GPU-accelerated databases are mind-blowing! ​ Imagine a database natively integrated with best-in-class AI foundational models: ​ • Zero warmup latency • Massive GPU-backed scalability • Ability to process your data with any model • Ability to train and fine-tune models on your data ​ There are 1.7 million deployments of PostgreSQL worldwide, one of the world's most popular relational database management systems. Now, the team behind @postgresml is making them 10 times more powerful. ​ Look at the attached image. ​ That's an SQL script running Llama 3 at the database label to generate titles for a novel! You can do the same in Python or JavaScript. ​ You can access models like Llama 3, Mistral, and Phi without leaving your database. This has massive speed and cost implications! ​ You can also integrate your database with Hugging Face transformers and train and run supervised and unsupervised algorithms on your data. ​ What's happening here is massive: ​ We are combining the compute and the data layers as a single GPU-accelerated database. You can access virtually infinite GPU processing cores and terabytes of GPU memory on the same hardware where your data is cached. ​ This will completely change how we start building RAG applications. ​ This post is sponsored by @postgresml xitter.cf/postgresml. Their service is one of the most cool innovations I've seen lately. They have much more information and several benchmark metrics on their site: postgresml.org. ​ Pay particular attention to their code examples. Mind-blowing stuff!