Guest: Devvret Rishi (LinkedIn)
Company: Predibase (Twitter)
Show: Let’s Talk

Predibase is a platform for developers who want to productionize open-source AI models. Developers can use an open-source model and fine-tune it on their data, then they can serve the model inside their organization. Predibase is built on a foundation of deep learning models that support many different types of use bases, and they primarily lead with fine-tuning and serving fine-tined open-source LLMs. Devvret Rishi, Co-Founder & CEO at Predibase, says, “We make it easy for anyone to fine-tune and train a model.”

Rishi has a long-standing academic background in Computer Science, Statistics, and Machine Learning. He noticed while at Vertex AI that many organizations were struggling to get any value out of ML. In 2020, he met his co-founders who had the idea of building a product base that was built on top of a tool one of the co-founders had made and open-sourced in 2019, Ludwig. They aimed to make it easier for organizations to build models.

Open source is really important to Predibase which is built on top of open-source projects and around the open-source ecosystem. They have built two open source projects, one for training machine learning models, Ludwig. The second one is called LoRAX, which provides a cost-effective way to deploy open-source fine-tuned models in the market. Some of the best models coming out today are open source and these projects allow organizations to leverage them to fine-tune them so that they can own their models and IP.

Predibase believes that the models and their weights should be open source. Training models from scratch can cost 10s of millions of dollars, but having open source models, the architecture, and the weights of the models open source lowers the barrier to entry and provides other companies with a good starting point. Companies can also use their data to fine-tune the models rather than using proprietary data. However, Predibase feels that it makes sense to commercialize the infrastructure, managing customers’ compute so that they can get the underlying access to it so that it works in production with their internal data.

Open source can be beneficial as it provides options for organizations that might not need models with 100 billion parameters. There are three key reasons why open source will likely dominate the market in the future: it prevents vendor lock-in, it enables you to take a prototype and fine-tune it within your organization to improve its performance, and it gives you the choice to pick a model that is the right size and fits the task.

Some of the key reasons Predibase’s customers come to them is because they have identified a use case and need to improve the efficiency and performance and bump it up to the next level so that it can be served in production most cost-effectively. They also want to own the model and brainstorm the use cases around it. Many developers are now starting to do machine learning themselves, which can be seen in communities like Kaggle but as we move to an API-centric model it is also now occurring more in organizations.

This summary was written by Emily Nicholls.