- entrepreneurship as experimentationl (dollars invested into bin (1, 1~3, 3~5, 5~10, >10)) - portfolio, risk adjusted return (random thinking and blow) - scores that vc partners assigned to each venture at the time of their first investment predict poorly which investments are successful ### abandon option of belief value is created via staged finance - first step of SAFE can be framed as "experiment". we use this idea to design conversational financial value inference tool that can be used for deal structuring. it optimizes potential of applying conversational inference, inspired by sequential Monte Carlo methods, to business negotiations and investment decisions. This approach involves decomposing complex problems into manageable sub-models, allowing for iterative updating of beliefs based on new information. It addresses resource-constrained rationality by providing tools to overcome cognitive limitations, enabling participants to tap into conditional beliefs and check consistency across different decision spaces. The process can be particularly valuable in persuading investors by breaking down gut feelings into more rational sub-beliefs and structuring deals that are mutually beneficial. By representing these models using probabilistic programming languages, it becomes possible to simulate and infer complex scenarios, ultimately leading to more informed and rational decision-making in entrepreneurial finance. decision theory (rational agency), probabilistic program's programmable inference, bayesian entrepreneurship's optimal stopping rule 1. analyze how having conversational inference interface itself increasesm value, just like how multistage financing's abandon option which rejects parameters whose prior predictions is outside boundary of one's belief itself adds value 2. connect information relaxation, simulation-based calibration (which is better btw bayesian calibration), test two choose one 3. how inferring financial value can drive action to optimize utility of founder? 4. persuading investor using iterative inference through designing proposal algorithm (annealed importance sampling, stratified sampling) q. how would hypothesis space complexity affect ADEV effectiveness? (dEU/dx - reverse mode vs forward mode) 1. conversational inference: 2. equity allocation with adev 3. updated generate_heatmap_data to include both founder_utility and safe_utility 4. add calculate_utility function 1. Complexity of the connection: The connection between conversational inference and sequential Monte Carlo methods is more nuanced and less direct than the table suggests. The transcripts reveal that this connection is still an area of active research and not fully understood. 2. Role of probabilistic programming: The transcripts emphasize the importance of probabilistic programming languages like Gen in implementing these methods, which wasn't mentioned in the original table. 3. Application to human-AI interaction: There's an interesting focus on applying these methods to model interactions between humans and AI systems, particularly in the context of negotiation and shared belief systems. 4. Challenges in implementation: The discussion highlights several challenges in implementing these methods, such as parameter tuning and the need for custom algorithms, which weren't apparent from the original table. Based on these surprises, here's how I'd update the table: | Level | Name | Question | Conversational Inference Example | | ----- | ---------------------------- | ------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | 1 | Computational theory | - What are the inputs and outputs to the computation?<br>- What is its goal?<br>- What is the logic by which it is carried out? | **Inputs**: Probabilistic models of beliefs, observations from interactions, shared knowledge base.<br><br>**Outputs**: Updated belief distributions, inferred intentions of other agents, optimal actions in negotiation contexts.<br><br>**Goal**: To model and update beliefs in interactive settings, particularly for negotiation and shared decision-making scenarios.<br><br>**Logic**: Combines sequential Monte Carlo methods with probabilistic programming to iteratively update beliefs based on new information and interactions. Aims to efficiently explore high-dimensional spaces of possible beliefs and intentions. | | 2 | Representation and algorithm | - How is information represented?<br>- How is information processed to achieve the computational goal? | **Representation**: Uses probabilistic programs (e.g., in Gen) to represent complex belief structures and interaction dynamics. Employs state-space models to capture temporal aspects of beliefs and actions.<br><br>**Processing**: Utilizes sampling-based algorithms for approximate inference (e.g., MCMC, sequential Monte Carlo, importance sampling). Employs cost-sensitive sampling and fast initialization with bottom-up recognition models. In specific, after assuming resource-rational models, we relax resource constrains using natural language interface as conversational inference which can be understood as relaxing MCMC to sequential Monte Carlo methods or row generation of adding constraints to converge to feasible solution (dual algorithm). The process involves breaking down gut feelings into rational sub-beliefs and structuring mutually beneficial deals by overcoming cognitive limitations and tap into conditional beliefs. May incorporate techniques from inverse planning and game theory to model multi-agent scenarios. | | 3 | Hardware implementation | - How is the computation realized in physical or biological hardware? | Implemented using probabilistic programming frameworks like Gen, potentially leveraging parallel processing capabilities of modern CPUs/GPUs. May require significant computational resources for complex models. Implementation details are still an active area of research, with ongoing work on using automatic differentiation of expected value for practical applications. | | Level | Name | Question | Conversational Inference Example | | ----- | ---------------------------- | ------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | 1 | Computational theory | - What are the inputs and outputs to the computation?<br>- What is its goal?<br>- What is the logic by which it is carried out? | **Inputs**: Probabilistic program, language model, investment scenarios, assumptions about funding benefits/drawbacks, ownership preferences, growth projections.<br><br>**Outputs**: Visualizations of investment impact, trade-off evaluations between ownership and growth, alignment strategies for investment with contrarian beliefs.<br><br>**Goal**: To implement conversational inference for rational meaning construction in business negotiations and investment decisions.<br><br>**Logic**: Combines probabilistic programming with language models to simulate various investment scenarios, evaluate financial implications, and test assumptions. Uses scenario modeling to assess investment impact, evaluate trade-offs, and align investment strategies with company principles. | | 2 | Representation and algorithm | - How is information represented?<br>- How is information processed to achieve the computational goal? | **Representation**: Complex problems are decomposed into manageable sub-models. Beliefs are represented probabilistically, allowing for iterative updating based on new information.<br><br>**Processing**: | | 3 | Hardware implementation | - How is the computation realized in physical or biological hardware? | The computation is implemented using:<br>- MIT inference stack (e.g., GenParse, GenSQL)<br>- Sequential Monte Carlo steering<br>- Embedding ignorance prior in language models<br><br>These tools are likely realized on standard computational hardware such as CPUs and potentially GPUs for parallel processing, leveraging high-performance computing infrastructure to execute complex probabilistic models and language processing tasks. |