Tom Davidson's compute-centric framework

What a Compute-Centric Framework Says About Takeoff Speeds | Coefficient Giving 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

 
 

 

 
 

 
 
 
 
 
 
 
 
 
 
 



 

 

 

 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

 
 
 

 
 

 
 Skip to Content 

 

 
 
 
 

 
 
 *+*]:mt-5">
 June 27, 2023 
 What a Compute-Centric Framework Says About Takeoff Speeds

 
 
 
 By
 Tom Davidson
 

 
 
 
 
 

 
 

 
 Editor’s note: This article was published under our former name, Open Philanthropy. Some content may be outdated. You can see our latest writing here . 

 This is Part 0 of a four-part report — see links to Part 1 . Part 2. Part 3 , and a  folder with more materials . 

 Abstract

 In the next few decades we may develop AI that can automate ~all cognitive tasks and dramatically transform the world. By contrast, today the capabilities and impact of AI are much more limited. Once we have AI that could readily automate 20% of cognitive tasks (weighted by 2020 economic value), how much longer until it can automate 100%? This is what I refer to as the question of AI takeoff speeds; this report develops a compute-centric framework for answering it. 

 First, I estimate how much more “effective compute” – a measure that combines compute with the quality of AI algorithms – is needed to train AI that could readily perform 100% of tasks compared to AI that could just perform 20% of tasks; my best-guess is 4 orders of magnitude more (i.e. 10,000X as much). Then, using a computational semi-endogenous growth model, I simulate how long it will take for the effective compute used in the largest training run to increase by this amount: the model’s median prediction is just 3 years. The simulation models the effect of both rising human investments and increasing AI automation on AI R&D progress. 

 How to read this report

 The right approach depends on your technical background:  

 
 Non-technical readers should first watch this video presentation ( slides ), then read this blog post , and then play around with the Full Takeoff Model here . 

 Moderately technical readers should first read the short summary , then play around with the Full Takeoff Model here , and then read the long summary . 

 If you have a background in growth economics, or are particularly mathsy, you might want to read this concise mathematical description of the Full Takeoff Model.  

 
 I do not recommend reading the full report top to bottom. Instead, treat its sections as providing longer discussions of the important modelling assumptions and parameter values of the model. 

 This section  lists particular sections of the full report that I think would be useful to read after the long summary. 

 Short summary

 In the next few decades we may develop AGI that could readily [1] The phrase “readily” here indicates that i) it would be profitable for organisations to do the engineering and workflow adjustments necessary for AI to perform the task in practice , and ii) they could make these adjustments withi

... (truncated, 98 KB total)