LLM Agents Mental Models.png

Introduction

“AI Agents” have become the latest buzzword in AI world and have always confused me about what does it mean. It does not help that there are a lot of products claiming to be “Agents” without having any of its core capabilities.

So, in order to dig deeper, I explored some research papers & courses that talk specifically about characteristics of LLM agents to build my mental model around the topic.

In this blog post, I aim to present LLM Agents & their core capabilities and also differentiating them standalone LLM applications.

Promise of LLMs

Data Efficiency. Reasoning Capability (debatable!). Generalisation.**

Traditional ML/DL AI systems required large amounts of labeled data for specific tasks. Furthermore, they had limited generalisable qualities and had to be re-trained for other tasks.

Whereas LLMs demonstrate remarkable generalisation across wide variety of tasks while being data efficient as they just need exposure to a few examples. This is because LLMs exhibit language based reasoning, generating intermediate steps to reach conclusions similar to humans! ML/DL systems learned from statistical pattern matching from massive datasets.

Therefore, LLMs excel in variety of use-cases such as text summarisation, translation, code generation ,etc

Standalone-LLM based Applications

shapes at 24-10-13 13.22.38.png

Standalone-LLM based applications integrate intelligence of LLMs into specific workflows, automating parts of it and thereby, enhancing productivity.

Use-cases : Summarisation, Code generation, Language translation, Content creation, etc.

Key capabilities : Enhanced User Experience & Productivity with few parts of workflow automated.

Key shortcomings :