pivot game as mvp to elicit feedback from pc and be

[hyunjimoon]

to elicit promising collaboration between probabilistic computing and bayesian entrepreneurship audience (each of their tech/market in #224 and #234), Matin (Vikash's phd student) and Angie thought pivot game to find product market fit would be nice mvp.

this thread's goal is to log development of pivot game.

feedback from both sides of the audience (be: charlie, scott, jb, (13 entrepreneurial learning scholars at AOM conference who gave me feedback on programmatizing entrepreneurial learning)/ pc: vikash, josh, andrew) is to be logged.

[hyunjimoon]

based on @jeanbaptiste 's suggestion on connecting expatiation with customization + inventing-reinventing from reid hoffman's startup pivoting talk (which sounded like re-configuring) https://claude.ai/chat/bde8a267-3204-4b73-a0c4-4c5de02dce03

Exaptation Concept	Entrepreneurial Customization	Emoji	Pivot Game Design Suggestion
Repurposing existing elements	Firms repurpose resources for novel value creation	🔄	Allow players to repurpose car models (e.g., turn a Luxury Sedan into a Sports model)
Adaptation to new contexts	Companies adapt strategies to new markets/needs	🌱	Introduce market shifts (e.g., sudden demand for Economy Sports cars)
Emergent functionality	Discovering unexpected value propositions	💡	Reveal hidden customer preferences for certain car features
Evolutionary perspective	Iterative experimentation and adaptation	🧬	Implement multiple rounds where car designs evolve based on market feedback
Context-dependent value	Customizing approach to specific market conditions	🎯	Vary rewards for different car types based on changing market conditions
Non-linear innovation	Developing unexpected paths to value creation	🌀	Introduce "wild card" events that create opportunities for unconventional car designs

[hyunjimoon]

@jeanbaptiste how does the above table look for environment design? graphics are as below.

[hyunjimoon]

[hyunjimoon]

logging and walkie-talkie with @mugamma

Action Items during 🪑 chat
[ ] Angie finalize product x market label with Matin by Thursday
[ ] Add visualizations to help understand the game state and player choices
[ ] Continue discussion on induction biases and approximation errors in modeling
[ ] Run experiments to compare performance of different player strategies over many games

Action Items during 👯 walkie-talkie
[ ] Develop a minimum viable product
[ ] Consider adding the pivot game to existing experiments on teaching scientific decision-making versus other approaches.
[ ] Continue the discussion on Slack about next steps regarding the pivot game and connections to longer-term research visions.

🪑chat

The conversation revolved around decision-making and problem-solving in uncertain markets, particularly in the context of entrepreneurship and AI. Speakers discussed various tools and techniques, such as chat GPT and modeling multi-agent interactions, to aid in informed decision-making. They emphasized the importance of understanding context, balancing prior beliefs and data-driven reasoning, and avoiding induction bias. The conversation also touched on the challenges of optimizing decision-making processes in AI product development and the role of prior knowledge and beliefs in reinforcement learning.

Using games to teach entrepreneurs how to think about business scenarios.

Entrepreneurs discuss using AI as thought partners, sharing beliefs and communication with machines.
Participants discuss a game-based approach to understanding firm-specific causal logic and its application in the automotive industry.
matin: Teaching tool for thinking about business scenarios, not automation.
matin: New class of tools help interact with information, like chat GPT.

Decision-making tools for complex problems.

matin discusses using symbolic algebra software like sympy for high-level planning of integrals, but notes that it doesn't perform lower-level calculations like polynomial multiplication and derivative calculation.
angie suggests finding a middle ground between continuous and discontinuous state models, and notes that discounting the continuous state is not an issue.
matin discusses a decision-making problem with a unique structure, where the goal is to maximize value and revenue over time.
matin mentions reinforcement learning as the current state-of-the-art approach for solving similar decision processes, involving neural policies and gradient descent.

Modeling principles for AI product development.

Speakers discuss modeling principles for targeting customers in AI product development.
angie emphasizes the importance of two-sided information in product development, contrasting with the lean startup approach that prioritizes customer feedback.
matin acknowledges the difference in information symmetry between the two sets, but suggests treating them as one for simplicity's sake.

Modeling product market fit with belief vectors and information asymmetry.

Entrepreneurs and markets have different beliefs about product-market fit, which can lead to confusion and misalignment.
The belief net vector of the entrepreneur and the belief network of the market are used to calculate the product-market fit in a simple linear model.
matin: Entrepreneurs have more information about product features than customers.
matin: Asymmetry in beliefs between entrepreneurs and customers exists in both scenarios.

Modeling multiple agents' interactions in a game.

matin: Emphasizes importance of modeling agent interactions in system dynamics.
angie: Proposes using machine learning to create a thought partner for decision-making.
matin explains market sentiment indicator in a game, using Tesla as an example.

Car type and market fit.

The speakers debate which market is better for their luxury sports car or economy sedan product.
Speakers discuss market segmentation for electric vehicles, focusing on range and luxury vs economy.

Reinforcement learning with market and product encoding.

matin explains the game's state and actions, including market and product choices.
matin discusses the reward situation after taking an action, using encoding notation.
Encoding actions as vectors with positive or negative signs affects reward distribution in MDP.

Game dynamics and strategies for maximizing rewards in a bandit problem.

matin explains the reward function in the game, which maximizes dynamic reward up to a finite horizon.
The move representation is discrete and can be printed as a string, with no loss of information.
matin explains bandit problem formulation with correlated arm distributions.

Optimal decision-making in a game with positive and negative outcomes.

The speaker uses Bayesian inference to determine the optimal move in a game, based on the noise of the previous move.
The speaker uses the difference between the previous moves to determine the signs of the two numbers, and selects the move with the highest probability of being the optimal play.
matin describes a simple ad hoc strategy for playing a game, involving adding a reward signal to the value function equation.
matin explains how AlphaGo, a Go-playing AI, is similar to this problem, with larger state sizes but similar concepts.
matin defines value function V and horizon value function HFA.

Reinforcement learning and belief states in a simple game.

angie wants to update the values of the cells based on their previous choices and observations.
matin explains that the beliefs about the value of the cells are correlated, and suggests using a Gaussian to Gaussian belief symbol.
matin: Observing noise in decision-making, like playing 212 games.
angie: Scientific entrepreneurs have weaker priors, willing to reject beliefs for surprising things.

Using Bayesian inference to make decisions in a game with uncertain rewards.

matin describes their prior beliefs in the game, with a high standard deviation but concentrated on positive outcomes.
matin's prior beliefs are influenced by their previous experiences and can change quickly with new data.
matin explains Bellman equation for decision-making in a game.
Reinforcement learning and value functions.
matin explains how the reward function works in the context of reinforcement learning, using the term "R" to represent the immediate reward and "VH" to represent the expected value of the next state.
angie asks questions to clarify the concept of conditional expectation notation, and how it relates to the updated beliefs of the agent.
matin explains value functions for actions in a probabilistic program.
matin discusses the importance of prior distributions in reinforcement learning, specifically in the context of softmax agents and greedy settings.
matin explains how tight prior distributions can lead to less movement in the value function, while loose prior distributions can result in more movement.

Using simulation to understand Bayesian updates and optimal exploration strategies.

angie shared a paper on overconfident bias and its connection to precision and estimation error in machine learning.
matin was interested in learning more about the analogy between precision error and induction bias, and the connection to universal approximation and inference algorithm.
angie suggests using a simple simulation to help develop intuition about Bayes rule (11:51)
matin believes a random NBA student may not have an intuition about Bayes rule (11:34)
matin explains how looking ahead in decision-making can lead to better outcomes (1:13:45)
angie asks for an example of how looking ahead can help in a specific situation (1:14:00)

👯 walkie-talkie

The conversation revolved around various topics in machine learning, reinforcement learning, and entrepreneurship. Speakers discussed the importance of exploration moves in reinforcement learning, the potential benefits and limitations of using neural networks, and the relevance of Bayesian reasoning in entrepreneurship. They also explored the concepts of overconfidence and optimism in decision-making, and the potential of machine learning to elicit beliefs from humans. Throughout the conversation, speakers shared their insights and experiences in academia and the startup world, highlighting the challenges and opportunities in these fields.

Outline

Reinforcement learning and optimal decision-making strategies in various scenarios.

Speaker discusses heuristic strategy for decision-making in a noisy environment.
Neural networks are used to calculate values and pick actions in decision-making processes.
Speakers discuss using neural networks to approximate value functions in reinforcement learning.

Technical debt in machine learning, with examples from AlphaGo and GPT.

matin discusses technical debt in machine learning, particularly in neural networks, and how it can accumulate quickly due to the complexity of the problem and the need for quick solutions.
The speaker argues that technical debt in machine learning can lead to high interest payments in the form of bugs, updates, and retraining, making it essential to pay it off to maintain productivity and deliver deadlines.
Speaker discusses using 80% solution to solve problem, rather than fully solving it.
matin discusses AlphaGo's limitations in certain situations, such as when facing high-level players who don't make a specific move.
matin mentions using GPT to solve a problem related to linear combinations of vectors, but encountering issues with numerical stability and incorrect normalization steps.

Neural networks and probabilistic programming for solving complex problems.

Speakers discuss the importance of understanding the reasoning process and value of a person in AI development.
angie describes neural networks as approximating from outward to inward, relaxing constraints or adding decision variables.
Probabilistic programming and symbolic approaches start with hard answers and gradually refine them, unlike neural networks that start with wrong answers and cut out garbage.
matin explains that neural networks are easier to scale but don't solve problems accurately.
matin discusses how probabilistic programs are harder to scale but provide exact answers for specific problems.

Bayesian reasoning and entrepreneurship.

angie compares probabilistic programming to social science, highlighting differences in precision and scalability.
Speaker discusses potential value of "Bayesian Entrepreneurship" program.
Speakers discuss the relevance of Bayesian and frequentist approaches in entrepreneurship, with a focus on the meaning of probability values.

Bayesian statistics and its applications in business and academia.

Speaker discusses different interpretations of quantum mechanics, including Copenhagen interpretation and parallel universes.
Philosophers debate meaning of probability in quantum mechanics, with some interpreting it as a distribution over states.
matin expresses skepticism about the Bayesian frequency thing, wondering if it applies in this context.
matin wonders if a simple game like this could be used in business courses or research on probabilistic programs.
Using machine learning to improve entrepreneurship education.
angie discusses experiment on scientific vs. natural entrepreneurship, finds that scientific approach leads to better outcomes and fewer pivots.
angie discusses designing machine learning models for collaboration with humans, considering industry-specific beliefs and decision-making.
Speakers discuss the challenge of differentiating optimism from overconfidence in entrepreneurship.

Optimism and overconfidence in AI research.

matin explains overconfidence as a belief state that affects decision-making in a game.
matin discusses optimism in reinforcement learning, mentioning regret minimization algorithm and epsilon GD policy.
matin and angie discuss publishing papers in AI, with matin emphasizing the importance of experimentation and results.
matin expresses uncertainty about staying in academia, citing pros and cons.
matin and 2 discuss the similarities between academic research and startups, with a focus on personal branding and selling ideas.

[jeanbaptiste]

Hi angie,On mobile phone, i reply when at my computer laterCheers! JB¡ǝuoɥdı ʎɯ ɯoɹɟ ʇuǝsLe 20 août 2024 à 08:35, Angie.H Moon ***@***.***> a écrit : @jeanbaptiste how does the above table look for environment design? —Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you were mentioned.Message ID: ***@***.***>