PromptLayer

We're #hiring a new Frontend Product Engineer in Manhattan, New York. Apply today or share this post with your network.

Had fun giving a talk at Arize Observe in SF. All about prompt evaluation as an iterative process 🍰

Ellipsis is building the ultimate coding agent and developer productivity tool. We are excited to share how the team is using PromptLayer to build their AI code review agent! As the Ellipsis team rapidly grew to handle 500K+ daily requests, they used PromptLayer to reduce debugging time by 90% and avoid spending hundreds of hours building an in-house solution. PromptLayer allows engineers to pinpoint errors in three clicks, and we give non-technical stakeholders insights and tools to help participate in this process. For an agentic product like Ellipsis, faster iterations and more collaboration translates to a seriously better product. Read the full case study here: https://lnkd.in/eHjbPjkd

ParentLab is revolutionizing parenting support through the product Era's AI-powered conversations that provide timely, personalized guidance to parents. PromptLayer's visual prompt CMS allows ParentLab's educators to rapidly iterate on the AI's voice, no technical background required. The results: 10x faster prompt deployment, 400+ engineering hours saved, and 700 prompt revisions in just 6 months. PromptLayer also provides critical analytics to track edge cases, usage, and costs as ParentLab scales. Read the full case-study here: https://lnkd.in/eA9UWkTv

Be sure to check out AI-Agent from the Gorgias team. It's a customer service chatbot that actually resolves conversations and takes actions. Exciting to see PromptLayer power amazing products. Congratulations to the team! Really a great LLM use-case and a very well made product.

Enjoyed joining the EverydayAI podcast last week with Jordan Wilson to discuss WHO will reap the most benefits from the AI revolution and WHAT makes someone a good prompt engineer. Some topics we discussed: 🧩 Domain knowledge is crucial for building differentiated, successful AI products that go beyond ChatGPT wrappers 💬 Prompt engineering is more about communication skills and "computational thinking" than coding ability 🏅 Companies should empower non-technical subject matter experts to lead AI projects and tailor products to their specific domain 🤖 The most powerful AI applications will combine cutting-edge language models with deep domain expertise 🚀 Adapting to rapidly evolving AI capabilities requires a modular "prompt routing" approach vs. over-relying on one model Check out the full episode and lmk what you think! Was fun to step out of the IDE and talk about the things we are learning at PromptLayer. See the full episode here https://lnkd.in/dgSCnTij  

Part 2 of the tutorial for building your own ChatGPT is out! Adding world information to your chatbot https://lnkd.in/d7urifQw

New PromptLayer job posting! Front-end Product Engineer DM me if interested or know someone 🍰

Prompting a large language model requires a bunch of tech know-how right? ↳ Super structured inputs ↳ RAG ↳ Fine-tuning Meh. Not so much. The best way to prompt your way to better results? Flex your domain expertise. Why the future of AI will be built by non-technical domain experts -- An Everyday AI chat with Jordan Wilson and Jared Zoneraich == == #prompts #prompting #promptengineer #promptengineering #llm #llms #larglanguagemodels #openAI #chatpgt #gpt4 #gpt4o #claude #gemini #copilot #microsoft #gemini #promptips

Why fine-tuning is (probably) not for you 👀 With the help of Pranav Kanchi, we spent some time talking to customers & researching the state of fine-tuning Fine-tuning can be really powerful, but is way overhyped (at least today). It makes building a good LLM application harder. The #1 thing you should worry about is how fast you can iterate on prompts & the AI application. Fine-tuning is slow. Training a new model takes time and is expensive. In today's world of LLMs, you want to just be trying tons of ideas and seeing what sticks. 🙅 Why you should avoid fine-tuning 🙅 1. RAG is usually just way more effective 2. Complex process: dealing with data, pipelines, and training 3. Slow iteration cycles means mistakes are less forgiving 4. Tons of hidden costs: Constant updates & dev time to name a few. 5. Large data needs. 6. Potential data privacy issues. These things might change. I think we are entering a world where everyone will be fine-tuning OSS models for their use-cases. We just aren't there yet. Everything in prompt engineering is a tool 🪛 Obviously there are some things that fine-tuning is actually the best tool in your toolbox. ✅ When fine-tuning is the answer ✅ 1. Tailoring specific output formats 2. Edits tone & writing style 3. Improves complex reasoning 4. Reduces token usage for complex prompts. 5. Can get GPT-4 level performance from GPT-3.5 #5 is the simplest and most common use case. Just train a cheaper model on your old, expensive outputs! Easy. So why not use both? Maybe... but don't start with fine-tuning. Fine-tuning has a lot of mindshare but it's mostly just VC Twitter buzz. You should be focused on shipping the best product as quick as possible. https://lnkd.in/evw8tXWd