Suggestions
Chiyoung Kim
BD @ ServiceTitan
Chiyoung Kim is currently the Manager of Business Development at ServiceTitan, where he focuses on designing, driving, managing, and optimizing partnerships. He has a background that includes completing an MBA, which has likely contributed to his expertise in business development and strategic partnerships. Chiyoung is based in the Washington DC-Baltimore area and has a significant professional network, boasting over 500 connections on LinkedIn.12
Highlights
A few weeks ago, I shared my tool to learn how to prompt engineer with Anthropic's Claude 3: https://lnkd.in/ey8CuUbT .
I'm writing a series of posts to break down my process building this, from building the prompt, testing it out on Google CoLab, and then deploying it via Streamlit. I'm hoping it can be a useful read for others trying to learn and deploy fun LLM applications.
You can check out Part 1: Building the Prompt, at https://lnkd.in/eN3EYgj6
Also, now that it seems like you can get your own Anthropic API key at https://lnkd.in/eTQgD8Km , you can use it to learn prompt engineering! Let me know if you have any questions about the process or the Streamlit app.
#promptengineering #generativeai
Coding a webapp with Anthropic's Claude, Part 5 (likely last?):
For latecomers to this series: I've been attempting to build a notes app end-to-end using only Claude prompts, with no human code contributions to this experiment.
Major update: I did it. It's fully deployed on AWS with user authentication, MongoDB Atlas backend, and a much improved user interface at: https://lnkd.in/ezY3atmE
Note that the landing page sucks--you'll need to scroll all the way to the bottom (to the bottom of the Claude-generated content!) to click "Sign Up".
The app isn't perfect, and it's going to need more work. I'll be continuing to add improvements including bug fixes and UI/UX improvements. I'm going to break the no-code rule going forward and begin implementing code myself (with help from Claude, of course).
*Clarification: It rounds up to 100% no-code. I made a few minor CSS and code fixes that would have taken up tokens. Claude got me started and halfway there with AWS deployment, and I fixed up a few errors.
Some thoughts (to be expanded in a longer reflection):
The limiting factor is tokens and time: the longer the conversation got on claude.ai, the longer the computation time with Claude 3 Opus. As a result, as conversations got longer, fixes incurred more and more token costs and took more time. Small bug fixes ended up taking longer than they would have by hand. With the longest conversations, I could only send ~10-20 messages using Claude 3 Opus before I had to wait for my tokens to be refreshed!
The biggest performance improvement comes from giving Claude the latest code snippets: this, however, can incur a significant amount of token costs, so it's probably best to do this when you start to see the model make a few mistakes.
Preserving context between conversations: As a result of all the code, however, I ended up maxing out the initial conversation's token limits and needed to migrate to a new conversation. Thankfully, Claude is smart enough to carry on the work if you give it a good enough prompt with a comprehensive overview of the code and the app.
Performance-token tradeoff can be optimized: if you're trying to optimize between performance and tokens, this is where purpose-built AI leveraging smarter functionality like RAG will excel.
"Specialized" Claude chats: Instead of having Claude be a generally talented developer, it's helpful to give Claude the role of a specific specialization. This makes it so that Claude will give you targeted code and require less context.
Larger apps definitely need specialized chats: at a certain point, you'll be able to get more juice from each Claude chat if you have focused conversations on specific parts of your app.
Next steps: general bug fixes and performance improvements, and adding way better landing page features. Maybe I'll add another LLM?
👇 Let me know what you think in the comments! #promptengineering #claudeai #aidevelopment #llm
