GPT-3-like models are now much easier to access and deploy than to develop

Abstract

GPT-3-like models are now becoming more accessible, with some models like BLOOM being publicly available for download. This raises questions about the resources needed to use and deploy these models. While the cost of deploying a model at scale for significant impact is probably an order of magnitude cheaper than training it, there are still barriers to deployment, including talent requirements and the need for substantial computing power. The article argues that focusing on shaping the development of large language models is more effective than attempting to control their deployment, because the process of development is more resource intensive, giving the AI governance community more leverage to impact the diffusion of this technology. – AI-generated abstract