In case you weren’t aware, it’s VMWorld! VMWorld is an annual conference hosted by VMware, the premiere enterprise-class data center virtualization technology company. What does all that mean? Basically, turtles all the way down. Virtualization allows you to run servers inside of larger servers, or containers inside of larger servers. They achieve this by abstracting physical hardware into virtual hardware. Or, as my girlfriend dubbed them: Inception machines. This sounds complicated but their are HUGE advantages to virtualization. This technology allows you to share hardware with multiple servers and applications, so you can get far better utilization of expensive hardware. It also gives you flexibility to move workloads around. If one physical server is running hot, you can move some workload off to lower usage servers.
How does this figure into OpenAI and GPT-3?
Well, at present, these giant models (LLMs or large language models) have to run on “bare metal” hardware. This means no virtualization. No virtualization means it is more expensive to set up and run AI/ML workloads, specifically these large models like GPT-3. In short, one of the reasons that GPT-3 DAVINCI is so expensive to use is because these cost-savings technologies have not scaled up enough yet.
Unfortunately, right now VMware is focusing on already-commercialized types of models, such as ASR, TTS, and object detection. These models are tiny in comparison to the likes of GPT-3. However, these advancements by VMware will be able to run smaller models, possibly up to the CURIE size. (Important note, this is not VMware’s opinion, this is my personal prediction, so please don’t cite this as news directly from VMware or NVIDIA). Then it’s only a matter of time before the tech scales up further.
In total, these advancements mean that LLMs will start to get cheaper to train and run soon. It will still be a few years before you see something the size of GPT-3 running in commercial environments. This will be a crucial step in making AI accessible to everyone. In point of fact, one of VMware’s explicitly stated goals is to democratize access to AI.
(Disclaimer, I do not work for VMware, I’m just a fanboy)