Garry Tan proposes that open-weight AI labs in the United States use knowledge distillation from frontier models to maintain competitiveness, reduce dependence on closed models, and give developers more options.
But this idea is not only about performance. It also raises questions about copyright, safety, and the right to use the outputs for further training. If done without clear boundaries, the labs that own the models could lose their business advantage, while developers could face the resulting risks.
Garry Tan proposes that open-weight AI labs in the United States use knowledge distillation from frontier models to maintain competitiveness, reduce dependence on closed models, and give developers more options.
But this idea is not only about performance. It also raises questions about copyright, safety, and the right to use the outputs for further training. If done without clear boundaries, the labs that own the models could lose their business advantage, while developers could face the resulting risks.
An Image That Helps Explain the Game
This image illustrates a large frontier model transferring its capabilities to a smaller open-weight model, set against the backdrop of AI competition between the United States and China. It shows that this game is not measured only by who can build the more capable model, but also by who can make those capabilities more broadly usable.
An Image That Helps Explain the Game
This image illustrates a large frontier model transferring its capabilities to a smaller open-weight model, set against the backdrop of AI competition between the United States and China. It shows that this game is not measured only by who can build the more capable model, but also by who can make those capabilities more broadly usable.
When API Costs Rise but Teams Still Need AI
Imagine a developer or startup that has to call a closed model every day. API costs keep rising, sensitive data has to leave the system, and the team cannot fully customize the model for its own work.
That is why Garry Tan wants to see AI labs in the United States distill capabilities from frontier models into open-weight models. The question is whether this approach will genuinely make AI more accessible to small teams—or simply shift the burden from API fees to infrastructure costs.
When API Costs Rise but Teams Still Need AI
Imagine a developer or startup that has to call a closed model every day. API costs keep rising, sensitive data has to leave the system, and the team cannot fully customize the model for its own work.
That is why Garry Tan wants to see AI labs in the United States distill capabilities from frontier models into open-weight models. The question is whether this approach will genuinely make AI more accessible to small teams—or simply shift the burden from API fees to infrastructure costs?
Where Garry Tan’s Proposal Fits on the AI Map
As an executive at Y Combinator, Garry Tan believes that the United States should not allow expensive closed frontier models to remain the only primary option. Instead, it should distill those capabilities into open-weight models that different teams can use and customize themselves.
This proposal sits between the convenience of paid services and the freedom of models that can be run independently. It also connects to the strategic competition between the United States and China, because open-weight models allow companies, developers, and organizations to build their own systems more quickly.
Where Garry Tan’s Proposal Fits on the AI Map
As an executive at Y Combinator, Garry Tan believes that the United States should not allow expensive closed frontier models to remain the only primary option. Instead, it should distill those capabilities into open-weight models that different teams can use and customize themselves.
This proposal sits between the convenience of paid services and the freedom of models that can be run independently. It also connects to the strategic competition between the United States and China, because open-weight models allow companies, developers, and organizations to build their own systems more quickly.
From Expensive Closed Models to Open Models That Can Be Extended
| Factor | Closed models via API | Distilled open-weight models |
|---|---|---|
| Access | Used through the provider’s service | Download the model weights and use them independently |
| Long-term cost | Service fees based on usage | Costs can be reduced when run on your own systems |
| Customization | Adjustments are limited by the API | Greater ability to customize and inspect |
This approach transforms the outputs of frontier models into smaller models suited to specialized tasks. Companies can therefore choose whether to pay API fees for convenience or run models themselves for greater control over their data and systems.
From Expensive Closed Models to Open Models That Can Be Extended
| Factor | Closed models via API | Distilled open-weight models |
|---|---|---|
| Access | Used through the provider’s service | Download the model weights and use them independently |
| Long-term cost | Service fees based on usage | Costs can be reduced when run on your own systems |
| Customization | Adjustments are limited by the API | Greater ability to customize and inspect |
This approach transforms the outputs of frontier models into smaller models suited to specialized tasks. Companies can therefore choose whether to pay API fees for convenience or run models themselves for greater control over their data and systems.
How Knowledge Distillation Could Change the Lives of AI Users
As models become smaller, small businesses gain more options for deploying them on their own servers. Lower per-request costs also help startups experiment with AI features more frequently without worrying about excessive budgets.
Offline operation is suitable for organizations handling sensitive data because information does not have to be sent outside the organization. Meanwhile, open weights allow researchers and developers to adapt models more precisely to specific languages, tasks, and industries.
How Knowledge Distillation Could Change the Lives of AI Users
As models become smaller, small businesses gain more options for deploying them on their own servers. Lower per-request costs also help startups experiment with AI features more frequently without worrying about excessive budgets.
Offline operation is suitable for organizations handling sensitive data because information does not have to be sent outside the organization. Meanwhile, open weights allow researchers and developers to adapt models more precisely to specific languages, tasks, and industries.
Three Options for Responding to Model Distillation
The decision is not simply about being “open” or “closed.” It requires balancing competition, safety, and the rights of model owners.
| Factor | Ban or restrict | Allow freely | Permit under conditions |
|---|---|---|---|
| Goal | Protect intellectual property | Accelerate competition | Balance both sides |
| Strength | Reduce the risk of models being misused | Help reduce monopolization | Allow research to continue |
| Limitation | Could slow development | Risk knowledge theft or unauthorized use of systems | Research and business criteria must be clearly separated |
Three Options for Responding to Model Distillation
The decision is not simply about being “open” or “closed.” It requires balancing competition, safety, and the rights of model owners.
| Factor | Ban or restrict | Allow freely | Permit under conditions |
|---|---|---|---|
| Goal | Protect intellectual property | Accelerate competition | Balance both sides |
| Strength | Reduce the risk of models being misused | Help reduce monopolization | Allow research to continue |
| Limitation | Could slow development | Risk knowledge theft or unauthorized use of systems | Research and business criteria must be clearly separated |
The Strengths of This Proposal—and What Could Harm the Industry
The idea of having open-weight AI labs distill capabilities from frontier models could increase competition and reduce development costs. Smaller companies and researchers would have more opportunities to build models for specialized tasks, and power would be less concentrated among major labs.
But the risks include models inheriting biases and vulnerabilities from their originals in ways that are difficult to inspect. Copyright also remains ambiguous, especially when companies aggressively extract capabilities from closed systems without taking responsibility for the consequences.
Pros
- +Increase competition in the AI industry
- +Reduce costs and create opportunities for small teams
- +Make it easier to build models for specialized tasks
Cons
- −May transfer biases and vulnerabilities
- −The copyright status of the knowledge remains unclear
- −Risk aggressively extracting capabilities from closed systems without accountability
The Strengths of This Proposal—and What Could Harm the Industry
The idea of having open-weight AI labs distill capabilities from frontier models could increase competition and reduce development costs. Smaller companies and researchers would have more opportunities to build models for specialized tasks, and power would be less concentrated among major labs.
But the risks include models inheriting biases and vulnerabilities from their originals in ways that are difficult to inspect. Copyright also remains ambiguous, especially when companies aggressively extract capabilities from closed systems without taking responsibility for the consequences.
Pros
- +Increase competition in the AI industry
- +Reduce costs and create opportunities for small teams
- +Make it easier to build models for specialized tasks
Cons
- −May transfer biases and vulnerabilities
- −The copyright status of the knowledge remains unclear
- −Risk aggressively extracting capabilities from closed systems without accountability
Costs That Do Not Appear on the Model’s Price Tag
Distillation involves more than GPU and energy costs. There are also hosting expenses, safety evaluation costs, and the labor required to maintain the model after it is deployed.
If the training data is not transparent, teams may face additional copyright risks and legal expenses. More importantly, inexpensive models can create social costs if they are used for fraud, system attacks, or harming others; being “good value” must therefore include accountability across the entire system.
Costs That Do Not Appear on the Model’s Price Tag
Distillation involves more than GPU and energy costs. There are also hosting expenses, safety evaluation costs, and the labor required to maintain the model after it is deployed.
If the training data is not transparent, teams may face additional copyright risks and legal expenses. More importantly, inexpensive models can create social costs if they are used for fraud, system attacks, or harming others; being “good value” must therefore include accountability across the entire system.
The Final Question Is Not Simply Who Gets There First
The key question may not be who builds the smartest model, but who creates an ecosystem that is genuinely open and responsible. Distilling knowledge from frontier models could help open-weight labs compete, but clear rules are needed regarding copyright, safety, and subsequent use.
If opening models removes the incentive for frontier-model owners to continue investing, everyone loses. Ultimately, the winner may be the lab that strikes a balance between developer choice, social responsibility, and a sustainable business.
The Final Question Is Not Simply Who Gets There First
The key question may not be who builds the smartest model, but who creates an ecosystem that is genuinely open and responsible. Distilling knowledge from frontier models could help open-weight labs compete, but clear rules are needed regarding copyright, safety, and subsequent use.
If opening models removes the incentive for frontier-model owners to continue investing, everyone loses. Ultimately, the winner may be the lab that strikes a balance between developer choice, social responsibility, and a sustainable business.
An Image That Helps Explain the Game
This image should depict the path from a large frontier model transferring its capabilities to a smaller open-weight model that developers can access and extend more easily, with a background conveying AI competition between the United States and China.
An Image That Helps Explain the Game
This image should depict the path from a large frontier model transferring its capabilities to a smaller open-weight model that developers can access and extend more easily, with a background conveying AI competition between the United States and China.
When API Costs Rise but Teams Still Need AI
Many developers and startups depend on closed models even as API costs continue to rise. Sensitive data still has to leave their systems, and teams cannot fully customize the models.
That makes Garry Tan’s proposal compelling: have open-weight AI labs in the United States distill knowledge from frontier models and make it available for others to extend. The question is whether this would genuinely make AI more accessible or simply move the cost to model training instead.
When API Costs Rise but Teams Still Need AI
Many developers and startups depend on closed models even as API costs continue to rise. Sensitive data still has to leave their systems, and teams cannot fully customize the models.
That makes Garry Tan’s proposal compelling: have open-weight AI labs in the United States distill knowledge from frontier models and make it available for others to extend. The question is whether this would genuinely make AI more accessible or simply move the cost to model training instead?
Where Garry Tan’s Proposal Fits on the AI Map
Garry Tan, an executive at Y Combinator, proposes that open-weight AI labs in the United States distill knowledge from frontier models and make it available for others to develop further. This position lies between closed models that are capable but expensive to access and open-weight models that organizations can deploy and adapt themselves.
In the broader picture, this is not merely about reducing costs. It also concerns strategic competition between the United States and China. If the United States has models that many people can access and extend, it could build a stronger AI ecosystem. However, the cost of training models remains an unavoidable barrier.
Where Garry Tan’s Proposal Fits on the AI Map
Garry Tan, an executive at Y Combinator, proposes that open-weight AI labs in the United States distill knowledge from frontier models and make it available for others to develop further. This position lies between closed models that are capable but expensive to access and open-weight models that organizations can deploy and adapt themselves.
In the broader picture, this is not merely about reducing costs. It also concerns strategic competition between the United States and China. If the United States has models that many people can access and extend, it could build a stronger AI ecosystem. However, the cost of training models remains an unavoidable barrier.
From Expensive Closed Models to Open Models That Can Be Extended
| Factor | Frontier model API | Distilled open-weight model |
|---|---|---|
| Access | Used through an API | Download and run independently |
| Cost and control | Depends on the provider | Can be adapted to the task and available resources |
| Inspectability | Primarily see the outputs | Greater ability to inspect and customize |
This approach changes the role of frontier models from services that must be called into sources of knowledge for creating smaller and cheaper models, suitable for organizations that want to control their own data and systems.
To be direct, the strength of open-weight models is not merely saving money. It is the ability for others to inspect, customize, and extend the models more deeply for real-world tasks.
From Expensive Closed Models to Open Models That Can Be Extended
| Factor | Frontier model API | Distilled open-weight model |
|---|---|---|
| Access | Used through an API | Download and run independently |
| Cost and control | Depends on the provider | Can be adapted to the task and available resources |
| Inspectability | Primarily see the outputs | Greater ability to inspect and customize |
This approach changes the role of frontier models from services that must be called into sources of knowledge for creating smaller and cheaper models, suitable for organizations that want to control their own data and systems.
To be direct, the strength of open-weight models is not merely saving money. It is the ability for others to inspect, customize, and extend the models more deeply for real-world tasks.
How Knowledge Distillation Could Change the Lives of AI Users
As models become smaller, small businesses have more opportunities to install AI on their own servers instead of constantly relying on external systems. Lower per-request costs also help startups experiment with new features more frequently and adjust course more quickly.
Tasks involving sensitive information, such as internal documents or customer data, can be run offline, reducing concerns about data leaks. At the same time, open model weights allow researchers and developers to adapt models more precisely to specific languages, tasks, and industries.
How Knowledge Distillation Could Change the Lives of AI Users
As models become smaller, small businesses have more opportunities to install AI on their own servers instead of constantly relying on external systems. Lower per-request costs also help startups experiment with new features more frequently and adjust course more quickly.
Tasks involving sensitive information, such as internal documents or customer data, can be run offline, reducing concerns about data leaks. At the same time, open model weights allow researchers and developers to adapt models more precisely to specific languages, tasks, and industries.
Three Options for Responding to Model Distillation
Banning or restricting distillation helps protect intellectual property and safety, but it may slow competition. Allowing it freely accelerates innovation and reduces monopolization, but increases the risk of models being misused.
| Factor | Restrict distillation | Allow free distillation | Permit under conditions |
|---|---|---|---|
| Intellectual property | Greater protection | Risk of imitation | Rights separated by use case |
| Competition | May slow down | Can accelerate significantly | Can progress within a framework |
| Safety | Easier to control | Must accept greater risks | Unauthorized access to systems is prohibited |
| Research | More limited | Open-ended | Permitted when conducted for research |
The third option attempts to preserve balance by separating research from commercial use and model theft.
Three Options for Responding to Model Distillation
Banning or restricting distillation helps protect intellectual property and safety, but it may slow competition. Allowing it freely accelerates innovation and reduces monopolization, but increases the risk of models being misused.
| Factor | Restrict distillation | Allow free distillation | Permit under conditions |
|---|---|---|---|
| Intellectual property | Greater protection | Risk of imitation | Rights separated by use case |
| Competition | May slow down | Can accelerate significantly | Can progress within a framework |
| Safety | Easier to control | Must accept greater risks | Unauthorized access to systems is prohibited |
| Research | More limited | Open-ended | Permitted when conducted for research |
The third option attempts to preserve balance by separating research from commercial use and model theft.
The Strengths of This Proposal—and What Could Harm the Industry
Having AI labs in the United States distill frontier models could increase competition, reduce costs, and distribute power beyond a small number of companies. Developers would also find it easier to build models suited to specialized tasks.
The risks are that models could inherit biases and vulnerabilities from their sources in ways that are difficult to inspect. Copyright remains ambiguous, and companies may aggressively extract capabilities from closed systems without taking responsibility for the consequences.
Pros
- +Increase competition and reduce costs
- +Build models suited to specialized tasks
Cons
- −May transfer biases and vulnerabilities
- −Ambiguous copyright and insufficient accountability
The Strengths of This Proposal—and What Could Harm the Industry
Having AI labs in the United States distill frontier models could increase competition, reduce costs, and distribute power beyond a small number of companies. Developers would also find it easier to build models suited to specialized tasks.
The risks are that models could inherit biases and vulnerabilities from their sources in ways that are difficult to inspect. Copyright remains ambiguous, and companies may aggressively extract capabilities from closed systems without taking responsibility for the consequences.
Pros
- +Increase competition and reduce costs
- +Build models suited to specialized tasks
Cons
- −May transfer biases and vulnerabilities
- −Ambiguous copyright and insufficient accountability
Costs That Do Not Appear on the Model’s Price Tag
Cheaper models do not mean that costs have disappeared. GPU usage, energy, hosting, model-maintenance labor, and safety evaluations before deployment are still required.
If the training data is not transparent, copyright risks and lawsuits may still follow. Social costs must also be considered if the model is used to create deceptive content, attack people, or support harmful activities; the model’s listed price is only one part of the total cost.
Costs That Do Not Appear on the Model’s Price Tag
Cheaper models do not mean that costs have disappeared. GPU usage, energy, hosting, model-maintenance labor, and safety evaluations before deployment are still required.
If the training data is not transparent, copyright risks and lawsuits may still follow. Social costs must also be considered if the model is used to create deceptive content, attack people, or support harmful activities; the model’s listed price is only one part of the total cost.
The Final Question Is Not Simply Who Gets There First
The competition over distillation may not end with who builds the smartest model, but with who makes it genuinely possible for others—including developers, businesses, and everyday users—to build upon it.
A healthy ecosystem must be open enough to foster innovation, provide accountability that can be verified, and preserve the incentive for frontier-model teams to continue investing in research. If everyone benefits from sharing without destroying the original source, this competition could move the industry forward more sustainably.
The Final Question Is Not Simply Who Gets There First
The competition over distillation may not end with who builds the smartest model, but with who makes it genuinely possible for others—including developers, businesses, and everyday users—to build upon it.
A healthy ecosystem must be open enough to foster innovation, provide accountability that can be verified, and preserve the incentive for frontier-model teams to continue investing in research. If everyone benefits from sharing without destroying the original source, this competition could move the industry forward more sustainably. Garry Tan proposes that open-weight AI labs in the United States use knowledge distillation from frontier models to maintain competitiveness, reduce dependence on closed models, and give developers more options.
But this idea is not only about performance. It also raises questions about copyright, safety, and the right to use the outputs for further training. If done without clear boundaries, the labs that own the models could lose their business advantage, while developers could face the resulting risks.
Garry Tan proposes that open-weight AI labs in the United States use knowledge distillation from frontier models to maintain competitiveness, reduce dependence on closed models, and give developers more options.
But this idea is not only about performance. It also raises questions about copyright, safety, and the right to use the outputs for further training. If done without clear boundaries, the labs that own the models could lose their business advantage, while developers could face the resulting risks.
An Image That Helps Explain the Game
This image illustrates a large frontier model transferring its capabilities to a smaller open-weight model, set against the backdrop of AI competition between the United States and China. It shows that this game is not measured only by who can build the more capable model, but also by who can make those capabilities more broadly usable.
An Image That Helps Explain the Game
This image illustrates a large frontier model transferring its capabilities to a smaller open-weight model, set against the backdrop of AI competition between the United States and China. It shows that this game is not measured only by who can build the more capable model, but also by who can make those capabilities more broadly usable.
When API Costs Rise but Teams Still Need AI
Imagine a developer or startup that has to call a closed model every day. API costs keep rising, sensitive data has to leave the system, and the team cannot fully customize the model for its own work.
That is why Garry Tan wants to see AI labs in the United States distill capabilities from frontier models into open-weight models. The question is whether this approach will genuinely make AI more accessible to small teams—or simply shift the burden from API fees to infrastructure costs.
When API Costs Rise but Teams Still Need AI
Imagine a developer or startup that has to call a closed model every day. API costs keep rising, sensitive data has to leave the system, and the team cannot fully customize the model for its own work.
That is why Garry Tan wants to see AI labs in the United States distill capabilities from frontier models into open-weight models. The question is whether this approach will genuinely make AI more accessible to small teams—or simply shift the burden from API fees to infrastructure costs?
Where Garry Tan’s Proposal Fits on the AI Map
As an executive at Y Combinator, Garry Tan believes that the United States should not allow expensive closed frontier models to remain the only primary option. Instead, it should distill those capabilities into open-weight models that different teams can use and customize themselves.
This proposal sits between the convenience of paid services and the freedom of models that can be run independently. It also connects to the strategic competition between the United States and China, because open-weight models allow companies, developers, and organizations to build their own systems more quickly.
Where Garry Tan’s Proposal Fits on the AI Map
As an executive at Y Combinator, Garry Tan believes that the United States should not allow expensive closed frontier models to remain the only primary option. Instead, it should distill those capabilities into open-weight models that different teams can use and customize themselves.
This proposal sits between the convenience of paid services and the freedom of models that can be run independently. It also connects to the strategic competition between the United States and China, because open-weight models allow companies, developers, and organizations to build their own systems more quickly.
From Expensive Closed Models to Open Models That Can Be Extended
| Factor | Closed models via API | Distilled open-weight models |
|---|---|---|
| Access | Used through the provider’s service | Download the model weights and use them independently |
| Long-term cost | Service fees based on usage | Costs can be reduced when run on your own systems |
| Customization | Adjustments are limited by the API | Greater ability to customize and inspect |
This approach transforms the outputs of frontier models into smaller models suited to specialized tasks. Companies can therefore choose whether to pay API fees for convenience or run models themselves for greater control over their data and systems.
From Expensive Closed Models to Open Models That Can Be Extended
| Factor | Closed models via API | Distilled open-weight models |
|---|---|---|
| Access | Used through the provider’s service | Download the model weights and use them independently |
| Long-term cost | Service fees based on usage | Costs can be reduced when run on your own systems |
| Customization | Adjustments are limited by the API | Greater ability to customize and inspect |
This approach transforms the outputs of frontier models into smaller models suited to specialized tasks. Companies can therefore choose whether to pay API fees for convenience or run models themselves for greater control over their data and systems.
How Knowledge Distillation Could Change the Lives of AI Users
As models become smaller, small businesses gain more options for deploying them on their own servers. Lower per-request costs also help startups experiment with AI features more frequently without worrying about excessive budgets.
Offline operation is suitable for organizations handling sensitive data because information does not have to be sent outside the organization. Meanwhile, open weights allow researchers and developers to adapt models more precisely to specific languages, tasks, and industries.
How Knowledge Distillation Could Change the Lives of AI Users
As models become smaller, small businesses gain more options for deploying them on their own servers. Lower per-request costs also help startups experiment with AI features more frequently without worrying about excessive budgets.
Offline operation is suitable for organizations handling sensitive data because information does not have to be sent outside the organization. Meanwhile, open weights allow researchers and developers to adapt models more precisely to specific languages, tasks, and industries.
Three Options for Responding to Model Distillation
The decision is not simply about being “open” or “closed.” It requires balancing competition, safety, and the rights of model owners.
| Factor | Ban or restrict | Allow freely | Permit under conditions |
|---|---|---|---|
| Goal | Protect intellectual property | Accelerate competition | Balance both sides |
| Strength | Reduce the risk of models being misused | Help reduce monopolization | Allow research to continue |
| Limitation | Could slow development | Risk knowledge theft or unauthorized use of systems | Research and business criteria must be clearly separated |
Three Options for Responding to Model Distillation
The decision is not simply about being “open” or “closed.” It requires balancing competition, safety, and the rights of model owners.
| Factor | Ban or restrict | Allow freely | Permit under conditions |
|---|---|---|---|
| Goal | Protect intellectual property | Accelerate competition | Balance both sides |
| Strength | Reduce the risk of models being misused | Help reduce monopolization | Allow research to continue |
| Limitation | Could slow development | Risk knowledge theft or unauthorized use of systems | Research and business criteria must be clearly separated |
The Strengths of This Proposal—and What Could Harm the Industry
The idea of having open-weight AI labs distill capabilities from frontier models could increase competition and reduce development costs. Smaller companies and researchers would have more opportunities to build models for specialized tasks, and power would be less concentrated among major labs.
But the risks include models inheriting biases and vulnerabilities from their originals in ways that are difficult to inspect. Copyright also remains ambiguous, especially when companies aggressively extract capabilities from closed systems without taking responsibility for the consequences.
Pros
- +Increase competition in the AI industry
- +Reduce costs and create opportunities for small teams
- +Make it easier to build models for specialized tasks
Cons
- −May transfer biases and vulnerabilities
- −The copyright status of the knowledge remains unclear
- −Risk aggressively extracting capabilities from closed systems without accountability
The Strengths of This Proposal—and What Could Harm the Industry
The idea of having open-weight AI labs distill capabilities from frontier models could increase competition and reduce development costs. Smaller companies and researchers would have more opportunities to build models for specialized tasks, and power would be less concentrated among major labs.
But the risks include models inheriting biases and vulnerabilities from their originals in ways that are difficult to inspect. Copyright also remains ambiguous, especially when companies aggressively extract capabilities from closed systems without taking responsibility for the consequences.
Pros
- +Increase competition in the AI industry
- +Reduce costs and create opportunities for small teams
- +Make it easier to build models for specialized tasks
Cons
- −May transfer biases and vulnerabilities
- −The copyright status of the knowledge remains unclear
- −Risk aggressively extracting capabilities from closed systems without accountability
Costs That Do Not Appear on the Model’s Price Tag
Distillation involves more than GPU and energy costs. There are also hosting expenses, safety evaluation costs, and the labor required to maintain the model after it is deployed.
If the training data is not transparent, teams may face additional copyright risks and legal expenses. More importantly, inexpensive models can create social costs if they are used for fraud, system attacks, or harming others; being “good value” must therefore include accountability across the entire system.
Costs That Do Not Appear on the Model’s Price Tag
Distillation involves more than GPU and energy costs. There are also hosting expenses, safety evaluation costs, and the labor required to maintain the model after it is deployed.
If the training data is not transparent, teams may face additional copyright risks and legal expenses. More importantly, inexpensive models can create social costs if they are used for fraud, system attacks, or harming others; being “good value” must therefore include accountability across the entire system.
The Final Question Is Not Simply Who Gets There First
The key question may not be who builds the smartest model, but who creates an ecosystem that is genuinely open and responsible. Distilling knowledge from frontier models could help open-weight labs compete, but clear rules are needed regarding copyright, safety, and subsequent use.
If opening models removes the incentive for frontier-model owners to continue investing, everyone loses. Ultimately, the winner may be the lab that strikes a balance between developer choice, social responsibility, and a sustainable business.
The Final Question Is Not Simply Who Gets There First
The key question may not be who builds the smartest model, but who creates an ecosystem that is genuinely open and responsible. Distilling knowledge from frontier models could help open-weight labs compete, but clear rules are needed regarding copyright, safety, and subsequent use.
If opening models removes the incentive for frontier-model owners to continue investing, everyone loses. Ultimately, the winner may be the lab that strikes a balance between developer choice, social responsibility, and a sustainable business.
An Image That Helps Explain the Game
This image should depict the path from a large frontier model transferring its capabilities to a smaller open-weight model that developers can access and extend more easily, with a background conveying AI competition between the United States and China.
An Image That Helps Explain the Game
This image should depict the path from a large frontier model transferring its capabilities to a smaller open-weight model that developers can access and extend more easily, with a background conveying AI competition between the United States and China.
When API Costs Rise but Teams Still Need AI
Many developers and startups depend on closed models even as API costs continue to rise. Sensitive data still has to leave their systems, and teams cannot fully customize the models.
That makes Garry Tan’s proposal compelling: have open-weight AI labs in the United States distill knowledge from frontier models and make it available for others to extend. The question is whether this would genuinely make AI more accessible or simply move the cost to model training instead.
When API Costs Rise but Teams Still Need AI
Many developers and startups depend on closed models even as API costs continue to rise. Sensitive data still has to leave their systems, and teams cannot fully customize the models.
That makes Garry Tan’s proposal compelling: have open-weight AI labs in the United States distill knowledge from frontier models and make it available for others to extend. The question is whether this would genuinely make AI more accessible or simply move the cost to model training instead?
Where Garry Tan’s Proposal Fits on the AI Map
Garry Tan, an executive at Y Combinator, proposes that open-weight AI labs in the United States distill knowledge from frontier models and make it available for others to develop further. This position lies between closed models that are capable but expensive to access and open-weight models that organizations can deploy and adapt themselves.
In the broader picture, this is not merely about reducing costs. It also concerns strategic competition between the United States and China. If the United States has models that many people can access and extend, it could build a stronger AI ecosystem. However, the cost of training models remains an unavoidable barrier.
Where Garry Tan’s Proposal Fits on the AI Map
Garry Tan, an executive at Y Combinator, proposes that open-weight AI labs in the United States distill knowledge from frontier models and make it available for others to develop further. This position lies between closed models that are capable but expensive to access and open-weight models that organizations can deploy and adapt themselves.
In the broader picture, this is not merely about reducing costs. It also concerns strategic competition between the United States and China. If the United States has models that many people can access and extend, it could build a stronger AI ecosystem. However, the cost of training models remains an unavoidable barrier.
From Expensive Closed Models to Open Models That Can Be Extended
| Factor | Frontier model API | Distilled open-weight model |
|---|---|---|
| Access | Used through an API | Download and run independently |
| Cost and control | Depends on the provider | Can be adapted to the task and available resources |
| Inspectability | Primarily see the outputs | Greater ability to inspect and customize |
This approach changes the role of frontier models from services that must be called into sources of knowledge for creating smaller and cheaper models, suitable for organizations that want to control their own data and systems.
To be direct, the strength of open-weight models is not merely saving money. It is the ability for others to inspect, customize, and extend the models more deeply for real-world tasks.
From Expensive Closed Models to Open Models That Can Be Extended
| Factor | Frontier model API | Distilled open-weight model |
|---|---|---|
| Access | Used through an API | Download and run independently |
| Cost and control | Depends on the provider | Can be adapted to the task and available resources |
| Inspectability | Primarily see the outputs | Greater ability to inspect and customize |
This approach changes the role of frontier models from services that must be called into sources of knowledge for creating smaller and cheaper models, suitable for organizations that want to control their own data and systems.
To be direct, the strength of open-weight models is not merely saving money. It is the ability for others to inspect, customize, and extend the models more deeply for real-world tasks.
How Knowledge Distillation Could Change the Lives of AI Users
As models become smaller, small businesses have more opportunities to install AI on their own servers instead of constantly relying on external systems. Lower per-request costs also help startups experiment with new features more frequently and adjust course more quickly.
Tasks involving sensitive information, such as internal documents or customer data, can be run offline, reducing concerns about data leaks. At the same time, open model weights allow researchers and developers to adapt models more precisely to specific languages, tasks, and industries.
How Knowledge Distillation Could Change the Lives of AI Users
As models become smaller, small businesses have more opportunities to install AI on their own servers instead of constantly relying on external systems. Lower per-request costs also help startups experiment with new features more frequently and adjust course more quickly.
Tasks involving sensitive information, such as internal documents or customer data, can be run offline, reducing concerns about data leaks. At the same time, open model weights allow researchers and developers to adapt models more precisely to specific languages, tasks, and industries.
Three Options for Responding to Model Distillation
Banning or restricting distillation helps protect intellectual property and safety, but it may slow competition. Allowing it freely accelerates innovation and reduces monopolization, but increases the risk of models being misused.
| Factor | Restrict distillation | Allow free distillation | Permit under conditions |
|---|---|---|---|
| Intellectual property | Greater protection | Risk of imitation | Rights separated by use case |
| Competition | May slow down | Can accelerate significantly | Can progress within a framework |
| Safety | Easier to control | Must accept greater risks | Unauthorized access to systems is prohibited |
| Research | More limited | Open-ended | Permitted when conducted for research |
The third option attempts to preserve balance by separating research from commercial use and model theft.
Three Options for Responding to Model Distillation
Banning or restricting distillation helps protect intellectual property and safety, but it may slow competition. Allowing it freely accelerates innovation and reduces monopolization, but increases the risk of models being misused.
| Factor | Restrict distillation | Allow free distillation | Permit under conditions |
|---|---|---|---|
| Intellectual property | Greater protection | Risk of imitation | Rights separated by use case |
| Competition | May slow down | Can accelerate significantly | Can progress within a framework |
| Safety | Easier to control | Must accept greater risks | Unauthorized access to systems is prohibited |
| Research | More limited | Open-ended | Permitted when conducted for research |
The third option attempts to preserve balance by separating research from commercial use and model theft.
The Strengths of This Proposal—and What Could Harm the Industry
Having AI labs in the United States distill frontier models could increase competition, reduce costs, and distribute power beyond a small number of companies. Developers would also find it easier to build models suited to specialized tasks.
The risks are that models could inherit biases and vulnerabilities from their sources in ways that are difficult to inspect. Copyright remains ambiguous, and companies may aggressively extract capabilities from closed systems without taking responsibility for the consequences.
Pros
- +Increase competition and reduce costs
- +Build models suited to specialized tasks
Cons
- −May transfer biases and vulnerabilities
- −Ambiguous copyright and insufficient accountability
The Strengths of This Proposal—and What Could Harm the Industry
Having AI labs in the United States distill frontier models could increase competition, reduce costs, and distribute power beyond a small number of companies. Developers would also find it easier to build models suited to specialized tasks.
The risks are that models could inherit biases and vulnerabilities from their sources in ways that are difficult to inspect. Copyright remains ambiguous, and companies may aggressively extract capabilities from closed systems without taking responsibility for the consequences.
Pros
- +Increase competition and reduce costs
- +Build models suited to specialized tasks
Cons
- −May transfer biases and vulnerabilities
- −Ambiguous copyright and insufficient accountability
Costs That Do Not Appear on the Model’s Price Tag
Cheaper models do not mean that costs have disappeared. GPU usage, energy, hosting, model-maintenance labor, and safety evaluations before deployment are still required.
If the training data is not transparent, copyright risks and lawsuits may still follow. Social costs must also be considered if the model is used to create deceptive content, attack people, or support harmful activities; the model’s listed price is only one part of the total cost.
Costs That Do Not Appear on the Model’s Price Tag
Cheaper models do not mean that costs have disappeared. GPU usage, energy, hosting, model-maintenance labor, and safety evaluations before deployment are still required.
If the training data is not transparent, copyright risks and lawsuits may still follow. Social costs must also be considered if the model is used to create deceptive content, attack people, or support harmful activities; the model’s listed price is only one part of the total cost.
The Final Question Is Not Simply Who Gets There First
The competition over distillation may not end with who builds the smartest model, but with who makes it genuinely possible for others—including developers, businesses, and everyday users—to build upon it.
A healthy ecosystem must be open enough to foster innovation, provide accountability that can be verified, and preserve the incentive for frontier-model teams to continue investing in research. If everyone benefits from sharing without destroying the original source, this competition could move the industry forward more sustainably.
The Final Question Is Not Simply Who Gets There First
The competition over distillation may not end with who builds the smartest model, but with who makes it genuinely possible for others—including developers, businesses, and everyday users—to build upon it.
A healthy ecosystem must be open enough to foster innovation, provide accountability that can be verified, and preserve the incentive for frontier-model teams to continue investing in research. If everyone benefits from sharing without destroying the original source, this competition could move the industry forward more sustainably.