AI tech stack (updated Mar '24)

Tl;dr:

• If you’re technical: use LangChain and Autogen
• If you’re not: GPT Sheets and SuperpowerGPT chrome extension
• Recommended: blog post detailing some use-cases, including a demo of an elegant approach to UX simplification for students. For those thinking "just tell me what I can do with this..."

Longer version with rationale:

• LangChain: At the core of my toolkit, LangChain stands out as a versatile framework designed for leveraging language models in a variety of applications. Its flexibility in integrating with other tools and services has enabled me to build complex, thought-driven AI applications with relative ease.
• AutoGen: Before going to LangChain, AutoGen has abstractions that make agent simulation (including multi-agent chats) extremely quick and easy. It’s also actively very developed and latest improvements to their GroupChat in particular have made it a strong contender.
• LangSmith: AI is great, but you want to be able to trust it. With Langsmith, debugging your prompts/agents/chains are possible. It’s the biggest Its intuitive interface and powerful analytics enable me to quickly identify and resolve issues, ensuring that my applications run smoothly and efficiently.
• ChromaDB: For vector storage, ChromaDB has been an invaluable asset. It’s got a simple api, and and makes it that much easier to quickly chunk, embed and store all my stuff. Prior to this, it was lots of fiddling around with more confusing interfaces that detracted from the main purpose.
• Open Interpreter: For those unpredictable tasks that arise, has been my go-to. It’s ChatGPT’s code interpreter on steroids. Its adaptability and precision in understanding and executing ad hoc commands have made it an indispensable part of my daily workflow.
• GPT Sheets: Integrating AI into spreadsheets has never been easier thanks to GPT Sheets. Even if you code, this tool is useful for quick prototyping. More importantly, it has democratised access to advanced prompting techniques like few-shot prompting, prompt chains etc., all within the familiar confines of a spreadsheet.
• GitHub Copilot and Aider-Chat: Everyone can be a 10X engineer now. Okay, perhaps not - but these definitely help any programmer do what they want to do more efficiently. Just be sure to not get lazy with it, and remained disciplined with thinking, planning, reading the outputs. Sharp tools lead to dull minds over time, if you’re not careful.
• Whisper: You type at 100wpm? Cool... but a child could probably speak twice as fast easily and it won't take years of practice. Transcription will become a larger input medium over time, especially as natural language interfaces mature. Whisper's advanced speech recognition capabilities (importantly, greater accuracy) now makes transcription a low hanging fruit for most people seeking to enhance their workflow.

Previous choices (only honorary mentions):

• Microsoft Guidance: Used this a lot, but was less reliable with the chat models (was better with LLMs). I gave it a go when they relaunched with v1 but never took to it again after LangChain.
• Quivr and PrivateGPT: Great UI, privacy, but I found that whenever I wanted to extend it, it was easier to do it from scratch instead (partially because of tech stack familiarity). Quivr also takes quite a lot of resources so the hosting costs might come into play if you’re using it yourself.
• Codium AI: Although innovative in its approach to coding assistance, the advent of tools like GitHub Copilot and Aider-Chat provide a more seamless experience and better understanding of code context and developer intentions IMO.
• SuperAGI/AutoGPT: Their ambitious scope often outpaced practical application feasibility, and so while cool, I struggled to get useful output from it. Other less ambitious projects offer immediate, tangible benefits and easier integration into existing workflows.
• GPT Engineer: Really wanted to love it but found Aider-Chat somehow just worked better for me, especially with interative development (rather than starting a project from scratch).
• GPT Researcher: Tavily researcher probably just was easier for me to use.

Special mention for these Agent based frameworks - still yet to give them a fair chance, I’ve just been more focused on Autogen but there are promising aspects to each of these:

• MetaGPT
• Crew.AI
• AgentGPT
• Fabric

Finally, hold out for a post on open-source models, fine-tuning and implementations e.g. Ollama, LM Studio. Deserves its own post!