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.
All Economic Causal Model Notes
Group 1: Growth
Commonalities, Cruxes, Uncertainties
Write a 1-paragraph reflection and summary of the model you’ve created.
Attendee 1A
The hardware-software model (Hardware and software: A new perspective on the past and future of economic growth) is a high-level representation of the production process. It assumes that output is produced through (i) purposefully initiated (ii) physical action. Therefore it considers two key factors of production: hardware and software, instead of the traditional capital-labor dichotomy. It assumes that both hardware and software are mutually complementary and essential in production. Hardware consists of physical capital and human physical labor. Software includes human cognitive work and digital software, including AI. Hardware and software are used both in final production and R&D.
Zooming into the factors, one can uncover a multiplicity of factors affecting the technology level as well as the supply of software factors.
Attendee 2D
Scaling-focused model: Focuses on breaking down the change in the economy in terms of three major factors: <1. General capability of deployed models. 2. How economically “useful” that capability is. 3. And the cost of using that deployed AI.> Progress in AI drives especially (1), application of AI drives (2), and technological investment drives (3). As HLMI in 2030 would most likely depend on “compute scaling”, this model is hence focused on what appears to be important factors in driving the relevance and power of such compute scaling.
Attendee 5B
TFP Submodels: Two parts of the bigger Application/economic effects model attempting to show the factors driving TFP via knowledge accumulation and human capital quality.
In human capital quality submodel, the key inputs are skill matches (do humans have the right skills to use AI well?), education, and management quality. In the latter case, this is not just human management but potentially AI supervision (of humans, or other AIs). If AI supervision becomes useful (likely in Scenario 3) it requires decent social ability from the AI and ways of getting work information into the system (and for training), requiring monitoring and measuring systems. Human acceptance of this is a key variable that might go in many directions.
AI reliability is a key driver for the need for human supervision of AI, and AI ability to act as a superviser for tasks. Benchmarking it is likely a very useful thing, and might be measured by checking how well AI handles long task chains before going off the track.
What are the most important cruxes you’ve found from this work? Answers to what questions would most help you resolve predictions from your economic model?
Attendee 1A
Frontier technological progress can be expected to proceed much faster than adoption and diffusion into production and R&D. Specifically, even drastic improvements in AI capabilities will not immediately transpire into TFP growth, because of partial automation that keeps humans in the loop, also people’s reluctance to adopt and legal/social constraints.
However, use of AI in R&D, and particularly AI R&D, can lead to a positive feedback loop. A cascade of recursive self-improvements leads to intelligence explosion, with scary implications (for real, not just on Halloween!).
What was the area of most uncertainty in your model?
Attendee 1A
There can be a lot more interactions than we have highlighted. The effects can be highly nonlinear. Some of the nonlinearities are described by the structure of the hardware-software framework, but there are also others. For example, there is a clear link between AI quality and market structure or capital allocation across firms. There is also a clear link between robotics and human labor allocation; AI quality and R&D efficiency.
Attendee 2D
There are important uncertainties in exactly what capability AI has now and will have in the future, and how far adaptation/last-mile-traveling will go at each point. Further “scaling” may be extremely powerful in the next 5 years, or maybe it will be somewhat useful but very very expensive.
What are the most key important things?
Regulatory complexity might be quite crucial
Attendee 1A
It is a huge area of uncertainty - whether scenarios like Scenario 3 will involve an AI takeover, or humankind will remain in control. Under intelligence explosion, the likelihood of AI takeover is high.
What measurables, if you had access to, would help you predict the value of your top-level metric most effectively?
Attendee 5B
TFP is a total vanilla metric. Is there anything in that literature that would be useful for AI? There should be existing work that should be repurposable to some extent.
Attendee 3D
One predictor for TFP in 2030 would be TFP in 2027. Is there something that we’d prefer knowing about 2027?
Attendee 1A
Even if we assume drastic changes in technologies, the previous year will still be heavily predictive of the TFP level.
Attendee 2D
Scaling focused model
Speculative metrics:
Then, you’d need some way of working from those factors to figure out the impacts on the economy. An additional model.
Group 2: Growth
Commonalities, Cruxes, Uncertainties
Write a 1-paragraph reflection and summary of the model you’ve created.
Initially we focussed on GDP-ish, but in the afternoon we shifted to one giant driver of GDP-ish: diffusion of AI. Even if all cognitive tasks are automatable, there are still important factors that determine whether they will be automated. We’ve therefore focused on “Share of AI-automated cognitive work by 2030”. There’s lots of factors that slow down diffusion. Our team focused on India, with its growing young population, less developed AI industry, and great variation in citizens’ circumstances. There are also reasons diffusion could be faster in India; size of service industry, for example.
What are the most important cruxes you’ve found from this work? Answers to what questions would most help you resolve predictions from your economic model?
The top level nodes are the most important nodes, by definition, but these get a lot of attention. We focused on diffusion as a crux for GDP-ish. For the things than fan out from diffusion, the most important sub-cruxes, those that give the strongest signal are:
What was the area of most uncertainty in your model?
The geopolitical node was very uncertain. Big impact scenarios are hard to predict and account for, but will be important.
Capabilities need to be in publicly available models to have big impact. If frontier models are monopolized by, e.g. DoD, we could be in scenario 3 but it’d feel like scenario 1 for most people.
We have uncertainty in how much trust matters. Specifically, trust by the public vs trust by the regulators & governments. The former can largely be manufactured. If capabilities are high enough, profits will find a way to convince people. However, one big bad event could undermine adoption. Is nuclear power a good reference class? Well, there were other alternatives to nuclear power but are there alternatives to automating the economy?
Diffusion will also be messy. Things will be chaotic. People initially didn’t care about the privacy implications of the web - not because there were good privacy practices but because people just didn’t think about it. Trust might deteriorate.
What measurables, if you had access to, would help you predict the value of your top-level metric most effectively?
Not the most important factor, but countries will be interested in: To what extent can good intentioned, pro-AI governments with lots of funding to allocate make a difference to diffusion? Where should resources be allocated?
Cost of AI is very important and needs to be measured. Inference might not just get cheaper and cheaper - O1 was an uptick. We expect it to go down, but we can’t be certain. Knowing this would be influential.
Energy use is also important and impacts cost, but we don’t know loads about it.
If in 3 years there was a complete overhaul of drug approval to accommodate basic breakthroughs from Alphafold, that’d be a big update about where the world is going in terms of diffusion. That’s a very conservative, highly regulated environment. It’s less so a driver of diffusion and more of a real-world observable signal that diffusion is happening. If it doesn’t happen, is this a huge missed opportunity? Can hype drive diffusion? If the public knows some AI research has developed a superdrug but it’s blocked by regulation, there’ll be huge pressure to skip regulation. More generally, where it’s legible to the public that there’s some hugely helpful thing that they don’t have access to, there’ll be huge political pressure to access it.
How good are open source models? Is the ecosystem thriving and they don’t go misused, that’ll hugely impact diffusion. It’ll encourage specialization, experimentation, less reliance on big actors etc. Very influential.
What are the topics that members of your table disagreed on the most?
Not much disagreement! Mostly exploratory discussion.
One small point of disagreement is the importance/impact of trust, and how manipulatable it is.
While talking about barriers to diffusion, there’s a non-trivial likelihood that people are directing AI to solve the barriers to diffusion. Most of what we’ve identified are status quo barriers, but there’s a recursive element that we haven’t integrated. For example, maybe capabilities is really the only crux that matters. Capable enough AI can manipulate trust, persuade people to and how to diffuse it, and will lead to rapid diffusion whose speed depends only on capabilities.
Group 3: Intra-US Inequality
Introduction to Economic Causal Models: What are the best top-level metrics to measure growth?
Palma: Income share of top x% / bottom x%
Gini for wealth
Ideating on Intermediate Variables: What are the most important variables that contribute to the previously selected top-level metric?
(for Palma ratio)
Commonalities, Cruxes, Uncertainties
Write a 1-paragraph reflection and summary of the model you’ve created.
In estimating the Palma ratio c. 2030, assuming no major policy changes, we found that changes depend largely on the capital-labor income gap and on the elasticity of substitution between human and AI labor. As AI is a form of capital, changes in AI capabilities and applications will change the value and income-generating potential of capital stocks. Changes in returns on non-AI capital, such as stocks and real estate, will also impact income returns on capital. Changes in labor income depend on the elasticity of substitution, which in turn depends on the proportion of job tasks that are automatable, the marginal cost savings of automating those tasks, and the non-financial costs of integrating AI (including e.g. tacit knowledge reduction and coordination barriers). Changes in unemployment rates – affected by rates of reskilling of displaced workers (including effectively displaced workers whose wages are decreased as most but not some of their labor is automated) – and corresponding wage changes based on changes in labor supply, also impact changes in labor income. Finally, changes in taxation per tax brackets will affect inequality by affecting the distribution of labor income across tax brackets.
What are the most important cruxes you’ve found from this work? Answers to what questions would most help you resolve predictions from your economic model?
What was the area of most uncertainty in your model?
Call centers, driving gigs, therapists,
What are the topics that members of your table disagreed on the most?
Can all jobs be automated?
Group 4: Inter-Country Inequality
Introduction to Economic Causal Models: What are the best top-level metrics to measure inequality?
Condition on adoption cohorts
Ideating on Intermediate Variables: What are the most important variables that contribute to the previously selected top-level metric?
Deep-Dive Into Variables & Observables
Composition of India’s economy
Sector
GDP Contribution (%)
Employment Distribution (%)
Agriculture
17.66%
42.86%
Industry
25.92%
26.12%
Services
48.44%
31.02%
Some stats
India’s demographics
Some references:
Past studies on IT and agricultural productivity
First order effects of AGI by 2030
Adoption/diffusion
Impact
BPO demand down 4 million (1%)
Divergence (India as two countries, rural vs cities)
Divergent effect -> inequality increase
Services productivity way up
Infrastructure 50% adoption
Public health
Industry productivity up
70% rural adoption
Agriculture automation
(Why hasn’t India already adopted other technologies: (a) political and (b) intl competition)
Low demand for automation because labour costs are low
Public health better
(India has 20% of world chip designers)
Education better
India’s AI industrial policy:
https://ainowinstitute.org/publication/analyzing-indias-ai-industrial-policy
Agriculture in India
Service Sector
Some data
Commonalities, Cruxes, Uncertainties
Write a 1-paragraph reflection and summary of the model you’ve created.
We focused on India as an economy with components reflecting multiple aspects of economic issues addressing LMICs (services, agriculture, manufacturing, extractives, tourism). We then considered first order effects of AGI by 2030, seeing decreases in BPO, increases in service productivity, increases in manufacturing and extractives productivity, increases in agricultural productivity, better diffusion of higher quality public health and education services, particularly to rural areas
Constraints on adoption/diffusion include the rural/urban divide (70%/30%), limited access to the internet (~52% access, lower than global median of 66%), potential political and international competition reasons for lack of economic advancement and technological adoption.
Impacts: better public health, potential increases in in-country inequality. Likely low demand for domestic automation in near term due to low cost of labor, but likely major reductions in outsourced/offshored services/BPO labor based in India. May see opportunities to take advantage of presence of 20% of world’s chip designers based in India (though working for int’l firms), or globally distributed AI talent with Indian citizenship
What are the most important cruxes you’ve found from this work? Answers to what questions would most help you resolve predictions from your economic model?
What was the area of most uncertainty in your model?
The extent to which global competitive pressures pushes India to undertake more drastic reform or not.
What measurables, if you had access to, would help you predict the value of your top-level metric most effectively?
What are the topics that members of your table disagreed on the most?
Share of GDP
Share of employment
Export (% GDP)
Agriculture
20%
40%
5%
Industries
30%
30%
5%
Services
50%
30%
10%
Sectors where it will enhance productivity the most:
Group 5: Quality of Life
Ideating on Intermediate Variables: What are the most important variables that contribute to the previously selected top-level metric?
Some References
Major harm expectations: mass employment, pathogens, nanotech risks, depressed/anxious/mental health,
Assumptions:
We’re interested in self-reported agency / sense of purpose.
People can be free to travel etc. but feel useless. This is somewhat decoupled from labor participation.
AI can help with mental health processes such as cognitive behavioral therapy, but these are mitigating existing hard contexts rather than addressing the contexts that create these.
A portion of this comes from entrenchment, and not being able to be a meaningful participant in the civic context / in the tribe. Isolation from city, decoupling from political choice that governs people’s lives, prevention of social mobility
Relatively few of the other groups will be looking at physical health (nuclear war, nano)
Commonalities, Cruxes, Uncertainties
Write a 1-paragraph reflection and summary of the model you’ve created.
We agreed on most things. Could have fleshed out more the causal connections for how manipulation ops affect physical safety. We came up with what seemed reasonable for measuring things out in the world that aren't necessarily well-measured currently, that we assume can be given the right resources. Seems like a decent assumption for this process, but requires some stable methodology to actually do. We focused on physical safety at the exclusion of all the other parts of well-being (e.g parts of the OECD better life index) and we would have liked to be able to do more than one drill-down on the features of well being.
We also did not get a chance to look at different dimensions / other breakdowns of quality of life, which would have been interesting.
More work of this type is needed! By teams of full-time people.
What are the most important cruxes you’ve found from this work? Answers to what questions would most help you resolve predictions from your economic model?
How our metrics change over time - the longitudinal aspect of these metrics would be interesting to explore.
We kind of assumed a general international melee / conflict, but did not explore in detail how that conflict would evolve or be exacerbated, by AI or otherwise.
What was the area of most uncertainty in your model?
Likelihood of war is a strategic inflection point as to what to focus on for our model.
What measurables, if you had access to, would help you predict the value of your top-level metric most effectively?
Labour participation rate / unemployment rate would be most likely first order indicators in the majority of world lines for QOL.
For physical safety, probably things around civil unrest and regime stability, even thought we didn't cover those comprehensively.
What are the topics that members of your table disagreed on the most?
Disagree on rate of AI diffusion in research.