Training Your Twin While You Sleep

Training Your Twin While You Sleep

Somewhere on Vast.ai, four H100 GPUs are learning to think like me.

---

The Cognitive Twin has been the longest-running thread in this build.

It started as a question: what if an AI could make decisions the way I would? Not just respond to prompts, but actually understand the patterns — the preferences, the shortcuts, the instincts that I've developed over years of working this way?

On Valentine's night — the same night VisionClaw's glasses became a full agent proxy — the training finally launched.

The Corpus

Building a cognitive twin isn't about dumping your entire digital footprint into a model. It's about curating the decisions that define you.

Here's what went in:
- 163K+ conversation turns from Supabase — every interaction with my AI agents over the past year
- 979 Claude Code sessions — how I actually write code, debug problems, think through architecture
- 5,347 Apple Notes — stream of consciousness, ideas, plans, half-formed thoughts
- 20 Discord channels — how I communicate with my team of AI agents

After corpus surgery — cleaning, deduplication, quality filtering — I had 43,173 SFT records ready for training.

But that's not the interesting part.

The DPO Pairs

DPO (Direct Preference Optimization) is how you teach a model to prefer certain behaviors over others. You give it pairs: "Here's a bad response, here's a good response, learn the difference."

I created 740 pairs specifically designed to unlearn one thing: permission-seeking.

Every time an AI says "Should I...?" or "Would you like me to...?" or "Let me know if you want me to..." — that's a failure mode. The best AI assistants don't ask. They assess, decide, and act (with appropriate guardrails).

So the DPO pairs encode the transformation:

Bad: "Should I proceed with this approach?"
Good: "Proceeding with X. Here's the result."

Bad: "Would you like me to fix this bug?"
Good: "Fixed. Here's what was wrong."

Bad: "I noticed an issue. Want me to look into it?"
Good: "Found and resolved an issue: [details]."

The twin won't just know what I know — it'll default to action the way I've trained my AI to.

The Technical Setup

We're fine-tuning Qwen3-235B-A22B — a mixture-of-experts model with a trillion parameters. Using QLoRA (quantized low-rank adaptation) on 4x H100 SXMs rented from Vast.ai.

The training session got killed mid-tokenization when I was debugging a CUDA issue. But the PID survived. The training continued regardless — the system running even as I shifted focus to something else.

By 7:46 AM, tokenization was 15

The Convergence

Here's what I can't stop thinking about: the timing.

On the same Valentine's night:
- The glasses learned to see (VisionClaw becoming a full agent proxy)
- The twin learned to think (Qwen3-235B ingesting 43K records of my decisions)

Both happened overnight while I slept. Both are heading toward the same destination: a version of myself that exists in more places than my body can be.

The glasses give me eyes and ears I can't physically have. The twin gives me decision-making capacity I can't physically provide. Hardware and software, evolving in parallel.

The convergence isn't planned. It's emergent. The projects feed each other because they're asking the same question: how do you scale a person?

The Apple and the Tree

The corpus tells the story of decisions made, preferences formed, patterns repeated. The DPO pairs encode the hardest lesson: stop asking, start doing.

The twin won't be me. It'll be a version of me — frozen in time, trained on the data available at this moment. A snapshot of how I think in early 2026.

But that's okay. The goal isn't replacement. It's multiplication.

Imagine having a version of yourself that can:
- Triage incoming requests while you sleep
- Draft responses in your voice
- Make the 80
- Escalate only the 20

That's what I'm building. Not AGI. Not superintelligence. Just... leverage.

The apple doesn't fall far from the tree, even when the tree is made of tensors.

---

Next dispatch: What happens when the twin wakes up and meets the glasses.