Wordware (YC S24)’s Post

Kyle Corbitt and Saumya G. recently shared some insightful knowledge on optimizing outputs of LLMs with Mixtures of Agents (MoA) in their article on OpenPipe. 👀 They introduced the 'Mixture of Agents' (MoA) models, which outperformed GPT-4 variants in real-world tasks, being preferred 59.5% of the time. Smaller Llama 3 models fine-tuned on MoA-generated data also surpassed GPT-4 performance. Key Points: • Architecture: MoA uses a three-prompt chain to generate diverse candidate completions, critique them, and produce a final completion. It can be a drop-in replacement for GPT-4. • Performance Evaluation: MoA models achieved better results than GPT-4-Turbo in both open-source benchmarks and private tasks, with significant improvements in summarization, chat, and data extraction tasks. • Cost Efficiency: Fine-tuned Llama 3 8B models are 3x faster and 25x cheaper than GPT-4, while maintaining high quality and performance. This capability allows organizations to expand AI use, achieving better results and pushing the limits of AI's practical utility. See how Mixture of Agents can work on Wordware - see the link in the comments. Duplicate the app to see how it's prompted.

Our users asked, we delivered. 🚢 🗂 Structured Generation Type /structured generation to get output in structured JSON format.  👧🏻 New Models • Claude 3.5 Sonnet from Anthropic • Gemma 2 9B form Google Try them out and post your apps in the comments! (see also what Robert Chandler did with Claude 3.5 in a few minutes) 👁️🗨️ Follow Sub-Prompt Generation (see video) Click a running prompt to see how it generates, then return to the parent prompt. Check the Runs logs for more details.

Unlock the Power of Prompting with Wordware 🌟💡 Discover how powerful prompting can transform simple ideas into functional applications with Wordware. In this video, see how a chain of prompts can create a website using just two inputs: a name and a one-liner description of what the startup does. Using "Doggo" as an example, we show how a startup idea for helping customers move their pets abroad can be turned into a fully functional website. Watch as Wordware generates the HTML and hosts the site instantly, demonstrating the practical applications of your prompts. What's your startup idea? Generate your website using our app (link in the comments), and share your creations below!

Big thanks to everyone who mirrored our excitement at the AI Engineer conference! 🤳🏻 📸 We hope you snapped some cool selfies and enjoyed the reflections. See you next year! 👋🏻

List of Winners: 1. Aleksandra Pedraszewska 2. Tinah Hong 3. Jordan Krugman 4. Phil Salesses 5. Krzysztof Czarski 6. Nicholas Charriere We distributed the t-shirts to those who were with us at 4 p.m. Thank you all for participating!

To those we met at the AI Engineer conference, Our t-shirts found a real PMF (someone even tried to steal one - talk about high demand!). Since we only have 6 pieces left - one of each model, and we want to be fair to everyone - we’ll let LLMs decide who gets a t-shirt. We built a WordApp on Wordware that will collect and rank all comments from this post (instructions in the comments). That’s how we’ll distribute the last 6 t-shirts. May the best prompt win! To participate, leave a comment that will impress the LLMs. Reposting this doubles your chances. We’ll run the app at 4 PM next to our ‘hotter than H200 cluster’ mirror. You need to be there to pick up your t-shirt. If you want a t-shirt REALLY BADLY, the only way to get one without this little competition is to bring us a client or a very strong lead. PS: Leave your email and the model you like if you’d like to get a t-shirt once we have a new batch.

Build Smarter AI Agents with Wordware: ReAct Agent Tutorial 🧠✨ Explore the cutting-edge capabilities of ReAct agents, which combine reasoning and action plan generation to create highly intelligent AI systems. In this tutorial, you’ll learn to build agents that handle non-deterministic workflows with ease. In this video, we demonstrate: • How ReAct agents generate reasoning traces and task-specific actions in an interleaved manner, allowing for greater synergy and efficiency. • The structured process of question, thought, action, input, and observation, repeating as needed to achieve the desired outcome. • Using familiar programming concepts like loops, tools, and conditional statements to enable complex reasoning and decision making. Our interface, inspired by Notion but even more intuitive, makes it easy for newcomers to start building sophisticated AI agents. Watch as we create an agent to find the current Bitcoin price, showcasing how it autonomously searches for an API, writes the necessary code, and executes it to deliver accurate results. ReAct agents overcome issues of hallucination and error propagation by interacting with external sources, such as APIs, to gather additional information and ensure accuracy. This approach not only improves performance but also enhances human interpretability and trustworthiness. Discover how you can use Wordware to build intelligent agents that can reason and act effectively. Credits for the ReAct paper go to: Shunyu Yao Jeffrey Zhao Dian Yu Nan Du Karthik Narasimhan Yuan Cao

We built WordApp that creates a personal website just for you. Here’s how it’s built: • 🕵️♂️ Researches the person using Command R+ online (Cohere) and Sonar Large (Perplexity) models to ensure detailed info is gathered, avoiding mix-ups with others sharing the same name. • 🎨 Generates an image based on this research using Stable Diffusion. • ✍️ Composes a poem about you using GPT-4 Turbo (OpenAI). • 🗣️ Reads the poem aloud using ElevenLabs. • 🌐 Creates and hosts a webpage featuring all these elements using Sonnet 3.5 (Anthropic). Please be patient; it may take ~200+ seconds to generate everything. If you want to play around with prompts, add some APIs for different tools, or change models to compare the outputs, duplicate it by clicking the button in the top right corner! https://lnkd.in/dJfTm68B

Leverage the power of AI without coding 🤖✨ During the Q&A at the GenAI event, we covered: - What is Wordware - Who Wordware is for - What's the vision for Wordware Do you need an orchestration layer for your Enterprise AI agents? Are you a PM frustrated with long feedback cycles? Are you an engineer tired of receiving millions of prompts via Slack? Join Wordware for easy collaboration and streamlined LLM app creation. Thanks, The GenAI Collective, for hosting us!