Cross-posted here on the EA Forum.
Transformative AI and Compute - A holistic approach - Part 3 out of 4
This is part two of the series Transformative AI and Compute - A holistic approach. You can find the sequence here and the summary here.
This work was conducted as part of Stanford’s Existential Risks Initiative (SERI) at the Center for International Security and Cooperation, Stanford University. Mentored by Ashwin Acharya (Center for Security and Emerging Technology (CSET)) and Michael Andregg (Fathom Radiant).
This post attempts to:
This article is Exploratory to My Best Guess. I've spent roughly 300 hours researching this piece and writing it up. I am not claiming completeness for any enumerations. Most lists are the result of things I learned on the way and then tried to categorize.
I have a background in Electrical Engineering with an emphasis on Computer Engineering and have done research in the field of ML optimizations for resource-constrained devices — working on the intersection of ML deployments and hardware optimization. I am more confident in my view on hardware engineering than in the macro interpretation of those trends for AI progress and timelines.
This piece was a research trial to test my prioritization, interest and fit for this topic. Instead of focusing on a single narrow question, this paper and research trial turned out to be more broad — therefore a holistic approach. In the future, I’m planning to work more focused on a narrow relevant research questions within this domain. Please reach out.
Views and mistakes are solely my own.
You can find the previous post "Forecasting Compute [2/4]" here.
Lastly, I want to motivate the topic of compute governance as a subfield of AI governance and briefly highlight the unique aspect of compute governance.
Compute has three unique features which might make it more governable than other domains of AI governance (such as talent, ideas, and data) (Anderljung and Carlier 2021):
Second, according to my initial research and talking to people in the field of AI governance, there seems to be more of a consensus on what to do with compute regarding governance: restricting and regulating access to compute resources for less cautious actors. This does not include a consensus on the concrete policies but at least in regards to the goal. Whereas for other aspects in the field of AI governance, there seems to be no clear consensus on which intermediate goals to pursue (see a discussion in this post).
Within this decade, we will and should see a switch in funding distribution at publicly funded AI research groups. Whereas AI and computer science (CS) research groups usually had relatively low overhead costs for equipment, this will change in the future to the increased need for spending more funding on compute to maintain state-of-the-art research. Those groups will become more like high-energy physics or biology research groups where considerable funding is being spent on infrastructure (e.g., equipment and hardware). If this does not happen, publicly funded groups will not be able to compete. We can already observe this compute divide (Ahmed and Wahed 2020).
For a list of research questions see some “Some AI Governance Research Ideas” (Anderljung and Carlier 2021). My research questions are listed in Appendix A, including some notes on compute governance-related points.
Compute is a substantial component of AI systems and has been a driver of their capabilities. Compared to data and algorithmic innovation, it provides a unique quantifiability that enables more efficient analysis and governance.
The effective available compute is mainly informed by the compute prices, the spending, and algorithmic improvements. Nonetheless, we should also explore the downsides of purely focusing on computational power and consider using metrics based on our understanding of the interconnect and memory capacity.
We have discussed components of hardware progress and discussed the recent trends such as Moore’s law, chip architectures, and hardware paradigms. Focusing on only one trend comes with significant shortcomings; instead, I suggest we inform our forecasts by combining such models. I would be especially excited to break down existing compute trends into hardware improvements and increased spending.
Limited research in the field of macro AI
My research is based on a small set of papers, whereas most focus on certain sub aspects. Overall, the research field of macro ML trends in used compute is, to my understanding, fairly small. Seeing more research efforts on compute trends and algorithmic innovation could be highly beneficial. This could lead to a better understanding of past trends, and forecasting future trends — for example, breaking down the trend into increased spending and hardware progress can give us some insights into potential upper limits.
Limited data for analyzing AI trends
Another limitation, and perhaps the cause of limited research, is that , there is also limited data available. Consequently, researchers first need to build the required dataset. I would be excited to see bigger datasets of compute requirements or experiments to measure algorithmic efficiency.
We share in this work our public ML progress dataset and a dataset using MLCommons training benchmarks (MLCommons 2021) for measuring the performance progress of modern AI hardware and ask others to share their insights and data.
ML deployment engineers
As the role of compute is significant for AI progress, there is a strong need for ML engineers who can efficiently deploy AI systems. This was also discussed by Olah in an 80’000 hours episode #107. Consequently, ML engineers should consider working at safety-aligned organizations and enable the deployment of gigantic models which are —ideally— reliable, interpretable and steerable.
An essential component for compute prices and spending are economic models — either based on spending, or the computing industry, such as the semiconductor industry. Interdisciplinary research on those questions could be of great benefit. Examples of such work are (Thompson et al. 2020; Thompson and Spanuth 2021).
I plan to work on aspects of this research in the future and would be especially interested in exploring collaboration or other synergies. Please reach out. The exact research questions are still to be determined.
Appendix A lists various research questions that I would be interested in exploring and also want others to explore.
The appendix "Compute Research Questions and Metrics [4/4]" will attempt to:
You can find the acknowledgments in the summary.
The references are listed in the summary.
It seems reasonable and somewhat likely to me that we will be regulating and restricting the export of AI hardware even harsher and might classify it legally as weapons within the next decades. ↩︎