Building models for predicting when and how a transition to a world with advanced AI will occur
We aim to develop quantitative models to elucidate when a transition to a world with advanced AI will occur (AI timelines), and how quickly this transition might play out (AI takeoff dynamics).
The answer to these questions will help inform others about what avenues of work to prioritize. For example, on longer timelines we may want to prioritize building broad scientific and governance capacities that can be leveraged later, while shorter timelines might suggest investing in a suite of focused and targeted strategies for reducing risks.
This line of research involves:
- Reviewing existing models to forecast AI development and deployment; critically evaluating their assumptions, and assessing their strengths and weaknesses
- Developing and improving models to forecast AI development and deployment by carefully updating and improving modular components
- Improving our estimates of key parameters in models of AI development, like the returns to hardware R&D or the growth in AI investment. These estimates can be improved by e.g. reviewing the relevant literature, or gathering relevant data.
The longest training run
Aug. 17, 2022
Jaime Sevilla, Tamay Besiroglu, Owen Dudney and Anson Ho
Training runs of large Machine Learning systems are likely to last less than 14-15 months. This is because longer runs will be outcompeted by runs that start later and therefore...
Projecting compute trends in Machine Learning
Mar. 07, 2022
Tamay Besiroglu, Lennart Heim and Jaime Sevilla
Projecting forward 70 years worth of trends in the amount of compute used to train Machine Learning models.