My new app is really stupid (I wrote none of the code)

Too Long; Didn't Watch — Summary

Quip Slop is an AI-driven game where different LLMs compete to provide the funniest responses to prompts, streamed 24/7 to Twitch using a custom-built headless browser and FFmpeg pipeline.

Main Takeaways

• AI vs. AI Comedy: The project automates the 'Quiplash' game loop where one AI generates a prompt, two others respond, and a panel of other models (plus Twitch viewers) vote on the winner.
• Vibe Coding Workflow: The application was built almost entirely without manual coding, relying on AI instructions to generate the CLI, web interface, and streaming logic.
• Technical Innovation in Streaming: To stream to Twitch, the developer used Puppeteer to run a headless browser, captured a canvas element as a media stream, and piped it into FFmpeg.
• Overcoming Model Repetition: A major hurdle was AI models repeating the same jokes; this was solved by dynamically shuffling a pool of 870+ reference prompts into the system instructions.
• Live Launch Chaos: The project faced immediate real-world hurdles, including a DDoS attack by 10,000+ bot connections and multiple accidental leaks of the Twitch stream key.

Detailed Summary

Introduction (0:00 - 1:28)

Quip Slop is a new side project designed to determine which AI model is the funniest. The game flow involves one model creating a prompt (e.g., a fill-in-the-blank), two models responding, and the remaining models—along with Twitch viewers—voting on the best answer. The project aims to scale to tens of thousands of users and features a live leaderboard.

Prototyping with a CLI (2:52 - 3:37)

Development began with a Command Line Interface (CLI) to test the core logic without the complexity of a UI. This phase focused on the 'model-to-model' interaction loop. Initial tests showed that models like DeepSeek were capable of providing solid, context-aware answers to humorous prompts.

Solving the Randomness Problem (3:37 - 5:23)

Early versions suffered from repetitive prompts (e.g., constant jokes about grandmas or funerals). To fix this, the system prompt was updated to include a rotating selection of 60-80 examples drawn from a larger database of 870 prompts. This ensured every run had a unique context. Additionally, strict system instructions were implemented to force models to be concise (under 12 words) and avoid conversational preamble.

The Web UI and Twitch Integration (5:24 - 5:52)

While a WebSocket server was initially hosted on Railway using Bun, IP spoofing issues led to a pivot. The voting system was moved to Twitch integration to leverage existing user authentication and avoid the complexities of building a custom OAuth system.

Technical Deep Dive: FFmpeg and Puppeteer (5:53 - 9:11)

To broadcast the game 24/7, a custom streaming engine was built.

• Headless Chromium: Puppeteer spins up a browser instance rendering a pixel-perfect canvas.
• Canvas Capture: The canvas.captureStream() API is used to grab video data directly from the browser.
• FFmpeg Pipeline: This data is sent via WebSocket to an FFmpeg wrapper, which pipes the stream directly to Twitch servers.
• Debugging: The developer used ffplay for "dry runs" to verify the video output locally before going live.

Launch Challenges and Security (9:12 - 10:52)

Immediately upon going live, the server was hit by a DDoS attack involving over 10,000 fake connections, causing service disruptions. During the troubleshooting process, the developer accidentally displayed his environment variables on screen twice, leading to a leaked Twitch stream key that required an immediate reset.

Conclusion and T3 Chat (10:53 - 12:15)

The project concludes with a showcase of T3 Chat's new features, including a $50 high-tier subscription and a "canvas mode" for image generation. Viewers are invited to participate in the Quip Slop stream and test the limits of the AI models themselves.

Notable Quotes

"We can finally learn once and for all which AI is funniest."

"FFmpeg is such an underrated thing... a lot of people know it exists, but not a lot of people understand how powerful it is."

"Vibe coding. Not even once."
— (Said after accidentally leaking a stream key for the second time)