What are foundation models?

Over the last few years, artificial intelligence (AI) has gone from a technical concept to an everyday occurrence. From your music recommendations to search engines like Google and Bing, AI is already embedded in many different parts of our lives.

The practical applications of AI aren't the only thing that has seen rapid advancement in recent years. Just a year ago, enterprise AI teams needed massive labeled datasets to train custom AI models — and they were often hand-labeled by people in a slow and often costly process for enterprise teams. Thanks to the advent of foundation models, however, the path to building AI has become far faster, easier, and more streamlined.

Definition of foundation models

Foundation models are models that are trained on a broad set of unlabeled data that can be used for different tasks with minimal fine-tuning. This term was originally popularized by the Stanford Institute for Human-Centered Artificial Intelligence and marks a shift in task-specific models to models that are built more broadly in order to set the foundation for different applications of AI. Using self-supervised and transfer learning, foundation models apply the information they’ve been trained on and extrapolate it to other situations.

Today, instead of having humans label all their data for training AI, enterprise teams can leverage foundation models to enrich and pre-label data and use humans in the loop for quality assurance. Pre-labeling data can reduce labeling time and costs by half. When teams use foundation models for this task, it can cut labeling costs by nearly 90%.

Enterprise teams can also fine-tune foundation models, particularly large language models (LLMs), instead of building AI systems from scratch. As they start with a baseline model of such advanced capabilities, the model development effort, time, and costs are greatly reduced.

The importance of foundation models

Accessing large amounts of high-quality labeled data has long been an expensive and time-consuming challenge for machine learning teams. Advanced ML teams often adopt partially automated labeling workflows to mitigate costs and accelerate model development. 

One such technique is pre-labeling — that is, feeding unlabeled data through a model (off the shelf, a model built specifically for this purpose, or an early version of the model in training) and using the results as a starting point for labeling data. Some ML teams also had their data first labeled via a software program instead of a model. 

Pre-labeling requires human labelers to review and correct the model’s errors rather than labeling data from scratch. When effectively incorporated into the ML workflow, pre-labeling can cut labeling time and costs by over 50%.

With the advent of foundation models, ML teams need to do much less work to create an effective and time-saving pre-labeling process. With such powerful off-the-shelf models available, many teams no longer need to build their own models for pre-labeling. 

Teams that were previously relying on programmatic pre-labeling can now upgrade their whole workflow with low effort. Those using a foundation model for pre-labeling have experienced nearly a 90% reduction in labeling time and costs.

How foundation models are used

LLMs are some of the most popular foundation models in use today, including GPT-4 from OpenAI and Claude from Anthropic. These algorithms can be effective starting points for many basic text-based enterprise AI efforts, such as:

• Extracting specific information from text documents

• Removing comments with certain words or phrases from online forums or comment sections

• Sentiment analysis from customer reviews and social media mentions

ML teams will still need to fine-tune LLMs as needed. Even so, the journey from ideation to deployment will likely be much faster when beginning with a foundation model than building and training a model from scratch.

Potential challenges of foundation models

Taking full advantage of foundation models’ potential still requires much data-centric labor for AI teams, including:

• Experimenting with and testing generative AI models

• Architecting workflows to automate labeling or reduce human labeling

• Choosing the right foundation model for a specific use case

These tasks are challenging and time-consuming for AI teams, even as they leverage foundation models to accelerate their work. To make these tasks easier and faster, AI teams can integrate a comprehensive, data-centric AI platform like Labelbox into their development processes.

As Labelbox CEO and Cofounder Manu Sharma discussed in his session at the Data+AI Summit, AI teams can use Labelbox Foundry to:

• Explore and experiment easily with foundation models’ capabilities on their data

• Use out-of-the-box and custom workflows to automate labeling and quickly enrich or pre-label data

• A/B test foundation models on conditions and parameters that matter for your specific requirements

• Securely incorporate people into the workflow for model evaluation and data quality assurance

Final thoughts on what are foundation models

Previously, when building new models, machine learning teams needed to ensure that there was a high-quality, well-labeled dataset to train models on. If this dataset didn't exist or your data was of low-quality, you'd need to either start over again or spend a large amount of both time and money to find the appropriate critical mass for that training dataset. 

As the sea level of AI rises with expanding foundation model capabilities, every enterprise can benefit from AI-enhanced processes faster and more easily. With the right tools to help teams integrate foundation models into their development process, enterprises can mitigate some of the toughest data roadblocks and rapidly evolve into an AI-powered organization. Learn more about how you can use Foundry for your enterprise AI journey.