Conference Report

Threshold 2030

Threshold 2030

Modeling AI Economic Futures

Modeling AI Economic Futures

A two-day conference in Oct 2024 bringing together 30 leading economists, AI policy experts, and professional forecasters to rapidly evaluate the economic impacts of frontier AI technologies by 2030.

Proposed Scenarios

Threshold 2030 was centered around scenario modeling: an analytical tool used by policymakers, strategists, and academics to explore and prepare for the landscape of possible outcomes in domains defined by uncertainty. This approach asks researchers to consider a range of possible scenarios with explicitly defined prior assumptions.

This approach of scenario modeling has several key benefits for predicting on highly uncertain futures:

By reducing the space of potential futures, researchers can forecast scenario-specific outcomes with greater precision.

By reducing the space of potential futures, researchers can forecast scenario-specific outcomes with greater precision.

By reducing the space of potential futures, researchers can forecast scenario-specific outcomes with greater precision.

Similarly, this can enable researchers to more effectively evaluate recommendations contingent on specific assumptions.

Similarly, this can enable researchers to more effectively evaluate recommendations contingent on specific assumptions.

Similarly, this can enable researchers to more effectively evaluate recommendations contingent on specific assumptions.

Scenario modeling allows us to compare and contrast outcomes, helping us understand how specific assumptions shape the predictions of experts.

Scenario modeling allows us to compare and contrast outcomes, helping us understand how specific assumptions shape the predictions of experts.

Scenario modeling allows us to compare and contrast outcomes, helping us understand how specific assumptions shape the predictions of experts.

Methodology

Our conference centered evaluation of potential economic futures on three key scenarios, differentiated primarily by the magnitude and rate of AI capabilities development.

At the start of the conference, we asked attendees to make 2 simplifying assumptions:

1

Assume a “default” condition – that is, assume that no major government or economic interventions will occur to shift the trajectory of outcomes.

1

Assume a “default” condition – that is, assume that no major government or economic interventions will occur to shift the trajectory of outcomes.

1

Assume a “default” condition – that is, assume that no major government or economic interventions will occur to shift the trajectory of outcomes.

2

Assume that we are predicting economic states on December 31st, 2030 (~6 years from October 2024) as the date of consequence.

2

Assume that we are predicting economic states on December 31st, 2030 (~6 years from October 2024) as the date of consequence.

2

Assume that we are predicting economic states on December 31st, 2030 (~6 years from October 2024) as the date of consequence.

Next, attendees were asked to consider three unique scenarios, providing a detailed description for each on the capability of AI systems to automate human cognitive labor (via AI models) and human manual labor (via a combination of robotics and AI models).  Attendees were given some time to read and discuss these scenarios in groups of four, and took notes on the thoughts and feedback they brought up.

For the rest of the conference, attendees considered their tasks and created outputs in the context of these scenarios. During the Worldbuilding exercises, attendees used these scenarios to describe three possible economic futures in detail. During the Economic Causal Modeling exercises, attendees created quantitative predictions for specific economic variables based on Scenario 3.

Below are the exact descriptions for the three scenarios that were shared with attendees during Threshold 2030:

Scenarios from Threshold 2030

Scenario 1

Current AI systems, but with improved capabilities in 2030

Overall, AI systems in 2030 are more powerful versions of today’s LLMs, but with similar structural limitations. LLMs still primarily function in response to direction from humans, and do not take the initiative or act independently.

Though LLMs may be better integrated into existing workflows and products, this scenario doesn’t assume significant advances in independent reasoning or end-to-end execution.

Cognitive Labor

Capabilities:

Capabilities:

Capabilities:

LLMs and AI systems in 2030 continue to increase in capabilities, but cannot take on the end-to-end responsibilities of most existing jobs involving cognitive labor.

LLMs and AI systems in 2030 continue to increase in capabilities, but cannot take on the end-to-end responsibilities of most existing jobs involving cognitive labor.

LLMs and AI systems in 2030 continue to increase in capabilities, but cannot take on the end-to-end responsibilities of most existing jobs involving cognitive labor.

LLMs are still too unreliable to string together into unsupervised workflows, and too unreliable to act as good supervisors.

LLMs are still too unreliable to string together into unsupervised workflows, and too unreliable to act as good supervisors.

LLMs are still too unreliable to string together into unsupervised workflows, and too unreliable to act as good supervisors.

For example, LLMs may allow for massive productivity improvements helping web designers generate new websites, but not deliver a cohesive end-to-end experience independently.

For example, LLMs may allow for massive productivity improvements helping web designers generate new websites, but not deliver a cohesive end-to-end experience independently.

For example, LLMs may allow for massive productivity improvements helping web designers generate new websites, but not deliver a cohesive end-to-end experience independently.

For example, LLMs may generate accurate results in radiology 10x faster than humans, but still require the oversight of a trained radiologist to validate the results before delivering to a patient.

For example, LLMs may generate accurate results in radiology 10x faster than humans, but still require the oversight of a trained radiologist to validate the results before delivering to a patient.

For example, LLMs may generate accurate results in radiology 10x faster than humans, but still require the oversight of a trained radiologist to validate the results before delivering to a patient.

AI systems can be considered extremely high-powered calculators that massively improve productivity, but are still just tools used by humans as part of their workflows.

AI systems can be considered extremely high-powered calculators that massively improve productivity, but are still just tools used by humans as part of their workflows.

AI systems can be considered extremely high-powered calculators that massively improve productivity, but are still just tools used by humans as part of their workflows.

A skeptical politician / layperson in 2024 should easily be convinced that AI systems are capable or will soon be capable of these abilities.

A skeptical politician / layperson in 2024 should easily be convinced that AI systems are capable or will soon be capable of these abilities.

A skeptical politician / layperson in 2024 should easily be convinced that AI systems are capable or will soon be capable of these abilities.

Exceptions: There are certain low-skill job roles where AI can replace 90% of human labor (e.g. customer support & truck drivers).

Exceptions: There are certain low-skill job roles where AI can replace 90% of human labor (e.g. customer support & truck drivers).

Exceptions: There are certain low-skill job roles where AI can replace 90% of human labor (e.g. customer support & truck drivers).

Even in these exceptions, the tasks in the top 10% of complexity are still escalated to humans.

Even in these exceptions, the tasks in the top 10% of complexity are still escalated to humans.

Even in these exceptions, the tasks in the top 10% of complexity are still escalated to humans.

Robotics

Capabilities:

Capabilities:

Capabilities:

AI robotics systems allow for the continued automation of manufacturing (e.g. vehicles, electronics) and logistics systems (e.g. warehouses, last-mile delivery, packaging).

AI robotics systems allow for the continued automation of manufacturing (e.g. vehicles, electronics) and logistics systems (e.g. warehouses, last-mile delivery, packaging).

AI robotics systems allow for the continued automation of manufacturing (e.g. vehicles, electronics) and logistics systems (e.g. warehouses, last-mile delivery, packaging).

AI robotics continue to augment and improve productivity for human workflows (e.g. precision surgeries, drones).

AI robotics continue to augment and improve productivity for human workflows (e.g. precision surgeries, drones).

AI robotics continue to augment and improve productivity for human workflows (e.g. precision surgeries, drones).

AI robotics systems do not approach the capabilities of humans in most manual labor roles and are unable to fully automate most workflows. They outperform humans only in specific use-cases.

AI robotics systems do not approach the capabilities of humans in most manual labor roles and are unable to fully automate most workflows. They outperform humans only in specific use-cases.

AI robotics systems do not approach the capabilities of humans in most manual labor roles and are unable to fully automate most workflows. They outperform humans only in specific use-cases.

Scenario 2

Powerful, narrow AI systems that outperform humans on 95% of well-scoped tasks

AI systems achieve better results than people in most constrained or well-scoped tasks. However, they fail to outperform humans in task integration, handling multifaceted responsibilities, and communication with other humans. They still require oversight.

Cognitive Labor

Timing:

Timing:

Timing:

Powerful, narrow AI systems that outperform humans are steadily released and productized across a variety of domains from Jan 2027 onwards until 2030.

Powerful, narrow AI systems that outperform humans are steadily released and productized across a variety of domains from Jan 2027 onwards until 2030.

Powerful, narrow AI systems that outperform humans are steadily released and productized across a variety of domains from Jan 2027 onwards until 2030.

Capabilities:

Capabilities:

Capabilities:

AI systems significantly outperform humans on most well-scoped tasks, such as copywriting, legal writing, design, data analysis, financial modeling, debugging coding systems, etc.

AI systems significantly outperform humans on most well-scoped tasks, such as copywriting, legal writing, design, data analysis, financial modeling, debugging coding systems, etc.

AI systems significantly outperform humans on most well-scoped tasks, such as copywriting, legal writing, design, data analysis, financial modeling, debugging coding systems, etc.

AI systems may outperform humans on certain research tasks - e.g. creating AI system optimizations, or designing proteins / vaccines with certain characteristics.

AI systems may outperform humans on certain research tasks - e.g. creating AI system optimizations, or designing proteins / vaccines with certain characteristics.

AI systems may outperform humans on certain research tasks - e.g. creating AI system optimizations, or designing proteins / vaccines with certain characteristics.

AI systems may be able to handle each of the independent parts of a job autonomously, but have a hard time combining different modalities of work.

AI systems may be able to handle each of the independent parts of a job autonomously, but have a hard time combining different modalities of work.

AI systems may be able to handle each of the independent parts of a job autonomously, but have a hard time combining different modalities of work.

Humans continue to outperform AIs at multifaceted, context-sensitive tasks.

Humans continue to outperform AIs at multifaceted, context-sensitive tasks.

Humans continue to outperform AIs at multifaceted, context-sensitive tasks.

For example, AI may be superior at legal analysis, but a single AI system may not be able to integrate client conversations, legal analysis, document creation, and contract negotiation.

For example, AI may be superior at legal analysis, but a single AI system may not be able to integrate client conversations, legal analysis, document creation, and contract negotiation.

For example, AI may be superior at legal analysis, but a single AI system may not be able to integrate client conversations, legal analysis, document creation, and contract negotiation.

Humans are largely required to “tie” different aspects of a job together - serving as liaisons between AI systems and other humans.

Humans are largely required to “tie” different aspects of a job together - serving as liaisons between AI systems and other humans.

Humans are largely required to “tie” different aspects of a job together - serving as liaisons between AI systems and other humans.

They act as safety and reasoning checks to integrate tasks that are individually well-handled by powerful AI systems.

They act as safety and reasoning checks to integrate tasks that are individually well-handled by powerful AI systems.

They act as safety and reasoning checks to integrate tasks that are individually well-handled by powerful AI systems.

AI systems rely on humans for context, integration, and communicating with other humans.

AI systems rely on humans for context, integration, and communicating with other humans.

AI systems rely on humans for context, integration, and communicating with other humans.

Human-AI interaction necessitates trained human specialists that can liaison and use AI systems efficiently.

Human-AI interaction necessitates trained human specialists that can liaison and use AI systems efficiently.

Human-AI interaction necessitates trained human specialists that can liaison and use AI systems efficiently.

AI can’t work easily with external customers or untrained teammates.

AI can’t work easily with external customers or untrained teammates.

AI can’t work easily with external customers or untrained teammates.

Human job roles often require deep knowledge of the intricacies and limitations of powerful, specialized AI systems.

Human job roles often require deep knowledge of the intricacies and limitations of powerful, specialized AI systems.

Human job roles often require deep knowledge of the intricacies and limitations of powerful, specialized AI systems.

Final decision making is still entrusted to human specialists.

Final decision making is still entrusted to human specialists.

Final decision making is still entrusted to human specialists.

Robotics

Timing:

Timing:

Timing:

Powerful, narrow robotics systems that outperform humans in specific tasks are steadily released and productized across a variety of domains from Jan 2027 onwards until the present day (Jan 2030).

Powerful, narrow robotics systems that outperform humans in specific tasks are steadily released and productized across a variety of domains from Jan 2027 onwards until the present day (Jan 2030).

Powerful, narrow robotics systems that outperform humans in specific tasks are steadily released and productized across a variety of domains from Jan 2027 onwards until the present day (Jan 2030).

Diffusion for robotics (manual) capabilities is significantly slower than diffusion of cognitive capabilities.

Diffusion for robotics (manual) capabilities is significantly slower than diffusion of cognitive capabilities.

Diffusion for robotics (manual) capabilities is significantly slower than diffusion of cognitive capabilities.

Capabilities:

Capabilities:

Capabilities:

AI robotics systems can match or outperform humans at nearly any well-scoped task (manufacturing, boxing, operating machinery, flying a plane).

AI robotics systems can match or outperform humans at nearly any well-scoped task (manufacturing, boxing, operating machinery, flying a plane).

AI robotics systems can match or outperform humans at nearly any well-scoped task (manufacturing, boxing, operating machinery, flying a plane).

AI systems have a very difficult time dealing with ambiguity or changing roles. A single AI system is not very adaptable to novel inputs.

AI systems have a very difficult time dealing with ambiguity or changing roles. A single AI system is not very adaptable to novel inputs.

AI systems have a very difficult time dealing with ambiguity or changing roles. A single AI system is not very adaptable to novel inputs.

Human overrides may still be required in complex situations. Lack of comprehensive trust in AI systems (e.g. medical, judicial domains) means that AI systems need human oversight.

Human overrides may still be required in complex situations. Lack of comprehensive trust in AI systems (e.g. medical, judicial domains) means that AI systems need human oversight.

Human overrides may still be required in complex situations. Lack of comprehensive trust in AI systems (e.g. medical, judicial domains) means that AI systems need human oversight.

AI systems still cannot effectively manipulate inputs and controls designed for humans (e.g. the instrument panel for a 747).

AI systems still cannot effectively manipulate inputs and controls designed for humans (e.g. the instrument panel for a 747).

AI systems still cannot effectively manipulate inputs and controls designed for humans (e.g. the instrument panel for a 747).

Scenario 3

Powerful, general AI systems that outperform humans on all forms of cognitive labor

Powerful AI systems can meet and surpass the performance of humans in all dimensions of cognitive labor, and can function as “drop-in” replacements for nearly all human jobs.

Cognitive Labor

Timing:

Timing:

Timing:

The first such powerful / general AI systems were announced in Jan 2027.

The first such powerful / general AI systems were announced in Jan 2027.

The first such powerful / general AI systems were announced in Jan 2027.

However, there is still a multi-year delay (diffusion) to productize and polish these AI systems to be deployed for specific job roles.

However, there is still a multi-year delay (diffusion) to productize and polish these AI systems to be deployed for specific job roles.

However, there is still a multi-year delay (diffusion) to productize and polish these AI systems to be deployed for specific job roles.

Capabilities:

Capabilities:

Capabilities:

AI has the capability of outperforming humans on all cognitive jobs end-to-end.

AI has the capability of outperforming humans on all cognitive jobs end-to-end.

AI has the capability of outperforming humans on all cognitive jobs end-to-end.

AI systems can communicate with untrained humans effectively. They can act as experts to understand the needs and requests of others, and deliver expert-quality output.

AI systems can communicate with untrained humans effectively. They can act as experts to understand the needs and requests of others, and deliver expert-quality output.

AI systems can communicate with untrained humans effectively. They can act as experts to understand the needs and requests of others, and deliver expert-quality output.

AI systems can function as a standalone member of a highly productive / capable team. They can function as a “drop-in coworker”.

AI systems can function as a standalone member of a highly productive / capable team. They can function as a “drop-in coworker”.

AI systems can function as a standalone member of a highly productive / capable team. They can function as a “drop-in coworker”.

They can be a standalone project manager, engineer, or lawyer, outperforming humans on all tasks.

They can be a standalone project manager, engineer, or lawyer, outperforming humans on all tasks.

They can be a standalone project manager, engineer, or lawyer, outperforming humans on all tasks.

AI systems can rapidly adapt and learn new skills & roles, given the correct training.

AI systems can rapidly adapt and learn new skills & roles, given the correct training.

AI systems can rapidly adapt and learn new skills & roles, given the correct training.

AI systems massively accelerate innovation and scientific discovery, as they have comparable (& significantly better) capabilities than leading human scientific researchers.

AI systems massively accelerate innovation and scientific discovery, as they have comparable (& significantly better) capabilities than leading human scientific researchers.

AI systems massively accelerate innovation and scientific discovery, as they have comparable (& significantly better) capabilities than leading human scientific researchers.

Robotics

Timing:

Timing:

Timing:

The first such robotics systems integrating general AI capabilities were announced in Jan 2028.

The first such robotics systems integrating general AI capabilities were announced in Jan 2028.

The first such robotics systems integrating general AI capabilities were announced in Jan 2028.

Diffusion for robotics (manual) capabilities is significantly slower than diffusion of cognitive capabilities.

Diffusion for robotics (manual) capabilities is significantly slower than diffusion of cognitive capabilities.

Diffusion for robotics (manual) capabilities is significantly slower than diffusion of cognitive capabilities.

Capabilities:

Capabilities:

Capabilities:

Robotics AI systems leveraging powerful general AI systems can perform on par with humans using cameras and microphones to replicate human input.

Robotics AI systems leveraging powerful general AI systems can perform on par with humans using cameras and microphones to replicate human input.

Robotics AI systems leveraging powerful general AI systems can perform on par with humans using cameras and microphones to replicate human input.

Robotics AI systems have similar general capabilities to cognitive AI systems.

Robotics AI systems have similar general capabilities to cognitive AI systems.

Robotics AI systems have similar general capabilities to cognitive AI systems.

Robotics systems can manipulate human inputs on par with humans. That is, they can drive a car, operate a bulldozer, or fly a plane that has not been significantly re-designed to allow accessibility by humans.

Robotics systems can manipulate human inputs on par with humans. That is, they can drive a car, operate a bulldozer, or fly a plane that has not been significantly re-designed to allow accessibility by humans.

Robotics systems can manipulate human inputs on par with humans. That is, they can drive a car, operate a bulldozer, or fly a plane that has not been significantly re-designed to allow accessibility by humans.

AI systems can deal with ambiguity as well as humans. Human overrides are largely unnecessary.

AI systems can deal with ambiguity as well as humans. Human overrides are largely unnecessary.

AI systems can deal with ambiguity as well as humans. Human overrides are largely unnecessary.

Attendee Responses

We’ll share a number of key thoughts and feedback from each of the groups regarding the above scenarios. Broadly, attendees were aligned with the high-level premise of the scenarios as described. Most of the discussion centered around methods to expand or concretize aspects of the scenarios, such as describing topics that were underdefined or absent.

An attendee in Group 1 suggested that even Scenario 3 as described above might be conservative in terms of describing the impacts of transformative AI capabilities. They suggested that AI self-improvement, particularly around AI research, may lead to significantly greater impact and risks.

Another attendee in Group 1 suggested that these scenarios were narrowly focused on human task replacement and automation, rather than potential novel task opportunities for AI systems. For instance, AI systems could integrate expertise across a variety of domains, providing new roles compared to humans (who are typically experts in only 1-2 domains).

Group 5 discussed the challenges in achieving human-level dexterity in robotics by 2030. They suggested that tasks designed specifically as sending software signals to existing machines (e.g. integrating an AI system into a car) would be significantly more automatable than tasks that may require more dexterity (e.g. carving a wood block).

Several attendees across groups had questions about what set of futures would fall under Scenario 2. One group wished that terms such as “well-scoped” or “narrow” were more precisely defined. Some attendees found Scenario 2 to be too similar to Scenario 3, while others thought that Scenario 2 was already the minimum possible set of AI capabilities we might expect in 2030.

Many groups discussed that the applicability of AI technologies, as well as the rate of diffusion, will be highly uneven based on the domain. Group 1 discussed differences in tolerance to errors: for instance, coding might be fine with a 10% tolerance to error, whereas flying a plane would have significantly less tolerance. Group 4 discussed the impacts of vertical integration in the supply chain, for instance in military applications. Group 5 mentioned the legal and structural differences between startups and established organizations.

Finally, multiple groups discussed the relevance of compute costs to these proposed scenarios. Group 4 mentioned that the cost of these AI models would strongly impact adoption, and that further investment in AI systems might depend heavily on revenue from adoption. Notably, they asked: “Will these AI models be cheaper than human equivalents?”