🧩 Build, 🚀 Launch, 🌏 Scale, ♻ Repeat | Health AI Entrepreneur | FastCompany: 100 Most creative people in business
It took us < 30 mins to integrate Llama3 to our health AI stack. It takes our customers < 3 mins. Within a week Llama3 went live with all the 101 GenAI healthtech customers. We’ve curated and integrated 15+ models over last 7+ months to enhance healthcare workflows. Don’t spend weeks or months integrating with just one model; instead have a “team of models” power your “network of copilots”. Great post by Prashanth Aditya Susarla about how to build a high performance AI product with team of LLMs.
Building a high-performance AI product with not just one, but a team of #LLMs In what has been an enriching journey building 101 GenAI over the last 8 months, one of the simple yet profound discoveries has been the "personalities" of the foundational models we work with. Much has been said about their scores on top benchmarks, their inference speed (or lack thereof, with the notable exception of #Groq's blazing avatar of the newest kid on the block - #Llama3) and other quantifiable characteristics. But there's little commentary on what they all feel like, or remind you of. Given the variety of use cases 101 GenAI has built and the workloads that our copilots are subjected to, we have a ringside view into some of the subtler characteristics of these models. The way different models respond to the same inputs perhaps gives us some sense of the "philosophy" behind their training. Some (e.g., #GPT family) models excel at translating requirements worded in natural language into computer programs or declarative queries, while others (#Claude3, #Mixtral) excel at reframing ambiguous, coarsely-worded prompts into more formal, standardized ones that are able to get the best out of a downstream model. Yet others (#Gemini) amaze with the breadth of pre-training corpora they've been fed. In a way, this almost feels like a teacher discovering the varied talents of a bunch of kids in their class, or a community getting to know their members and the diversity in how they respond to the same stimulus. No wonder there's so much fascination with this rapidly evolving field. We don't see a winner-take-all situation here. The way forward seems to be building AI-powered applications that deeply understand the strengths and weaknesses of individual foundational models, and putting together a framework that achieves the best out of a carefully curated and constructed network of copilots. This is much like how champion sports teams or high-powered organisations are built. The parallels seem almost uncanny, yet obvious in hindsight! It takes 101 GenAI less than half a day (sometimes only as much as half an hour) to integrate and deploy to production a new inference service provider and all the model families they serve. Within hours to less than a week, our benchmarks provide us such insights into these models' characteristics. With these inputs, we're able to determine the best network of copilots that can solve a particular use case. It won't be long before these recommendations and application design start getting solved by some of these models working as a team, with the input simply being our customers speaking out their requirements aloud to our listener. And we can't wait to get there as fast as we can! If all of this sounds exciting, come hit us up! My partners-in-crime and best buddies Janardan Prasad and Praveen Dua will be thrilled to share more at Booth 310 at the TiE Silicon Valley #AI Ubiquity Conference (May 1-3, '24).
