PowerUp Coaching — Living Knowledge Base

2025-09-11_Mastermind

5 min read

2025-09-11 AI Mastermind

Table of Contents

Session Overview

The September 11th session opened with a live introduction to K2Think.ai — a Dubai-built, open-source mixture-of-experts model running on Cerebrus hardware that Lou had discovered and was immediately impressed by for its reasoning depth and processing speed (approximately 10x faster than Claude or OpenAI on equivalent hardware). The session positioned this as a model worth adding to the “specialist tool” arsenal rather than replacing primary models, and introduced the concept of using a deep-reasoning model to generate a brief that is then handed to writing-focused models.

The session’s main technical substance was a live demonstration of Claude Code and the PRD-first development workflow. Lou walked through an active coding experiment — a Pinecone database interface for comparing RAG retrieval quality — showing the group how Claude Code initialized from the codebase, managed its own claude.md context file, and executed incrementally from a clean brief through to a working Streamlit prototype. The key methodological contribution was the explicit sequencing: concept document → PRD → phased implementation plan → phase-by-phase execution with Git commits.

Members Donald and Kasimir shared their Docker and N8N local installations, and the session confirmed that several members were actively deploying local development environments and preparing to move to cloud hosting. This represented meaningful progress from earlier sessions where the infrastructure discussions had felt abstract.

High-Signal Moments

  • K2Think.ai introduced as a high-speed reasoning model from Dubai’s sovereign AI program, open source, running on Cerebrus architecture, beating all open-source models and competing with GPT-4o on reasoning benchmarks
  • Workflow introduced: “Use K2Think.ai at temperature=1 for the thinking/brief phase, then hand the brief to Claude or GPT for the writing/execution phase” — separating reasoning from generation
  • PRD-first development workflow demonstrated live: brainstorm → save to file → generate PRD → generate phased plan → execute phase 1 → git commit → branch for phase 2
  • Streamlit hack shared: “I always tell it to make a very simple Streamlit interface — it’s not React or Tailwind, but it gets up and running super quick so you can test everything’s working before spending time on the user interface”
  • Cole Medin’s RAG template architecture confirmed as the foundation for Lou’s legal AI app and the shared AIMM instance (Qdrant + Ollama + Open Web UI + Docker)
  • “Those four — Qdrant, Docker, Ollama, Open Web UI — you can build pretty much any application you want that requires data”
  • Kasimir’s insight on cost discipline: running N8N locally first to validate the workflow before committing to cloud hosting, saving subscription costs during development
  • Discussion of available AI coding agents: Cursor, Warp, CLIM, RooCode vs. Lou’s preference for VS Code with Codex/Gemini/Claude CLIs installed

Open Questions

  • K2Think.ai is new — how does it perform on coaching-specific tasks (synthesizing frameworks, generating client exercises, nuanced communication) vs. its demonstrated strength in reasoning/analysis?
  • At what point does a Streamlit prototype need to be replaced with a production UI, and what is the right migration path for non-developers?
  • How should Git be introduced to members who have never used version control, given how central it is to the PRD-first workflow?
  • What is the decision framework for choosing between Cursor/RooCode/Claude Code — when does each shine?
  • Is the Hugging Face leaderboard the right place to track emerging open-source models, or are there better monitoring resources?

Suggested Follow-Through

  • Lou: Share K2Think.ai link in Telegram with a brief prompt recommendation for first-time use
  • Lou: Create a short “first app” tutorial using the PRD-first workflow with a simple example that members can replicate in under 2 hours
  • Members: Check Cole Medin’s GitHub repo for the Open Web UI + Qdrant + Ollama Docker Compose template and test deployment locally
  • Members interested in coding: Install VS Code and at least one AI CLI (Claude Code, Codex, or Gemini CLI) and run a simple experiment using the PRD-first workflow
  • Follow Hugging Face leaderboards monthly to track open-source model quality for specific use cases

Additional Resources

Links & Tools Shared in Chat

Books & Articles Mentioned

  • None

Ideas from Chat

  • Donald Kihenja: “Going forward, I wouldn’t recommend any young person study computer science. I would tell them master English (or their native language) and logic/critical thinking instead.” — a strong provocation about where durable human value lies in an AI-saturated world
  • Don Back: “Information is essentially free — Transformation/implementation are the valuable pieces” — a useful reframe of where coaches and knowledge entrepreneurs should focus
  • Don Back: delivering a seminar to Chemistry students facing AI disruption — exploring how to humanly share the message that their technology skills alone won’t be sufficient

Graph View