Language Models at Oscar

It is hard to “waterfall” development around language models.

On a daily basis, new research or new language models are published that could be applied to Oscar’s stack. And yet, both the fundamental capabilities and the limitations of language models aren’t fully understood by anyone, not even theoretically. This is why we chose to take an iterative approach: generate lots of use cases, test fast, build a library of what works, learn a lot, share a lot.

• It's not enough for just tech to have a command of the possibilities for AI, you need the entire company to have at least some familiarity. We've organized workshops and roadshows to unearth AI-driven opportunities in the far reaches of the business.

• Everyone in the tech org (and every non-technical team who wants it) has access to a HIPAA-compliant Oscar-internal web app with test data and a gRPC service that talk to the OpenAI models - so everyone can test out ideas, immediately, from radical ideas to small but clever productivity hacks (like using AI to upgrade or deprecate Python libraries).

• We’ve set up an agile but robust AI governance process to ensure we’re launching these features responsibly. This includes guiding teams on ways to measure and maximize the accuracy and performance of LLMs for their use case, and assessing regulatory requirements and ethical considerations.

• We’ve decentralized development in each team, but have a centralized list of ideas, centralized prompt library, centralized discussion channel and knowledge sharing (like in this rapid fire run through of use cases). By centralizing our efforts, everyone can refine from what others are learning. We prefer to over-share.

With that said, let’s publish.

Internally, Oscar tech, operations and business development is organized into what we call the “Oscar Program Structure”. It is the most systematic answer to the question: what are all the components you need to execute well on to manage health risk? All Oscar tech sits somewhere in this tree structure. Click into the structure, and find out how our LLM use cases support those drivers. This is a live document which will keep getting updated as we launch new use cases, pilots and features. Keep checking back and give us feedback & ideas.