Applications of high-capability models
November 2021
On the last page, I ran through a series of rough order-of-magnitude estimates of the scale and capabilities reachable by future large training runs, and concluded that there’s a reasonable chance that models will match or exceed human capabilities at all but the most difficult-to-train tasks in the next 20 to 50 years.
On this page, I’ll try to make this scenario and its consequences more concrete by sketching out how high-capability models could actually be trained and used.
What training could look like
For many practical tasks, it is unlikely that high-capability models will be trained from scratch. The largest and most capable models today have already moved to a multi-stage training process, where a model is pre-trained once on large amounts of data and then fine-tuned for many different tasks.
As models grow larger, it seems likely that multi-stage training will become more common, with earlier stages imparting general skills and preparing the model to learn more sample-efficiently during later stages. The training process for practically useful high-capability models might look something like this:General pre-training
Unsupervised learning or meta-learning on large amounts of general data, games, and procedurally generated tasks
Very high compute and data requirements (see Estimates of future training runs)
Most likely performed by a few deep learning companies specializing in large models
Domain specialization
Unsupervised learning or meta-learning on domain-specific data
E.g. generative modeling, offline RL, learning to imitate human domain experts
Lower compute and data requirements
Task training
Fine-tuning on the target task using RL, supervised learning, or prompt engineering (to select among behaviors learned during earlier stages)
May need to use online RL or other ongoing adaptation mechanisms
Applications and outcomes
As a reminder, I’m thinking here about models that combine human-like conceptual understanding with superhuman reasoning and planning abilities. Broadly, I think the most significant applications will take advantage of the facts that (1) once trained, models can be copied many times to perform large-scale tasks that would be infeasible with human labor, and (2) in some domains, models could achieve better performance than any human expert.
I think we should expect the highest-impact applications to come from large companies and governments, since they have both the resources available to invest in models and the potential to benefit from scalable high-quality decision-making. These organizations could use models in many ways:
Large-scale operations: Organizations could apply fleets of copied models to large-scale datasets and tasks, e.g. data gathering and analysis, server and software management, advertising, shipping and supply-chain logistics, surveillance, social media manipulation, and physical or cyberspace military operations. Model-run operations would scale up more easily with increased resources, be easier to manage at very large scale, and could make more consistent, faster, higher-quality decisions than large-scale human operations; it seems plausible that the capabilities of model-run operations would quickly outpace competing human-run organizations.
Research and development: Organizations could rapidly develop new products and technological capabilities by applying models to problems in areas like software engineering, electronics, biotechnology, materials science, manufacturing process design, or robotics. Research and development might be sped up significantly by using copies of models to investigate many research directions at once, or by training models to super-human levels of expertise.
Strategic planning and management: Some of the most important decisions organizations make are about which projects to pursue, how to respond to new problems and opportunities, how to invest their resources, and how to compete or transact with other organizations. It seems plausible that the same kinds of reasoning and planning abilities that have allowed models to achieve super-human performance in board games could be applied to organization-level planning and management. Strategic models of this type could give organizations large advantages across many areas of commerce, industry, finance, infrastructure management, politics, and military operations.
In each of these three categories, it seems likely that the activities of a model-run organization could become so complex and fast-moving that unaided humans – at competing organizations, outside institutions, or even within the same organization – wouldn’t be able to understand what they are doing or make competitive strategic decisions, in much the same way that unaided humans have difficulty understanding or competing with a Go- or chess-playing model today.
Overall, it looks the companies and governments most willing to delegate large-scale operations, R&D, and strategic planning to high-capability models would gain large and growing competitive advantages (although if training safety problems are not solved, I expect human stakeholders to ultimately lose control of these organizations).
As a result of these advantages, it seems likely that after a period of adoption, growth, reinvestment, and competition, model-run organizations would end up in effective control of a large fraction of the world’s resources, technological capacity, and political and cultural power, making their behavior one of the largest factors shaping the future development of civilization.