's brand and future potential.
θ	Entrepreneur's preferences	Parameters representing the entrepreneur's value system, risk aversion, time preference, or strength of bequest motive.	Elon Musk's preference for high-risk, high-reward strategies and long-term vision for sustainable transport, even at the cost of short-term profitability.

#190 , #191 , #192

architecture for learning and making decision in 📦 modular and 🎲uncertain world

[hyunjimoon]

.

[hyunjimoon]

if we linearize it, and express the above as max c'x s.t. Ax = b, evolution of A matrix is fat2skinny from nail-scale-sail. simulation and calibration is most helpful when strategy-operation is reversible. https://claude.ai/chat/32012e69-6125-46fc-8c5b-6837d45b854f

[hyunjimoon]

Entrepreneurial Uncertainty	Machine Learning Analogy	Key ML References
Technological uncertainty	Inductive bias in algorithm selection	Mitchell (1997), "Machine Learning"; Domingos (2015), "The Master Algorithm"
Customer uncertainty	Generalization to unseen data	Vapnik (1998), "Statistical Learning Theory"; Goodfellow et al. (2016), "Deep Learning"
Organizational uncertainty	Bias-variance tradeoff	Geman et al. (1992), "Neural Networks and the Bias/Variance Dilemma"; James et al. (2013), "An Introduction to Statistical Learning"
Competition uncertainty	Adversarial examples	Szegedy et al. (2013), "Intriguing properties of neural networks"; Goodfellow et al. (2014), "Explaining and Harnessing Adversarial Examples"
Regulatory uncertainty	Domain adaptation	Ben-David et al. (2010), "A theory of learning from different domains"; Csurka (2017), "Domain Adaptation in Computer Vision Applications"

[hyunjimoon]

2 🟩Bundle of choice probabilistic reasoning can inform

Moon24_proprietary_bundle of choice forms feasibility space and desirability vector

lay out two definitions and three algorithms to design simulation tool to calculate cost of committing.

Definition 1 (Choice Sets and Groups) A choice set is a collection of potential alternatives with similar calculation structure of expected value. For easier formulation of committing to certain choice as optimization problem with state and action, we group sets based on entrepreneurial decision making architecture for product market fit.

Definition 2 (Choice Triplet) A choice triplet consists of interrelated choices across three groups of choice set:
• Supply: The core technological and organizational capabilities and approaches
• Product: features that matches supply and demand choice group, abstracting underlying fulfillment choice process determined by technology, organization, customer, competitor choice set
• Demand: The target market segment shaped by customers and competitors

Algorithm 1 (forming choice set)
Algorithm 2 (defining feasibility space)
Algorithm 3 (desirability vector over feasibility space)

[hyunjimoon]

Aspect	Nail It	Scale It	Sail It
Decision Space Complexity	High: Complex due to high uncertainty and many unknowns, requiring rapid experimentation and flexibility	Medium: More structured but still complex, balancing known factors with new challenges as the business grows	Low: Highly constrained and potentially over-complicated due to established processes, bureaucracy, and high opportunity costs for changes
Matrix Analogy	"Fatty" matrix (many actions, fewer constraints)	"Square" matrix	"Skinny" matrix (fewer actions, more constraints)
Freedom (Strategy Space)	Large set of potential paths	Focused set of growth strategies	Limited strategic options
Constraint (Resource Allocation)	Minimal resource sharing	Rationed allocated resources	Complex resource management
Decision Variables (x)	Many possible actions	Balanced set of actions	Few established processes
Feasibility Function (π(x; ψ) > 0)	Rapid experimentation and pivoting	Structured growth and replication	Bureaucracy and ossification
Uncertainty (Value Creation)	High uncertainty in desirability and utility	More concrete value proposition	Established processes and systems
Noisy Learning (Value Capture)	Larger set of unknowns (theta) compared to observables (y)	Balance between known factors and new challenges	Extensive data and established patterns
Statistical Error (Customer, Technology)	High - Customer intimacy, Problem-driven	Medium - Customer segmentation, Revenue-driven	Low - Customer data, Profit-driven
Approximation Error (Organization, Competition, Regulatory, Industry)	High - Flat structure, Minimal processes	Medium - Early hierarchy, Process creation	Low - Complex hierarchy, System-centric
Optimization Error (Product, Market)	High - Flexible, prepare to pivot	Medium - Capabilities replicators	Low - Consistent experimenters

[hyunjimoon]

3 🦄Complexity probabilistic reasoning can lower

Moon24_propriety_complexity of entrepreneurial decision making.pdf

formally defines Entrepreneurial decision making and investigate non additive (interaction term) and opportunity-based (uncertainty in future world state) makes problem complex. It reduces knapsack (NP complete) problem to EDMNO.

Definition 1 (Entrepreneurial Decision Making (EDM)). Given a set of constraints and performance measures over multiple time periods, along with a utility function U, does there exist a sequence of decisions that satisfies all constraints and achieves a utility of at least L, where the utility depends on the performance measures across time periods?

Definition 2 ( Entrepreneurial Decision Making with Non-additive Opportunity-dependent utility (EDMNO)). Given a set of constraints, performance measures, and opportunity states over multiple time periods, along with a non-additive utility function U, does there exist a sequence of decisions that satisfies all constraints and achieves a utility of at least L for some opportunity state, where the utility depends non-additively on the performance measures across time periods?

Theorem 1. EDMNO is NP-complete.
(proof incomplete (cld), but think it'd work)

---

• depending on the complexity, different learning modules (bayesian vs behavioral vs evolutionary) should be applied which i explained with tesla's supply chain, demand, cash flow, outsourcing (battery) example in abductive classification of ent_learning/pivoting model #242 (reply in thread)
• clockspeed (which organizes eight uncertainty across technological, customer, organizational, competition, regulatory, industry, market, product into three buckets of fast, medium, slow) can help me design this modular function. before getting real @chasfine's feedback, i chatted with clockspeed charlie claude to conclude "Fast clockspeed environments often have higher complexity due to rapid change and uncertainty. However, slow clockspeed industries can also be highly complex.
  The integration/modularization cycle I discuss in Clockspeed does relate to complexity. As modular structures become too complex to manage, there is often pressure to integrate to reduce coordination costs. This can temporarily slow clockspeed as the integrated structure stabilizes. Conversely, as integrated structures become too rigid, modularization allows for more rapid innovation, potentially accelerating clockspeed."
• producing value from analyzing impasse/infeasibility using dual algorithms which maintains optimality till finding feasibility (cutting plane alg) https://claude.ai/chat/92109b1c-2c71-4d14-b13c-1c53212a126f

[hyunjimoon]

lifecycle of startup, star, engine, human cld which i'm trying to make more animationy here

using 4module-5loop-3tradeoff-semiconductor cld,

[hyunjimoon]

evolving strategy:operations mapping

fat square thin A

Nail It (Fat A):
In this stage, the expected value optimization problem has many decision variables but few constraints. This could be expressed as maximizing E[U(x)] subject to Ax ≤ b, where x is a high-dimensional vector of decision variables, A is a fat matrix with few rows (constraints) and many columns (decisions), and U is the utility function. The fatness of A reflects the high degrees of freedom in exploring possible strategies.
Scale It (Square A):
Here, the expected value optimization becomes more balanced. It could be expressed as maximizing E[U(x)] subject to Ax ≤ b, where A is now a square matrix. This reflects a more structured decision space with a balance between exploration and exploitation. The number of constraints has increased, representing more defined business processes and market understanding.
Sail It (Thin A):
In this stage, the expected value optimization has fewer decision variables but many constraints. It could be expressed as maximizing E[U(x)] subject to Ax ≤ b, where A is a thin matrix with many rows (constraints) and fewer columns (decisions). This reflects a more constrained decision space with established processes and regulatory requirements.
The evolution from fat to square to thin in the context of optimizing expected value represents the changing nature of the entrepreneurial decision space. Initially, there's high uncertainty and many possible strategies (fat A). As the business model becomes clearer, the decision space becomes more structured (square A). Finally, as the business matures, it operates within a more constrained environment with established processes and regulations (thin A). This evolution affects how the expected value E[U(x)] is optimized at each stage, with the changing structure of A reflecting the shifting balance between exploration and exploitation in the entrepreneurial journey.

edit:

Charlie thinks "I think scaling has the thinnest A. The scaling Firm knows what it needs to scale and needs to be very focused on that. One in sailing, the firm needs to be doing much more exploration while exploiting, to assure that it does not get disrupted or fall behind."

angie replied:

fat -> square -> thin was a descriptive statement.
Scale: I view "knows what it needs" as no degeneracy i.e. x is uniquely determined as A^-1b meaning A is square.
Sail needs exploration as they are skinny (too many constraints from organizational/product complexity). Conscious effort is needed like artificially adding new decision variables that are less affected by existing constraints (IBM pc, alphabet) to be closer to square.
Nail needs anchoring (optimism/leap of faith) to operate efficiently with less degeneracy.
quiz: how many squares are there in nss_matrix.png?
answer: three
I also made nss_driver.png and hope to find a story that ties nss_driver with nss _constraint.png in my dream tonight :)

Reliability, Commitment, and Irreversibility

from analyzing tesla's strategic choices cld
higher reliability, commitment, irreversibility, the stronger the link between strategy2operations, which is different from my expectation that lower irreversibility would have stronger s2o link🤔

Factor	Low	High	Impact on Strategy-Operations Link
Reliability	Initial battery pack design: Frequent changes and updates required as the technology was still evolving.	Lithium-ion battery technology choice: Tesla committed to this technology early, making it a cornerstone of their strategy.	High reliability strengthens the link as it allows for consistent execution of strategy through operations.
Commitment	Initial manufacturing plan: Outsourcing to various suppliers with limited long-term commitments.	Elon Musk's financial commitment: Invested heavily in multiple funding rounds, showing long-term dedication to the strategy.	High commitment strengthens the link as it ensures resources and focus remain aligned with the strategic direction.
Irreversibility	Early prototype iterations: Allowed for frequent changes and pivots in design.	Investment in specialized manufacturing equipment: E.g., for carbon fiber body production at Lotus.	High irreversibility strengthens the link as it forces operations to closely follow strategic decisions.

Examples of how these factors strengthened the strategy-operations link:

1. Reliability: Tesla's commitment to lithium-ion battery technology

   • Strategy: Focus on high-performance, long-range electric vehicles
   • Operations: Consistent investment in battery R&D and manufacturing capabilities
   • Link: The reliable performance of lithium-ion batteries allowed Tesla to consistently deliver on its strategic promise of high-performance EVs.

2. Commitment: Elon Musk's financial and personal investment

   • Strategy: "The Secret Tesla Motors Master Plan" outlining a long-term vision
   • Operations: Continuous funding and support for operations even during challenging times
   • Link: Musk's commitment enabled Tesla to maintain its strategic direction despite operational setbacks.

3. Irreversibility: Investment in specialized EV manufacturing

   • Strategy: Produce a high-end electric sports car (the Roadster)
   • Operations: Partnering with Lotus and investing in EV-specific manufacturing processes
   • Link: The specialized investments made it crucial for operations to align closely with the strategic goal of producing the Roadster.

This table and these examples demonstrate that as reliability, commitment, and irreversibility increase, the link between Tesla's strategy and operations became stronger. This alignment was crucial in allowing Tesla to overcome early challenges and execute its ambitious vision for electric vehicles.

[hyunjimoon]

Joshua is skeptical about nail in sail "However, incumbent firms, despite being able to redeploy resources using authority, are constrained in adoption as exploration cannot mitigate the costs of disagreement." from attached paper
gans22_theory_visionary_disruption.pdf

for nail in sail to succeed, resource effect should outweigh organizational rigidity (causing disagreement) effect so I wonder whether the nail in sail design I suggested
which has non-overlapping decision variables and constraints (e.g. ibm's pc devision with chairman's support (resource effect) can sidestep Joshua's skepticism

based on intuition from physical simulation #100 (comment) my prior on success: fail ratio is 1:3!

[hyunjimoon]

Descriptive Model: "How Firms Actually Operate"

Phase	Observed Behavior	Typical Outcomes	Common Patterns
Nail	Most startups over-optimize early, trying to build perfect products	90% failure rate, often from running out of runway	Founders often caught between "move fast" advice and desire for perfection
Scale	Companies struggle with growing pains, oscillating between fast and slow modes	High variance in outcomes, successful ones find balance organically	Middle managers create processes, sometimes prematurely
Sail	Large firms generally over-invest in analysis, building complex models	Gradual decline in adaptability, vulnerable to disruption	Established processes become increasingly rigid

Prescriptive Model: "How Firms Should Operate"

Phase	Recommended Approach	Expected Outcomes	Implementation Steps
Nail	Embrace imperfection: Set clear "good enough" thresholds for decisions	Higher survival rate through runway preservation	1. Set explicit decision time limits 2. Define MVP criteria 3. Track decision velocity
Scale	Systematically increase model complexity only where data shows value	Sustainable growth with controlled burn rate	1. Measure decision ROI 2. Implement staged processes 3. Balance local/global optimization
Sail	Maintain startup-like pockets while optimizing core business	Sustained innovation alongside stability	1. Create innovation sandboxes 2. Set process bypass criteria 3. Measure process overhead

Aspect	Nail (Startup)	Scale (Mid-size)	Sail (Large Corp)
Decision Speed	Lightning fast ("Just Do It")	Balanced	Methodical (Often too slow)
World Model	Simple, incomplete	Pragmatically complex	Highly sophisticated (Sometimes over-complex)
Burn Rate	High risk of burning too fast	Optimized burn	Risk of under-utilizing resources
Resource Focus	Cash runway	Market expansion	Operational efficiency
Typical Pitfall	Running out of money from bad decisions	Growing pains from shifting between modes	Analysis paralysis from over-thinking
Fixed Cost Impact	Minimal - lean operation	Growing concern	Major constraint
Decision Framework	Intuition-driven	Data + Intuition blend	Process-driven
Innovation Style	Rapid experimentation	Structured iteration	Carefully planned initiatives
Resource-Rationality Level	Low (Degree 0.2-0.3)	Medium (Degree 0.4-0.7)	High (Degree 0.8-0.9)
Success Metric	Survival and product-market fit	Growth and market share	Profitability and stability
Optimal Time Horizon	Days/Weeks	Months/Quarters	Years
Key Trade-off	Speed vs. Survival	Growth vs. Stability	Efficiency vs. Agility
Core Challenge	Making decisions fast enough to survive	Finding balance between speed and process	Avoiding organizational inertia

[hyunjimoon]

using gpt "choosing the right order to build things" then"figuring out how to combine all these exact requests"

[hyunjimoon]

[hyunjimoon]

.

[hyunjimoon]

optimization vs action, sequential vs parallel, bayes vs evol learning

#242 suggests tesla examples for each, #246 (comment) explains low test cost (relative to pivot)💸, high technological and market uncertainty 🎲 , low correlation btw tech innov an customer segment 🧩 true value as conditions where parallel entrepreneurship and evolutionary learning is better

1. Heimann93_ChallengerOrgStr.pdf which discusses organizational reliability and structure, focusing on parallel versus serial configurations. While originally applied to NASA, these concepts are relevant to business environments, particularly in uncertain fields. Parallel structures, where multiple components operate independently, are advantageous in highly uncertain scientific or technological areas. In these contexts, the cost of a type II error (missing a potential breakthrough) often outweighs the cost of type I errors (pursuing some unproductive paths). For example, in pharmaceutical research or rapidly evolving tech sectors, pursuing multiple development paths concurrently can increase innovation chances by reducing type II errors, albeit at the risk of increased type I errors. The paper's k-out-of-m unit network concept is particularly applicable to portfolio management in venture capital or R&D resource allocation, allowing businesses to spread risk across multiple projects while maintaining a success threshold. This approach is especially beneficial in highly uncertain fields, as it mitigates type II errors by increasing the likelihood of capturing rare but valuable opportunities. However, businesses must balance this approach with resource constraints and the potential costs of type I errors from pursuing multiple paths simultaneously.

Business Situation	Description	Relation to Process/Concept
High uncertainty environments	Fields with significant scientific or technological uncertainty	Aligns with complex system development process (c)
Rapid technological change	Industries experiencing fast-paced innovation	Similar to spiral product development process (b)
High-stakes innovation	Potential payoff for breakthrough is extremely high	Relates to Type II error concept in Heimann paper
Platform technologies	Core technology applicable across multiple domains	Reflected in generic product development process (a), but with multiple concurrent instances
Resource-rich environments	Capital and talent readily available	Leverages resources across multiple ventures
Interdisciplinary fields	Breakthroughs occur at intersection of domains	Facilitates cross-pollination of ideas
Long development cycles	Extended product development timelines	Helps maintain momentum and diversify risks
Regulatory uncertainty	Fields influenced by evolving regulations	Allows hedging against policy changes
Market uncertainty	Unclear which application or segment will gain traction	Enables testing of multiple value propositions
Talent development	Organizations aiming to cultivate entrepreneurial talent	Provides more learning opportunities across ventures

Bayes vs evolutionary

Let L be the learning approach, where L ∈ {LB, LE}
LB = Bayesian Learning
LE = Evolutionary Learning
For Horizon 1 (H1): P(LB|H1) > P(LE|H1)
For Horizon 2 (H2): P(LB|H2) ≈ P(LE|H2)
For Horizon 3 (H3): P(LE|H3) > P(LB|H3)
Where P(L|H) is the probability of using learning approach L given horizon H.

Formal Characteristics of Learning Approaches:

Bayesian Learning (LB):

Representation: Graphical models (e.g., Bayesian networks)
Evaluation: Posterior probability P(H|D), where H is the hypothesis and D is the data
Optimization: Probabilistic inference (e.g., MCMC, Variational inference)

Evolutionary Learning (LE):

Representation: Genetic programs or solution populations
Evaluation: Fitness function f(s), where s is a solution
Optimization: Genetic algorithms (selection, crossover, mutation)

Decision Rule for Learning Approach:

Choose LE if max(Ut, Uc, Um, Uo, Ucomp, Ur, Ui) > threshold T
Choose LB if max(Ut, Uc, Um, Uo, Ucomp, Ur, Ui) ≤ T
Where T is a threshold determined by the organization's risk tolerance and prior knowledge.

Tesla Roadster Example:

For the Roadster (primarily a Horizon 2 opportunity):
U = f(Ut, Uc, Um, Uo, Ucomp, Ur, Ui)
= f(high, medium, high, high, medium, medium, high)
max(Ut, Uc, Um, Uo, Ucomp, Ur, Ui) = high > T
Therefore, LE (Evolutionary Learning) is more appropriate, but elements of LB (Bayesian Learning) can be incorporated where uncertainty is medium or lower.

---

1. Regulatory Uncertainty:

   • This fits best under Approximation Error (Value Capture) because:
     • Regulatory actions, especially antitrust regulations, directly affect a firm's ability to capture value.
     • It influences the competitive landscape and the firm's ability to accumulate market power.
     • Regulatory uncertainty impacts how firms can structure their operations and competitive strategies.

2. Industry Uncertainty:

   • This also aligns well with Approximation Error (Value Capture) because:
     • Industry structure (integral vs. modular) significantly affects how firms can capture value.
     • It's influenced by technological architecture, which relates to how firms can position themselves in the value chain.
     • Industry uncertainty impacts the competitive dynamics and the firm's ability to capture a share of the industry's value.

This placement emphasizes that while customer needs and technological possibilities define the potential value (Value Creation), the actual value a firm can capture is significantly shaped by its ability to navigate organizational, competitive, regulatory, and industry-wide challenges. The optimization of product-market fit (Value Delivery) then determines how effectively this captured value can be delivered to customers.

---

Industry	Search Strategy	Reason for Parallel (if applicable)
Personal computers	Parallel	Fast product tech clockspeed suggests low test costs 💸, high uncertainty 🎲, and potentially low correlation 🧩 due to rapid changes
Computer-aided software engineering	Parallel	Similar to personal computers, with fast product tech clockspeed 💸🎲🧩
Toys and games	Parallel	Very fast product tech clockspeed indicates low test costs 💸, high uncertainty 🎲 in customer preferences
Athletic footwear	Parallel	Fast product tech clockspeed, low test costs 💸, high uncertainty in customer preferences 🎲
Semiconductors	Parallel	Fast product and process tech clockspeed suggests low test costs 💸 and high uncertainty 🎲
Cosmetics	Parallel	Relatively fast product tech clockspeed, low test costs 💸, high uncertainty in customer preferences 🎲
Bicycles	Sequential	Slower clockspeeds suggest higher test costs and lower uncertainty
Automobiles	Sequential	Medium clockspeeds and high development costs favor sequential approach
Computer operating systems	Sequential	Medium clockspeeds and high development costs favor sequential approach
Agriculture	Sequential	Slower clockspeeds and high implementation costs favor sequential approach
Fast food	Parallel	Fast product tech clockspeed suggests low test costs 💸, high uncertainty in customer preferences 🎲
Beer brewing	Sequential	Very slow process tech clockspeed suggests high pivoting costs
Airlines	Sequential	High costs and slow hardware clockspeed favor sequential approach
Machine tools	Sequential	Medium clockspeeds and specialized nature favor sequential approach
Pharmaceuticals	Sequential	Long development times and high costs favor sequential approach
Aircraft (commercial)	Sequential	Very slow clockspeeds and high costs strongly favor sequential approach
Tobacco	Sequential	Slow process and organization clockspeeds favor sequential approach
Steel	Sequential	Very slow clockspeeds strongly favor sequential approach
Aircraft (military)	Sequential	Long development times and high costs strongly favor sequential approach
Shipbuilding	Sequential	Very slow clockspeeds strongly favor sequential approach
Petrochemicals	Sequential	Slow clockspeeds and high costs favor sequential approach
Paper	Sequential	Slow clockspeeds favor sequential approach
Electricity	Sequential	Extremely slow clockspeeds strongly favor sequential approach
Diamond mining	Sequential	Extremely slow clockspeeds strongly favor sequential approach

This analysis suggests that industries with faster clockspeeds, particularly in product technology, are more likely to benefit from parallel search strategies. These industries typically have lower costs for testing new ideas (💸), higher uncertainty in both technological innovations and customer preferences (🎲), and potentially lower correlation between the value of technological innovations and customer segments (🧩). Industries with slower clockspeeds and higher development or implementation costs are more likely to benefit from sequential search strategies.

[tomfid]

Interesting that Manzi mentions the BCG learning curve story. One of Ventana's early projects was figuring out Douglas Aircraft lost money every time they introduced a new highly successful plane. The extremely short version of the answer is that they were pricing for learning curves, but neglecting rookie effects from the big increase in hiring needed to scale up. So I think the model's predictive power can be rescued if you add structure to pick up other important features.

[tomfid]

It's hard to expand a model to handle everything in items 1-10 on the list. But I don't think that means you toss out the models and switch to experimentation. You can't afford to do a 2^10 trial either. So as Charlie said, you can use the models to rule out ideas that are universally dumb, find the boundaries of understanding where more information is needed, and design experiments that will be productive.

[tomfid]

theory: SD and probabilistic computing can complement each other

hypothesis1: object and relation can be extracted from text format of SD model xmile and mapped to probabilistic language of thought syntax

hypothesis2: conversational support implemented from hypothesis1 can improve user interface for reality check function, increasing the model quality (as Tom hopes)

hypothesis3: SD's knowhow in 1) eliciting and quantifying uncertainty (table function, stock and flow diagram) across several levels of aggregation, 2) gaming interactive simulation function can facilitate developing PLOT technology

hypothesis4: SD community, capable of formal model building, appreciates quantifying uncertainty can be good lead user of PLOT technology

I particularly like #2. First, writing Reality Checks can be hard (just managing the variables and syntax needed) and it's not very exciting for the user. Having an LLM take a statement like "no inventory -> no sales" and put it into equations would be very helpful. Second, an LLM has a lot of general knowledge that could be used ton construct a bunch of basic RCs for obvious constraints, leaving the more exotic and interesting things for the modeler.

[hyunjimoon]

could you give me three examples of reality check that can be applied to your cwd model (p.15~ from stancon slide)?

[hyunjimoon]

my hypothesis on the environment complexity chooses optimal learning modes (seq or parallel, bayes or evol; this thread + #244 (comment)) earned score💰by Andrew Lo's research, especially "three specific environments favor the emergence of finite memory—those that are Markov, nonstationary, and environments where sampling contains too little or too much information about local conditions." from Lo21_bayes_origin.pdf.

with binary choice mathematical model in
Brennan11_origin of behavior.pdf
and Brennan12_evol_bdd_ration_intell.pdf figure summary:

he explains under nonstationary stochastic environment + evolution induces bayes heuristic with finite memory

---

tangentially, I wonder how andrew's recent paper Brennan18_Mutation.pdf defines optimal degree of irrationality to advocate variety under stochastic environment, can justify startup pivoting as expanding spandrel to increase variety/flexibility.

[hyunjimoon]

---

Scott's two hypotheses on when simulation is not helpful

1. I already know everything. (e.g. Bob Langer knows everything. He has a playbook, and just executes playbook) ~ Sail it stage?
2. Suppose we were to give high school students an entrepreneurship course. Do you think AI assisted tools are helpful for them? They have never understood business. in that situation AI is total disaster, because, because it substitutes for the actual like, if you've never interacted with them,
   to summarize, if you know everything, you don't need AI because you already know what to do. And if you're just learning about how to think about the world, then the AI substitutes for real world experience, and that's also costly. So when is AI most valuable, it's when you have, like, an intermediate level knowledge of a particular idea, and then you're really choosing making a and you want to simulate different alternative scenarios.

Factor	Useful 👍	Less Useful 👎
Knowledge Level on industry	🧠 Intermediate	🎓 Expert-level or 🐣 Complete novice
Knowledge Level on business	🧠 Intermediate	🎓 Expert-level or 🐣 Complete novice
Experience	🌱 Some real-world experience, but still learning	🏆 Vast experience or 📚 No practical experience
Uncertainty	🌪️ Moderate to high uncertainty	🏞️ Very low or extremely high uncertainty
Time Horizon	🔭 Medium to long-term planning	⏱️ Immediate decisions or 🔮 Very long-term visions
Market Scope	🌍 Expanding to adjacent or new customer segments	🏠 Serving well-understood existing customers or 🌌 Completely unknown markets
Company Stage	🚀 Early Scale It (transitioning from Nail It)	🏛️ Late Sail It or 🐣 Very early Nail It
Founder Orientation	🏗️ Balanced approach (some compete, some collaborate)	🤝 Pure collaboration or 💥 Pure disruption
Product-Market Fit	🧩 Refining fit, testing specific hypotheses	✅ Fully validated proposition or ❓ No idea about fit
Industry Clockspeed	⚡ Medium to fast-clockspeed industries	🐢 Very slow-clockspeed or 🚀 Extremely fast-changing industries
Resources	💡 Growing team with some expertise	🏢 Large, highly experienced team or 👤 Solo founder with no resources
Decision Complexity	🕸️ Moderately complex choices with some known variables	🚦 Very simple decisions or 🌪️ Extremely complex, unknowable scenarios
Data Availability	📊 Some data, but gaps in knowledge	📈 Abundant, comprehensive data or 📉 No relevant data at all
Technology Maturity	🔬 Emerging tech with some proven applications	🏭 Fully mature, well-understood tech or 💡 Purely theoretical tech
Customer Behavior	🎭 Evolving with some predictable patterns	🧘 Completely stable and predictable or 🌪️ Totally unpredictable
Competitive Landscape	🏁 Dynamic, with mix of established and new players	🏰 Static, unchanging landscape or 🌋 Chaotic, unpredictable competition