AIMM Session — July 10, 2025: The Vibe Coding Trap and the Grandma Test

“Every time you want to learn something, save up your two grand and hire some guy — or gal — to do it for you and record it. That’s the best use of money I’ve ever found. Better than any course.” — Lou

30-Second Summary

This week’s call opened with a confession: Lou has been deep in a coding rabbit hole, and he’s not sure he’d recommend it to anyone who doesn’t have to be there. From that honest starting point, the group unpacked one of the most important strategic tensions knowledge entrepreneurs face — the seductive pull of learning everything versus the business case for outsourcing intelligently. The conversation ranged from ROI frameworks and the “Grandma Test” for investment decisions, to a live demo of the Dia browser as a research-to-AI context machine. Hot topics included managing multi-session AI projects, which LLM to use for which job, and one member’s story of building a voice recruiter bot that collapsed elegantly into four lines of code.

1. The Vibe Coding Reality Check

The hype is fun — the maintenance is not.

Lou opened with a personal project debrief: he’s been building a local AI application using Qdrant (vector database) and N8N, and the experience has only strengthened his conviction that vibe coding is a trap for most knowledge entrepreneurs.

The hidden cost isn’t the software — it’s the distraction. Every open-source package you self-host (Qdrant, N8N, Llama) comes with a maintenance tail: security updates, version conflicts, broken dependencies. What starts as “free” becomes a second job.

Lou’s practical recommendation: Browse N8N’s workflow repository first. Find something close to what you need. Then go to Fiverr or Upwork and say “implement this for me.” You’ll pay $150–200, save a day of work, and never have to touch it again when the next update drops.

“Instead of thinking ‘I could save a couple thousand bucks if I learned this myself,’ ask: if I spent that time on client conversations, how long would it take to earn that money back?”

Alex Flueck added nuance from experience: after 30 years running Linux HPC clusters for his research lab, he can do all of this — but he agreed with Lou’s framing entirely. Knowing the landscape well enough to hire well is the real skill worth developing.

Don Back (in chat): “Clients don’t understand technology debt until it bites them in the…“

2. The “Should You?” Framework: Three Kinds of ROI

Not every investment needs to pay off in dollars.

Jay Drobez pressed Lou on how to think about ROI when making AI investment decisions. Lou’s answer introduced a three-part framework:

• Financial ROI — straightforward: will this generate more than it costs? Be brutally honest. The test: could you invest ten times as much and still believe the numbers?
• Emotional ROI — are you doing this because it genuinely energizes you? Valid. But don’t let it masquerade as a business decision.
• Intellectual ROI — the murkiest category. You believe it makes you more valuable, more rounded, more fluent. Lou: “Sometimes I just do it because I want to. I don’t calculate it.”

The key move: earmark it consciously. Mazie Zdanowicz reframed it cleanly — treat AI education like a college fund for your career. Set a dollar figure, protect it, and stop feeling guilty about spending it.

Mazie: “Not every course has to have an ROI for the business. Allow yourself a little.”

Waldon Moss (in chat): “I enjoy learning the concepts, but need to be careful about allowing it to become a distraction. Want to stay informed enough to proficiently vet vendors, contractors, etc.”

3. The Grandma Test

If you wouldn’t invest your grandmother’s money in it, you don’t believe in it enough.

Lou’s most quotable moment of the session: you think you’re going to 10X your output with this AI tool or workflow. Great. Now imagine your grandmother handed you $100,000 and said, “Give me half back when it works.” Would you take that deal?

If the answer is “maybe not,” you’ve just revealed that you don’t actually believe the projection.

Donald Kihenja (in chat): “The ‘Grandma test!’ — if you think you can justify spending on something based on ROI, then ask yourself, ‘Would I take my grandma’s money if she wanted to invest in this?‘“

4. The Outsource-to-Learn Hack

Stop paying for courses. Start paying for implementations with the camera rolling.

Hire someone to build your actual system — and ask them to record the whole process. You get:

• A working implementation tailored to your exact use case
• A video walkthrough with zero extraneous content
• The technical vocabulary to hire the next person intelligently
• A reference you can rewatch rather than a general-purpose course you have to adapt

Bonus negotiation: Offer to be their case study. They can share the recording with their audience; you get a discounted rate and a real-world example.

“I could have paid someone $2,000 to do this for me and record it. That would have been the best educational investment I ever made.”

5. Managing Multi-Session AI Projects (The Collective Pain Point)

Nobody has a great solution yet — but here’s what’s working.

Donald Kihenja surfaced a problem the whole group recognized: when you’re using multiple AI tools across multiple sessions, retracing your steps becomes its own project.

What’s working (imperfectly):

• Don Back: Copy-paste key outputs into a titled Word doc immediately.
• Donald Kihenja: Export to Google Docs, paste the URL to the original chat at the top, and manually add each key prompt as you go.
• Lou: The Paste clipboard manager (available in SetApp) lets you stack multiple copies with Cmd+Shift+C.
• Dirk: Uses Zapier to write Claude outputs directly to Google Drive docs.

Lou floated the idea of an MCP + N8N workflow that auto-exports a conversation to Google Drive on command.

6. Choosing Your AI Model Stack (Without Chasing the Leaderboard)

Pick by function, not by this week’s benchmark.

Lou’s current mental model:

• ChatGPT (4o / O3) — personal memory, knows Lou’s perspective, good for brainstorming in his voice
• Claude — writing for publication, voice refinement, coding (Claude Code with Max plan is “miraculous”)
• Gemini — free terminal access, good enough for simple projects
• DeepSeek — reasoning-heavy tasks
• GPT-4o Mini / O4 Mini — labeling, summarizing, categorization (cost-efficient)
• OpenRouter — if you genuinely need to swap models dynamically

The key insight: the top 10 models are separated by 0.2–1% on benchmarks. Swapping models every week to chase marginal gains means rewriting and re-testing your prompts constantly. Not worth it.

Development philosophy: Make it work → make it right → make it fast.

7. The Dia Browser Demo: Live Tabs as AI Context

What if your browser tabs were your knowledge base?

Using the Dia browser (from The Browser Company, makers of Arc), Lou showed how open browser tabs automatically become context for AI conversations — without copy-pasting, without uploading files.

Lou described it as “NotebookLM but for live browsing — less cumbersome, no copy-paste required.”

The power move: Open documentation pages for every tool you’re using, save them to a bookmark folder, then import them all as context at once. You get a workspace that references current documentation — not the AI’s training data.

Donald (in chat): “For any non-Mac users who want a similar experience, there’s also Sidekick browser.”

Hot Takes & Resources

• Dirk’s ElevenLabs Experiment: Built a voice chatbot for candidate screening. ElevenLabs support’s AI-written response analyzed his code and collapsed his entire chatbot implementation into four lines. “Even better than what I built. Super fast.”
• Kasimir’s Niche: There’s a white space at the intersection of consciousness development and AI.
• The “How Hole”: Donald named it. Don Back described falling into it. The place you go when you’re fascinated by how something works and lose hours to it.
• Claude Code on Max Plan: Lou says he’s considering upgrading to the $200/month Max plan. “Easy decision — I’m saving hours and hours of writing code.”

Try This Before the Next Call

The Outsource-to-Learn Experiment

Pick one AI workflow or tool you’ve been meaning to implement. Instead of taking a course:

1. Find a freelancer on Fiverr, Upwork, or a relevant community who has done something similar
2. Ask them to implement it for you and record the whole session
3. Offer to be their case study for a discounted rate
4. Watch the recording with your specific use case in mind

Start with $200–500. You’ll get a working implementation and a personalized tutorial.