watsonx.ai

Watsonx.ai is an enterprise AI studio from IBM for foundation models (large pre-trained models), generative AI, and machine learning. Powered by foundational models, watsonx.ai allows users to train, validate, tune, and deploy models of their own models using significantly less data. Watsonx.ai is part of watsonx, a broader AI and data platform for enterprises, announced by IBM on May 9, 2023. Watsonx also includes watsonx.data, a data store aiming to have the flexibility of a data lake and the performance of a data warehouse, and watsonx.governance, a toolkit for responsible, transparent, and explainable AI workflows. Upon release, users can sign up for Watsonx.ai's waiting list with general availability expected in July 2023.

IBM aims for watsonai.x to accelerate its customers' ability to derive value from new AI tools. Foundation models offer a versatile foundation for a variety of AI applications, allowing users to quickly perform tasks with only limited annotated data. IBM has stated what is included in watsonx.ai will continue to evolve. At release, watsonx.ai capabilities include the following:

• Generate content without requiring code

• Build a classifier without training; read and classify text inputs based on the user's description
• Summarize dense text into short overviews personalized to the user's preferences
• Extract specific information from long and complicated documents

Example of watsonx.ai generating a marketing email

Users can access a suite of seven IBM foundation models (some of which are open-source) in watsonx.ai or bring their own models for training and tuning. IBM is also partnering with Hugging Face to include thousands of the company's models, datasets, and libraries in watsonx.ai. As part of the agreement between IBM and Hugging Face, IBM is contributing open-source AI development software to Hugging Face and making multiple in-house models available from Hugging Face's AI development platform.

IBM's foundation models in the watsonx.ai studio have been incorporated into a range of the company's products. Each model family has a geology-themed name:

• Granite—models based on decoder-only GPT-like architecture for generative tasks
• Sandstone—encoder-decoder models suited for fine-tuning specific tasks
• Obsidian—models utilizing a modular architecture developed by IBM Research to provide high inference efficiency
• Slate—encoder-only models that, while not generative, are effective for many enterprise NLP tasks

All IBM watsonx.ai models are trained on IBM's curated data lake using Vela, the company's custom-designed cloud-native AI supercomputer.