I am hosting another meetup at Fox.Build for AI with a focus on where AI is headed and the content generation that will inevitably be happening in the next 5-10 years. The internet is about to get very noisey with AI generated content and I want to show you the technologies enabling that and the experimentation that is happening nowadays online. I will touch upon how to make AI generated visuals. Also, I want to discuss being able to identify AI generated content.
Open source text-to-image generating AI Model
Stable Diffusion was an AI model that was a collaborative effort between a few different interested parties. Companies, non-profits, and researchers all contributed to making it.
At a high-level, computers have 2 types of compute. CPUs and GPUs - which one is used mainly for general purpose computations of everything on your computer and one is intended specifically for high-throughput of very similar calculations on graphics such as images and video. Right now, the models that have come out are really only run on GPUs and can't be run on CPUs that generate images or cool content. GPUs and even CPUs that could one day run future models can be relatively expensive. GPUs for training or running inferences can range from a p3.2xlarge $3.06/hr cost to p4d.24xlarge $32.7726/hr on AWS. I am cherry picking Nvidia GPU instances because most researchers use Nvidia GPUs.
Even if you were to get inferences running on a CPU, the cost would still be anywhere from $0.034/hr to $0.1445/hr (c6g.medium to c7g.xlarge). That's on the order of $24.48-$104.04 per month. However, the unseen costs there is the herculean lift it is to make AI models run on CPUs. Getting an AI model size down isn't easy. It's difficult.
Stablility.ai itself is looking like it is having problems with the costs. OpenAI, the company behind ChatGPT, celebrated when they got their costs down for their offerings because running inferences on GPUs isn't cheap. It also isn't cheap to deal with lawsuits.
Cleary, it is expensive to run GPUs. People don't always realize that. I believe that displacing the burden of running GPUs to the client-side or customer-side is likely to be important given how expensive it is to run GPUs on the server. It's likely to be something to consider during this transitional phase before even better GPUs and cheaper GPUs become the norm.