[Feb 5] Implementation Experiences with Domain LLMs

A lot of theoretical work is happening but delivering it to end customers is still a bit of challenge. This week we focus on implementation experiences.

A shout out to my collaborator Rajesh Parikh for creating a new Agent Comparer service that intends to help Agents make decisions on which LLM to use when and why.

When Doctors With A.I. Are Outperformed by A.I. Alone

Author: Eric Topol and Pranav Rajpurkar

URL: https://erictopol.substack.com/p/when-doctors-with-ai-are-outperformed

This article uncovers some surprising findings where A.I. systems working independently have significantly outperformed doctors using A.I. tools. It argues for a more strategic division of tasks between humans and A.I. in healthcare to improve patient outcomes.

This pointed me to an opportunity - deconstructing workflows to correctly identify task boundaries and finding the right collaborative setup for each task.

Key Points:

• A.I. can outperform doctors in diagnostic tasks
• Atleast three possible configurations: AI without human, AI shadowing human, AI assisting human
• A.I., even if better, can trigger trust issues if collaboration setup is not constructed well
• Rethinking the human-A.I. collaboration model

Hippocratic AI

Author: Hippocratic AI Team

URL: https://www.hippocraticai.com/about

Hippocratic AI aims to tackle the global shortage of healthcare workers by developing a safety-focused Large Language Model. Now it has an appstore and agents. Their LLM is documented here: Polaris: A Safety-focused LLM Constellation Architecture for Healthcare

What I loved about the company was the structure, timeline, and offering which could be a template for other similar efforts.

Key Points:

• Addressing 15 million healthcare workforce gap
• 100+ agents covering some 20 different areas
• They are allowing licensed medical practitioners to create agents
• Revenue share with creators

Domain Specialization as the Key to Make Large Language Models Disruptive: A Comprehensive Survey

Author: Chen Ling et al.

URL: https://arxiv.org/abs/2305.18703

This article dives into the world of domain specialization for large language models (LLMs) and examines how they can address specific challenges across various applications. The authors systematically categorize techniques and application domains, emphasizing the potential and hurdles that specialized LLMs face.

This paper has a good summary of all the different ways we can use domain knowledge with the LLMs. This is a neat mindmap

Key Points:

• New taxonomy for domain specialization techniques
• Applications in healthcare, education, and legal fields
• Future trends point to more tailored AI solutions

Implementation Experiences

These are a collection of interviews/talks where creators of domain-specific models/agents discuss the learnings from their efforts

• Unlocking the Power of Domain-Specific AI: Bloomberg GPT
• Enhancing NLP with Galileo & John Snow Labs
• Leveraging Large Language Models in Finance