13h • General discussion
AI Developer Accelerator — Coaching Call - September 8th
AI Developer Accelerator — Coaching Call - September 08
Meeting Purpose
Key Takeaways
- Astra (GPT-6) is a powerful "computer use" agent, excelling at UI testing and complex research by autonomously navigating websites. This capability is so effective that Ty Wells is using it to replace a human QA role.
- Fable 5.1 (Claude) remains the preferred model for creative "thinking" tasks, especially when tuned to "medium" thinking to manage token costs. The consensus is that Fable excels at uncovering "unknown unknowns" and blind spots.
- The market for AI services is saturated in common areas (e.g., marketing, finance), creating a strong incentive to find niche applications. This requires leveraging one's unique business network and non-technical expertise to identify problems AI can solve.
- Personalized AI assistants, fed with continuous context (e.g., daily voice notes, screen activity), are proving highly effective for generating novel ideas. This approach turns the AI into a proactive "thinking partner" that surfaces opportunities from daily life.
Topics
AI Model Performance & Usage
- Strengths: Excels at "computer use" tasks, such as UI testing and complex web research.
- Weaknesses: Can produce inconsistent code ("moments of brilliance" mixed with "dumbass things").
- Cost: High token consumption necessitates the $100/month plan for serious use.
- Strengths: Preferred for creative "thinking" tasks and uncovering "unknown unknowns."
- Cost Management: Tuning to "medium" thinking is a key strategy to manage token costs.
- Strengths: A highly cost-effective alternative for agentic tasks.
- Example: Patrick Chouinard reduced his Recap Flow run cost from ~$1.00 to $0.09 by switching to GLM 5.3 Flash.
Project Updates & Challenges
- Problem: A RAG agent using GLM 5.3 is failing its 50-question test harness due to inconsistent answers.
- Hypothesis: The issue is either a poor model choice or a flawed RAG setup.
- Recommendation: First, establish a baseline with a top-tier model (e.g., Fable) to isolate the RAG system's performance before optimizing for cost.
- Project: Built AgentDeck, an observability surface to monitor all live agents (Claude, Codex, Hermes).
- Functionality: Tracks token usage, session history, and unresolved attention requirements, consolidating agent interactions into a single inbox.
- Tech Stack: Prometheus, Alertmanager, Loki, Grafana, NATS, Auth0, Traefik.
- Project: Launching an AI-powered ERP on Sept 15, with Astra handling UI QA.
- Key Feature: A "Business Advisor" that guides new businesses through vetting their ideas and helps existing businesses achieve goals by analyzing their data.
- Problem: Building a local-only RAG system for university students but needs a security review.
- Challenge: Claude (Opus vs. Fable) is providing conflicting security advice.
- Solution: Ryan C offered to run the code through a security review suite.
- Strategy: Offering a photo automation service for free to mid-tier venues (country clubs) to build a portfolio and generate content.
- Future Feature: Considering adding printing capabilities, a feature requested by a venue coordinator.
Finding Market Opportunities
- Challenge: The AI market is saturated in common areas, making it difficult for new entrants to find a niche.
- Strategies:
- Niche Markets: Find underserved markets (e.g., cemeteries, outdated internal software).
- Leverage Existing Networks: Use non-technical industry connections to find problems.
- Solve Past Frustrations: Use AI to solve problems that were previously too costly or complex (e.g., translating cryptic part numbers).
- Build Credibility: Create targeted projects and content based on real-time market needs (e.g., using a GrokBot to find trending LinkedIn topics).
Advanced AI Workflow Techniques
- Daily News Radar: Schedule a task for ChatGPT to summarize news relevant to your projects and goals.
- Interactive Discussion: Engage in a daily verbal conversation with the AI about the news to refine its understanding and generate ideas.
- Contrarian Prompts: Instruct the AI to be "confrontational" and surface opposing viewpoints to avoid tunnel vision.
Next Steps
- Hemal Shah: Post RAG agent issue in the forum; test GLM 5.3 Flash.
- Tim Hildenbrandt: Connect with Ryan C for a security review.
- Ryan C: Test Astra's "computer use" capabilities for video editing.
- Juan Torres: Develop outreach strategy for mid-tier venues.
- Kris Larson: Explore niche markets; use a GrokBot to find LinkedIn topics; consider joining a local business association.
- All: Experiment with personalized AI workflows (e.g., daily news radar, continuous context).
Action Items
- Post RAG eval/test-harness details in community forum (Paul, Patrick) - WATCH (5 secs)
- Evaluate GLM 5.3 Flash for RAG agent (Patrick) - WATCH (5 secs)
- DM Tim Hildenbrandt in community re: security review; request code/context - WATCH (5 secs)
- DM Ryan C in community re: security review; share code/context - WATCH (5 secs)
- Build venue outreach list; start email campaign; visit venues; offer free trials - WATCH (5 secs)
- Evaluate Astra Computer Use for video editing workflows - WATCH (5 secs)
- Watch Nate B. Jones Astra video (Patrick) - WATCH (5 secs)
0
0 comments
AI Developer Accelerator — Coaching Call - September 8th
skool.com/ai-developer-accelerator
Master AI & software development to build apps and unlock new income streams. Transform ideas into profits. 💡➕🤖➕👨💻🟰💰
Leaderboard (30-day)
1
+7
2
+7
3
+6
4
+5
5
+3
Powered by