Overview ​

The evolution of LLM is nothing short of extraordinary. Hype cycle or not, AI agents will have a profound way we work & interact with the world.

GCP Duet AI demos were a revelation. I started exploring the open-source LLM and wrote an articlebuilding up an intuition around code generation. I am keeping a repo where I explore various libraries and concepts.

I separate AI engineering from others because it's highly exploratory. The big ideas are emerging - leveraging AI Agents to manage tasks and control flows. My ideal AI Agent - here's my dashboard; tell me what I need to know for the day - I don't have to provide statistical boundaries.

What follows here is just attempts to keep pace!

AI Engineering (Rise of LLM Apps)

The rise of AI Agents is inevitable. The idea of creating an adaptable application that can learn and remain flexible without code change is remarkable. Furthermore, the fragile prompt engineering process will give way to abstractions like 'reasoning', demonstrating a paradigm shift in application and data analytics will happen.

Industry Landscape & Current State ​

AI Agents vs Developers

I strongly believe that this is the future that AI Agents will play a profound role going forward. The idea of 'reasoning agents' to manage control flow is going to happen. For instance, one of my projects is trying to determine if a statistical outlier is something we care about, so we create deterministic rules (e.g., if this event took place, ignore the alert). However, a more scalable approach is pointing an AI agent at the problem and, with a bit of human feedback, allowing it to learn the patterns non-deterministically.

State of AI Report

An incredible overview of meta trends in the AI industry - here are a few ideas I took away. The best models employ RLHF - humans evaluating multiple responses to steer the reward mechanism to the best and safest answers. Open Source LLM models are potentially blossoming with Meta/Falcon releases. LLM Agents interacting with applications is here. Multimodal models are here (e.g., Vision in Robotics). In reality, very few players are building and contributing to LLM research - the reasons are obvious, the implications less so.

Chollet - How I Think about LLM Prompt Engineering

Fascinating mental models that prompting is like a key into a vector space. Different prompts navigate to other areas of the vector space.

Thought Provoking Articles, Libraries ​

  • e2b ai agents versus developers
  • Standford DSPy: DSPy unifies techniques for prompting and fine-tuning LMs — and approaches for reasoning, self-improvement, and augmentation with retrieval and tools. Abstractions around thought processes to avoid brittle and hard code prompts.

Libraries Explored ​

  • haystack - is an end-to-end NLP framework that enables you to build applications powered by LLMs, Transformer models, vector search and more.
  • langchain - library aims to assist in the development of LLM based applications.
  • LlamaIndex - is a data framework for LLM-based applications to ingest, structure, and access private or domain-specific data.
  • semantic-kernel - framework powering Copilot.
  • TaskWeaver: Data Analytics AI agent-based framework that manages memory, logging, and orchestration between User, Planner, and Code Generator via Chat. I worked through tutorials, documentation, and jotted notes down. My key takeaway is that agent-based systems are the future. This implementation is clever, interweaving custom plugins that supplement LLM models.
    plotai

    plotai It is a great library that interacts with Openai to automate the creation of plots in matplotlib with a simple command. note: need to specify the key in .env file

    python
    from plotai import PlotAI
plot = PlotAI(df)
plot.make("make a scatter plot")
    from plotai import PlotAI
plot = PlotAI(df)
plot.make("make a scatter plot")

Company Inventory ​

I am primarily focused on business intelligence and automated data analysis. ​

  • Julius - AI data analyst & more
  • getdot.ai - Make Data-Driven Decisions, Fast
  • datagpt.com - The First Conversational AI Data Analyst Ask DataGPT any question and get analyst-grade answers in seconds.