Current Research
Current Research
Transcendence: Generative Models Can Outperform The Experts That Train Them
Transcendence: Generative Models Can Outperform The Experts That Train Them
Transcendence: Generative Models Can Outperform The Experts That Train Them
Founding recently worked with the Harvard Kempner Institute to outline the conditions necessary for transcendence in generative models, in which artificial models outperform the human experts they were trained on. Our research is further explained in this website.
Founding recently worked with the Harvard Kempner Institute to outline the conditions necessary for transcendence in generative models, in which artificial models outperform the human experts they were trained on. Our research is further explained in this website.
Founding recently worked with the Harvard Kempner Institute to outline the conditions necessary for transcendence in generative models, in which artificial models outperform the human experts they were trained on. Our research is further explained in this website.
Social Environment
Design
Social Environment
Design
Social Environment
Design
Founding is currently engaged in advancing research into Social Environment Design (SED), a novel approach to leveraging AI for automated policymaking described more completely in this website. Learn more about our research below. If you'd like to help us develop SED capacities or learn about relevant work in this space, see the SED Contributor Framework.
Founding is currently engaged in advancing research into Social Environment Design (SED), a novel approach to leveraging AI for automated policymaking described more completely in this website. Learn more about our research below. If you'd like to help us develop SED capacities or learn about relevant work in this space, see the SED Contributor Framework.
Founding is currently engaged in advancing research into Social Environment Design (SED), a novel approach to leveraging AI for automated policymaking described more completely in this website. Learn more about our research below. If you'd like to help us develop SED capacities or learn about relevant work in this space, see the SED Contributor Framework.
SED Process
AI holds promise as a technology that can be used to dramatically improve government and economic policy-making. Founding is pursuing a specific research agenda towards this end by introducing Social Environment Design (SED), a general framework for the use of AI for automated policy-making. The framework extends mechanism design to capture a fully general economic environment, including voting on policy objectives, and gives a direction for the systematic analysis of government and economic policy through AI simulation.
A position paper more particularly describing SED is under review and will be posted on arXiv in the coming weeks. At a high level, the process contains three primary steps as follows (see diagram):
The process begins with voting, where human or AI players report preferences on social welfare objectives to a voting mechanism (1). This defines an objective for the Principal, who designs a mechanism (2) that parameterizes an N-player Partially Observable Markov Game (POMG). The players are the same as the voters. This POMG unfolds over several timesteps T (3). Following the POMG, game state information is extracted to initiate a new round of voting, with the last POMG state used as the first game state of the new round. This whole process is repeated for τ timesteps.
Research Agenda
We believe that Social Environment Design should be further studied and deployed broadly in society. Towards this end, we are currently pursuing research and actively seeking collaboration towards:
Aggregating preferences and democratic representation in voting mechanisms
How can we ensure that social welfare objectives reflect collective preferences while still respecting minority views? What is fair in this context?
How can we ensure that preference elicitation is effective across diverse cultural, ethical, and socioeconomic paradigms found in the real world, in data-sparse contexts?
Effective Human Oversight and AI Accountability
What standards must be met to ensure AI decisions can be audited, and how do we achieve those standards?
How should AI be allowed to affect the real world, and who is responsible for effects of AI-driven decisions?
Exploring socioeconomic interactions within these systems
What should be the goal of the Principal agent given that it is unlikely for SED optimizations to converge to conventional game-theoretic equilibria?
How can we ensure the Principal agent can react-to and stabilize socioeconomic shocks that will occur in the continual-learning environment of the real-world?
Analyzing empirical scaling laws
Can the framework handle simulating economies with thousands or millions of agents?
What is the role of scale?
When is simulation useful and when does it fail?
Implementation
By solving these challenges, Founding hopes to extend the capacities of SED implementations in pursuit of more effective and pluralistic governments and social organizations. As we continue research in this direction, we intend to implement the system as a service for product and service optimization at nonprofits and government agencies. The data gathered from this work will allow for the creation of more effective simulacra of these entities’ beneficiaries, improving feedback and the general function of SED for social use cases en route to our final goal.
SED Process
AI holds promise as a technology that can be used to dramatically improve government and economic policy-making. Founding is pursuing a specific research agenda towards this end by introducing Social Environment Design (SED), a general framework for the use of AI for automated policy-making. The framework extends mechanism design to capture a fully general economic environment, including voting on policy objectives, and gives a direction for the systematic analysis of government and economic policy through AI simulation.
A position paper more particularly describing SED is under review and will be posted on arXiv in the coming weeks. At a high level, the process contains three primary steps as follows (see diagram):
The process begins with voting, where human or AI players report preferences on social welfare objectives to a voting mechanism (1). This defines an objective for the Principal, who designs a mechanism (2) that parameterizes an N-player Partially Observable Markov Game (POMG). The players are the same as the voters. This POMG unfolds over several timesteps T (3). Following the POMG, game state information is extracted to initiate a new round of voting, with the last POMG state used as the first game state of the new round. This whole process is repeated for τ timesteps.
Research Agenda
We believe that Social Environment Design should be further studied and deployed broadly in society. Towards this end, we are currently pursuing research and actively seeking collaboration towards:
Aggregating preferences and democratic representation in voting mechanisms
How can we ensure that social welfare objectives reflect collective preferences while still respecting minority views? What is fair in this context?
How can we ensure that preference elicitation is effective across diverse cultural, ethical, and socioeconomic paradigms found in the real world, in data-sparse contexts?
Effective Human Oversight and AI Accountability
What standards must be met to ensure AI decisions can be audited, and how do we achieve those standards?
How should AI be allowed to affect the real world, and who is responsible for effects of AI-driven decisions?
Exploring socioeconomic interactions within these systems
What should be the goal of the Principal agent given that it is unlikely for SED optimizations to converge to conventional game-theoretic equilibria?
How can we ensure the Principal agent can react-to and stabilize socioeconomic shocks that will occur in the continual-learning environment of the real-world?
Analyzing empirical scaling laws
Can the framework handle simulating economies with thousands or millions of agents?
What is the role of scale?
When is simulation useful and when does it fail?
Implementation
By solving these challenges, Founding hopes to extend the capacities of SED implementations in pursuit of more effective and pluralistic governments and social organizations. As we continue research in this direction, we intend to implement the system as a service for product and service optimization at nonprofits and government agencies. The data gathered from this work will allow for the creation of more effective simulacra of these entities’ beneficiaries, improving feedback and the general function of SED for social use cases en route to our final goal.
SED Process
AI holds promise as a technology that can be used to dramatically improve government and economic policy-making. Founding is pursuing a specific research agenda towards this end by introducing Social Environment Design (SED), a general framework for the use of AI for automated policy-making. The framework extends mechanism design to capture a fully general economic environment, including voting on policy objectives, and gives a direction for the systematic analysis of government and economic policy through AI simulation.
A position paper more particularly describing SED is under review and will be posted on arXiv in the coming weeks. At a high level, the process contains three primary steps as follows (see diagram):
The process begins with voting, where human or AI players report preferences on social welfare objectives to a voting mechanism (1). This defines an objective for the Principal, who designs a mechanism (2) that parameterizes an N-player Partially Observable Markov Game (POMG). The players are the same as the voters. This POMG unfolds over several timesteps T (3). Following the POMG, game state information is extracted to initiate a new round of voting, with the last POMG state used as the first game state of the new round. This whole process is repeated for τ timesteps.
Research Agenda
We believe that Social Environment Design should be further studied and deployed broadly in society. Towards this end, we are currently pursuing research and actively seeking collaboration towards:
Aggregating preferences and democratic representation in voting mechanisms
How can we ensure that social welfare objectives reflect collective preferences while still respecting minority views? What is fair in this context?
How can we ensure that preference elicitation is effective across diverse cultural, ethical, and socioeconomic paradigms found in the real world, in data-sparse contexts?
Effective Human Oversight and AI Accountability
What standards must be met to ensure AI decisions can be audited, and how do we achieve those standards?
How should AI be allowed to affect the real world, and who is responsible for effects of AI-driven decisions?
Exploring socioeconomic interactions within these systems
What should be the goal of the Principal agent given that it is unlikely for SED optimizations to converge to conventional game-theoretic equilibria?
How can we ensure the Principal agent can react-to and stabilize socioeconomic shocks that will occur in the continual-learning environment of the real-world?
Analyzing empirical scaling laws
Can the framework handle simulating economies with thousands or millions of agents?
What is the role of scale?
When is simulation useful and when does it fail?
Implementation
By solving these challenges, Founding hopes to extend the capacities of SED implementations in pursuit of more effective and pluralistic governments and social organizations. As we continue research in this direction, we intend to implement the system as a service for product and service optimization at nonprofits and government agencies. The data gathered from this work will allow for the creation of more effective simulacra of these entities’ beneficiaries, improving feedback and the general function of SED for social use cases en route to our final goal.
SED Process
AI holds promise as a technology that can be used to dramatically improve government and economic policy-making. Founding is pursuing a specific research agenda towards this end by introducing Social Environment Design (SED), a general framework for the use of AI for automated policy-making. The framework extends mechanism design to capture a fully general economic environment, including voting on policy objectives, and gives a direction for the systematic analysis of government and economic policy through AI simulation.
A position paper more particularly describing SED is under review and will be posted on arXiv in the coming weeks. At a high level, the process contains three primary steps as follows (see diagram):
The process begins with voting, where human or AI players report preferences on social welfare objectives to a voting mechanism (1). This defines an objective for the Principal, who designs a mechanism (2) that parameterizes an N-player Partially Observable Markov Game (POMG). The players are the same as the voters. This POMG unfolds over several timesteps T (3). Following the POMG, game state information is extracted to initiate a new round of voting, with the last POMG state used as the first game state of the new round. This whole process is repeated for τ timesteps.
Research Agenda
We believe that Social Environment Design should be further studied and deployed broadly in society. Towards this end, we are currently pursuing research and actively seeking collaboration towards:
Aggregating preferences and democratic representation in voting mechanisms
How can we ensure that social welfare objectives reflect collective preferences while still respecting minority views? What is fair in this context?
How can we ensure that preference elicitation is effective across diverse cultural, ethical, and socioeconomic paradigms found in the real world, in data-sparse contexts?
Effective Human Oversight and AI Accountability
What standards must be met to ensure AI decisions can be audited, and how do we achieve those standards?
How should AI be allowed to affect the real world, and who is responsible for effects of AI-driven decisions?
Exploring socioeconomic interactions within these systems
What should be the goal of the Principal agent given that it is unlikely for SED optimizations to converge to conventional game-theoretic equilibria?
How can we ensure the Principal agent can react-to and stabilize socioeconomic shocks that will occur in the continual-learning environment of the real-world?
Analyzing empirical scaling laws
Can the framework handle simulating economies with thousands or millions of agents?
What is the role of scale?
When is simulation useful and when does it fail?
Implementation
By solving these challenges, Founding hopes to extend the capacities of SED implementations in pursuit of more effective and pluralistic governments and social organizations. As we continue research in this direction, we intend to implement the system as a service for product and service optimization at nonprofits and government agencies. The data gathered from this work will allow for the creation of more effective simulacra of these entities’ beneficiaries, improving feedback and the general function of SED for social use cases en route to our final goal.