Truffle

I asked ChatGPT for startup ideas 5 times. All 5 were some variation of "AI-powered [X] management tool."

My friends got the same ones. We're all converging on identical ideas — using the same AI, getting the same output. A 2024 study (Doshi & Hauser, Science Advances) showed that AI boosts average creative quality but reduces the variance of novelty across people. The floor rises, but everyone clusters toward the same ceiling.

For indie makers, your edge is originality. And AI is quietly eroding it.

But there's a deeper issue: AI only answers what you ask. It can never find what you never thought to ask.

TOYOTA'S JIT WAS BORN FROM A SUPERMARKET

The often-told story: in the 1950s, Toyota's Taiichi Ohno visited an American supermarket. Customers took only what they needed; the store restocked only what was sold. He transferred this causal structure to the car factory — downstream pulls, upstream responds. Just-In-Time manufacturing was born.

A supermarket and a factory look nothing alike. What they share is a causal structure: demand pulls, supply responds. This kind of reasoning — transferring causal structure, not surface similarity — is called structure-mapping in cognitive science (Gentner, 1983).

The problem: this process relies on serendipity. If Ohno hadn't walked into that supermarket, JIT might never have happened.

I TRIED THIS WITH MY OWN NOTES

I keep notes about dance, investing, software dev, startups — different areas of my life. I extracted causal relations from each and compared them structurally.

My dance notes said: "Improvising to music with no choreography. There's no concept of failure. You just respond to the music."

My investing notes said: "Auto-sell rule at -10% loss. Portfolio dropped -15% in the crash. Recovered after one month."

Extracting the structures, I found matching causal chains:

• Dance: Repetition → body moves automatically

• Investing: Rule-making → emotions get removed

• Dance: Stop thinking → you can improvise

• Investing: Mechanical rules → detach judgment from feelings

• Dance: No concept of failure → only process matters

• Investing: Evaluate by profit/loss → ???

The first two pairs have matching causal structures. The third is where it gets interesting.

Both describe the same structural position: "what counts as success." In dance, it's process — did you stay in the flow? In investing, it's results — did you make money? That mismatch is the blindspot.

In cognitive science, this is called a "candidate inference" — a relation that exists on one side of a structural alignment but is missing on the other. It's the thing you never thought to ask about.

The question it surfaced: "What if you tracked how well you followed your own rules, instead of tracking profit/loss?"

I wouldn't have asked ChatGPT this. Not because ChatGPT can't answer it — but because the question requires knowing about my dance experience to even form. You can't prompt for what you don't know is missing.

And it's practical. There's evidence in investing that process adherence maximizes long-term returns. But almost every investor stares at P&L every day. A "process compliance dashboard" redefines what failure means — exactly the structural transfer from dance.

WHAT I'M BUILDING

So I built a CLI tool called Truffle that automates this: extract causal relations from your markdown notes, compare structures across folders, and surface the gaps as questions — not answers. The "aha" moment belongs to you.

WHERE I AM NOW

I compared three methods on the same test data (43 AI-generated notes, 12 evaluated outputs — tiny sample, I know):

- Method A (baseline): Ask an LLM "find blindspots" → 53% good rate

- Method B: Classify notes into 54 categories → 63%

- Method C (structure-mapping) → 83%

The gap between A and C keeps me going. But the most important finding was this:

Method C initially scored 55% — same as the LLM baseline. The breakthrough came from one change: filtering out generic causal relations during extraction. "Explaining helps understanding" is true but useless — it applies to everything. Once I kept only domain-specific relations, quality jumped. Generic in, generic out.

HONEST LIMITATIONS

- Test notes were AI-generated, not real messy human notes

- 12 outputs evaluated — small sample, self-evaluated (obvious bias risk)

- The mapping algorithm has issues (string similarity can't capture semantic correspondence)

- Haven't tested on anyone else's notes yet

WHAT'S NEXT

- Testing on my own real notes

- Improving the mapping algorithm (semantic similarity)

- Weekly improvement cycles targeting 70%+ good rate on real data

- Eventually: Obsidian plugin

Revenue: $0. Users: just me. Building this because I want a tool that finds ideas I can't find by talking to AI.

QUESTIONS FOR YOU

- Do you feel like AI brainstorming gives you the same ideas as everyone else?

- Have you ever had a breakthrough from connecting two completely unrelated areas of your life? I'd love to hear the story.

Building this in public. Will share results — including failures — as I go.