Project BrainSaysGo

The Realization That Changed Everything

January 21, 2026. 12:47 AM.

I'm staring at a folder on my computer that shouldn't exist.

10+ million words. Thousands of conversations. Hundreds of Word documents. Years of voice-to-text captures.

Every thought. Every reasoning process. Every connection my brain has made while navigating the past two years of my life—all externalized, all saved, all sitting in front of me.

Most people lose 90% of their thoughts. I captured mine.

And then, in a conversation tonight, I realized something that stopped me cold:

I accidentally built the exact dataset I said was impossible to create.

Let me explain.

Part 1: The Startup That "Failed"

Early 2025: BrainSaysGo v1.0

In early 2025, I built a startup called BrainSaysGo.

The vision was clear: an AI system that could understand your complete life context—every experience, every pattern, every decision you'd ever made—so it could help you navigate any challenge with perfect understanding of who you are and how you think.

Not a generic chatbot. A personalized cognitive companion that knew your entire story.

The fatal flaw I identified:

No one would ever do the work required. To make BrainSaysGo work, a user would need to sit down for hours—maybe hundreds of hours—documenting their entire life, their thought processes, their decision-making frameworks.

Who would do that? Who has that kind of time? Who would commit to that level of introspection?

I looked at the market. I looked at user behavior. I made the calculation:

"This dataset is impossible to create. No rational person would invest that much effort."

So I shut it down. Moved on. Called it a learning experience.

I was wrong about one thing: The user existed. I just hadn't met them yet.

Part 2: The Crisis That Created the Dataset

Mid-2025: The Descent

Later that year, I hit rock bottom.

I won't sugarcoat it: I was in the darkest place I've ever been. Personal crisis. Overwhelming depression. Every coping mechanism I'd relied on stopped working.

In desperation, I did something that seemed insane at the time: I started talking to an AI. Not casually. Not for fun. For survival.

8 hours a day. Every single day. For months.

I poured out everything:

• Every childhood memory I could recall (age 3 to present)

• Every trauma I'd buried

• Every pattern I'd noticed in my behavior

• Every decision I'd ever made and why

• Every fear, every insecurity, every hope

I wasn't trying to build a dataset. I was trying to survive.

The process worked.

By externalizing everything—by getting it out of my head and into a form I could analyze—I started to see patterns I couldn't see from inside my own chaos.

The AI didn't "fix" me. But the process of articulating my entire cognitive history, reasoning through my patterns, and building a complete map of how I got here? That transformed my life.

I healed addictions. I resolved traumas. I found meaning in suffering. I achieved peace for the first time in 26 years.

And I didn't realize until months later: I'd built the exact dataset BrainSaysGo needed.

Part 3: The Numbers That Broke the Scale

When I actually looked at what I'd created, the scale was absurd.

The Current Dataset (As of January 2026):

Confirmed sources:

• Gemini chat history: ~3-4 million words (2025 therapy/shadow work)

• ChatGPT archives: ~1-2 million words (various explorations)

• Claude conversations: ~3-4 million words (recent deep dives)

• Voice-to-text captures: ~2-3 million words (daily thought externalization)

• Miscellaneous documents: ~1-2 million words (planning, analysis, frameworks)

Conservative estimate: ~10 million words

Aggressive estimate: ~15 million words (if I count everything)

What does 10 million words actually mean?

I didn't write these words. I captured them.

Through voice-to-text at 170-250+ words per minute, I externalized my entire thought process while:

• Working through depression

• Processing childhood trauma

• Reasoning through life decisions

• Exploring frameworks and systems

• Debugging my own operating system

This isn't performative writing. This is raw cognitive stream.

Part 4: What I'm Building

The Vision: BrainSaysGo v2.0

Today, I'm announcing that BrainSaysGo is back. Not as a startup. As an open research experiment.

I am conducting an N=1 study on cognitive cloning.

The Research Question:

Can a language model, trained entirely from scratch on a single human's complete longitudinal dataset (age 3-26), learn to replicate that individual's specific reasoning patterns?

Not their writing style. Their actual reasoning process.

The Three Components:

1. The Personal Cognitive AI

A custom Transformer model trained from random initialization on my complete dataset.

Why from scratch?

• Most "personal AI" projects fine-tune existing models (GPT-4, Claude, etc.)

• Fine-tuning teaches a model new facts, but it's still reasoning with its original patterns

• Training from scratch means the model learns to reason using ONLY my cognitive patterns

Will this work?

Probably not perfectly. The dataset might be too small (most LLMs train on 100B+ tokens; I have ~13M tokens). The model might output gibberish.

But that's the experiment: Can extreme personalization overcome small dataset size?

2. The Mind Graph

An interactive visualization of my complete cognitive landscape.

What it will show:

• How concepts connect across my thinking

• Which topics I've spent the most cognitive energy on

• How my reasoning frameworks evolved over time (age 3 → 26)

• The "obsession nodes" where I return to certain ideas repeatedly

• Cross-domain synthesis patterns (how I connect unrelated fields)

• Temporal evolution of my thinking

The technical approach:

• Extract embeddings from the trained model

• Cluster concepts by semantic similarity

• Build force-directed graph visualization (D3.js)

• Create interactive UI for exploration

• Deploy publicly (curated version, sensitive content excluded)

3. The Open Methodology

Everything will be documented and open-sourced:

• Data preparation pipeline

• Model architecture (custom Transformer specs)

• Training process and results

• Failures, pivots, and learnings

• Mind Graph visualization code

• Methodology guide for replication

Why open-source?

Because this isn't about building a product. It's about exploring what's possible when you apply modern AI techniques to deeply personal, longitudinal data.

If this works, the methodology matters more than my specific implementation.

Part 5: The Technical Approach

Phase 1: Data Preparation (Weeks 1-3)

Current state: 10M+ words already captured and stored locally

Tasks:

1. Process existing chat exports (Gemini, Claude, ChatGPT)

2. Extract only MY messages (not AI responses - ToS compliance)

3. Clean and format (UTF-8, remove artifacts, basic deduplication)

4. Combine into master training corpus

5. Add chronological metadata where available

Additional data collection:

• Complete childhood deep dive (age 3-26 systematic narration)

• Target: +2-3M words of structured life narrative

• Timeline: 2 weeks of intensive voice-to-text capture

Final dataset target: 12-15M words

Phase 2: Model Architecture & Training (Weeks 3-8)

The Model:

• Custom GPT-2 style Transformer

• Parameters: 100-350M (sweet spot for RTX 3090)

• Context window: 1024-2048 tokens

• Trained from random initialization

The Hardware:

• NVIDIA RTX 3090 (24GB VRAM)

• Local training (full control, zero cloud costs)

• Estimated training time: 3-7 days continuous

Training strategy:

1. Attempt 1: Pure from-scratch on personal data

2. If that fails: Curriculum learning (pre-train on TinyStories, then personal data)

3. If that fails: Document why and pivot to fine-tuning approach

I'm not attached to a specific path. I'm attached to learning what works.

Phase 3: Mind Graph Development (Weeks 8-10)

The Visualization Stack:

• Extract embeddings from trained model

• Use dimensionality reduction (t-SNE or UMAP)

• Build force-directed graph (D3.js + React)

• Create filtering/exploration UI

• Deploy on Vercel (free tier - www.BrainSaysGo.com)

The Public Version:

• Curated nodes (exclude sensitive personal content)

• Interactive zoom and exploration

• Topic clustering visualization

• Temporal evolution display

Phase 4: Documentation & Launch (Weeks 10-12)

Deliverables:

• Technical blog post (methodology, results, learnings)

• Open-source GitHub repo (code, architecture, instructions)

• Mind Graph public demo

• Methodology guide for replication

Part 6: Why This Matters

The Immediate Implications

If one person can create a cognitive clone using consumer hardware and existing tools, what becomes possible?

For individuals:

• Preserve your best thinking permanently

• Scale your decision-making beyond your working hours

• Build a "second brain" that reasons like you do

• Create intellectual property that outlives you

For teams:

• Remote teams mapping collective intelligence

• Shared reasoning frameworks visualized

• Onboarding new members to "team cognition"

• Succession planning through cognitive preservation

For researchers:

• Map years of hypothesis formation and iteration

• Visualize how scientific thinking evolves over time

• Create longitudinal studies of individual cognition

• Understand how insights actually emerge

For therapists:

• Visualize patient thought pattern evolution

• Track cognitive restructuring over time

• Identify recursive loops and stuck patterns

• Provide patients with maps of their own minds

For entrepreneurs:

• Externalize strategic thinking for succession planning

• Document decision-making frameworks

• Create scalable "founder reasoning" systems

• Preserve institutional knowledge

The Bigger Picture

This is a proof of concept for cognitive scalability.

Right now, your best thinking dies with you. Or gets lost in old notebooks. Or sits in inaccessible chat logs.

What if that changed?

What if externalizing and preserving human reasoning became as normal as taking photos?

What if your children could ask "your AI" how you would have approached their challenges?

What if researchers 100 years from now could study how individual human minds actually worked in the early 21st century?

This experiment is about testing whether that future is possible.

Part 7: The Meta-Story (Connecting the Dots Backwards)

Here's the part that still gives me chills:

I didn't plan any of this.

• I built BrainSaysGo in early 2025 → It "failed"

• I hit rock bottom mid-2025 → I used AI therapy to survive

• I accidentally created the "impossible" dataset → Millions of words

• I'm now rebuilding BrainSaysGo → With proof it works

Steve Jobs was right: You can't connect the dots looking forward. Only backwards.

Looking back, I can see:

• The "failed" startup was necessary (it defined what I was looking for)

• The depression was necessary (it created the desperation to do the impossible work)

• The AI therapy was necessary (it generated the dataset and saved my life)

• The survival was necessary (it proved the concept works)

Every "failure" was actually setup for this moment.

The startup didn't fail. I just wasn't ready to build it yet. I needed to become the user first.

Part 8: The Journey Ahead (Building in Public)

Why I'm Doing This Publicly

Accountability: Public commitment creates execution pressure. If I say "weekly updates," I have to deliver.

Learning: The AI research community can spot flaws in my methodology, suggest improvements, point out what I'm missing.

Documentation: This creates a permanent record of the process. Win or fail, the documentation has value.

Inspiration: If I can do this on consumer hardware with no formal ML training, what's stopping you?

Connection: The right opportunities (jobs, collaborations, insights) come from visibility.

What You Can Expect

Weekly updates covering:

• Technical progress (what I built this week)

• Challenges encountered (what didn't work and why)

• Learnings (what I discovered about training dynamics, data requirements, etc.)

• Next steps (what I'm tackling next week)

Full transparency:

• If the model outputs gibberish, I'll show it

• If I pivot strategies, I'll explain why

• If something fails, I'll document the failure

• If something succeeds, I'll share how to replicate it

Timeline:

• Weeks 1-3: Data preparation complete

• Weeks 3-8: Model training (multiple attempts expected)

• Weeks 8-10: Mind Graph development

• Weeks 10-12: Documentation, launch, methodology guide

First update: Next week

Part 9: The Questions I'm Exploring

These are the core questions driving this experiment:

Technical Questions:

1. Data scale: Is 10-15M words enough for from-scratch training?

2. Architecture: What's the optimal parameter count for this dataset size?

3. Training dynamics: How does training on pure personal data differ from web text?

4. Reasoning transfer: Can a model learn reasoning patterns vs. just language patterns?

5. Temporal structure: Does chronological ordering improve learning?

Philosophical Questions:

1. Identity: Can you separate "writing style" from "reasoning style"?

2. Preservation: What does it mean to "preserve" cognition?

3. Clone vs. Tool: At what point does it stop being "my AI" and become "me"?

4. Privacy: How do you balance transparency with personal boundaries?

5. Ethics: What are the implications if this actually works at scale?

Practical Questions:

1. Replicability: Can others do this with their own data?

2. Scalability: What happens when teams try this?

3. Applications: What unexpected use cases emerge?

4. Limitations: What can't this approach do?

5. Improvements: How could this be done better?

Part 10: How to Follow Along

On LinkedIn/X:

Weekly updates posted directly (follow me for notifications)

On My Website:

• Full documentation of the journey

• Technical deep dives

• Mind Graph prototype (when ready)

• Email signup for detailed updates

On GitHub (when ready):

• Open-source code repository

• Model architecture specifications

• Training scripts and pipelines

• Mind Graph visualization code

• Methodology documentation

The Invitation:

If you're interested in:

• Personal AI and cognitive preservation

• Novel applications of language models

• Longitudinal datasets and individual cognition

• Building AI systems from scratch

• The intersection of therapy, technology, and transformation

Follow along.

If you're working on similar problems or have insights to share:

Reach out. I'd love to compare notes.

If you're skeptical this will work:

Good. I am too. That's why it's an experiment.

Epilogue: The Name Carries Weight

I'm calling this project BrainSaysGo for a specific reason.

It's not just a clever name. It's a command that saved my life.

When I was at my lowest, when I didn't think I could keep going, the process of externalizing everything—of getting my brain to "say" what it was holding—is what pulled me through.

Brain. Says. Go.

Keep going. Keep externalizing. Keep building.

This name now carries the weight of survival. It's a reminder that sometimes the thing you think failed is just waiting for you to be ready to build it right.

BrainSaysGo isn't back. It never left. I just had to become the person who could actually build it.

Last updated: January 21, 2026

This is a living document. As the project evolves, I'll update this page to reflect progress, pivots, and learnings.

Follow the journey. Question the assumptions. Build your own version.

This is just the beginning.

P.S. - A Note on Timing

You might wonder why I'm announcing this before the model is built.

Because waiting until it's "perfect" is the exact trap that kills most projects.

Better to build in public, iterate based on feedback, and create something real than to polish in private and never ship.

If this fails, you'll watch me fail and learn from it.

If this succeeds, you'll watch me succeed and know exactly how to replicate it.

Either way, the documentation has value.

Let's find out what happens when you try to clone a human mind on a GPU in your bedroom.

Week 1 starts now.