Conference Report
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:
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:
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
Robotics
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
Robotics
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
Robotics
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?”