Matt Duff

@DeRonin_ If you’re building @karpathy’s personal wiki right now, here’s a BIG unlock: Tell your LLM (Claude / Codex) to interview you & log your internal state every time you make a wiki entry. E.g Your priorities, how you see key relationships, what matters to you right now.

🚨 Karpathy just dropped a blueprint on how to clone your brain in 2026 most people read this and think "cool scifi, see you in 2040" reality check.. every tool Karpathy describes is already online, you just have to connect the dots By the end, you'll know how to: - capture your mind into structured data - fine-tune an LLM that thinks like you - give it your face, voice, and personality - ship it as an API anyone can talk to So, let's discuss your roadmap step by step Step 1: Data Dump the biggest mistake most people make is trying to "write" their personality into a prompt that's not cloning. that's cosplay instead you should: - record 20-40 hours of you talking (solo monologues + interviews) - dump every tweet, dm, email, voice note, blog post you ever wrote - journal for 2 weeks on HOW you make decisions, not just what you decide - record yourself reacting to random content in real time the goal isn't quantity.. it's capturing the reasoning patterns behind your words —————— Step 2: Structure the Brain raw data is useless. you need to turn it into training pairs instead you should: - transcribe everything with Whisper - use Claude or GPT to extract Q/A pairs from your interviews - label each response with context: mood, topic, audience - separate "public voice" from "private voice" datasets this is the part 99% of people will skip.. and this is exactly why their clone will sound generic —————— Step 3: Fine-tune Your LLM you don't need to train a model from scratch. you just need to bend an existing one toward you instead you should: - start with Llama 3.3 or Qwen 2.5 as the base - run LoRA fine-tune on your Q/A dataset (Unsloth makes this free on Colab) - test it against a held-out set of your real responses - iterate until it hits 80%+ similarity on style and reasoning if you can't afford the compute.. use OpenAI's fine-tuning API on gpt-4o-mini for under $50 —————— Step 4: Give It a Face and Voice text-only clone is mid. the real unlock is multimodal instead you should: - clone your voice with ElevenLabs (3 min of audio is enough) - build your avatar with HeyGen or Synthesia (30 min of video) - connect the fine-tuned LLM output to voice → avatar pipeline - add a lip-sync layer so it actually feels like you —————— Step 5: Wrap It in an Agent a clone that just replies to prompts is a toy. a clone with memory and tools is a product instead you should: - give it a vector DB of everything you've ever said - add RAG so it can pull your real opinions on any topic - plug in tools: email, calendar, twitter, stripe - deploy behind an API endpoint your clients/audience can actually talk to —————— Step 6: Ship It As a Startup Karpathy literally gave you the pitch deck in one tweet instead you should: - niche down: don't clone everyone.. clone coaches, lawyers, therapists, creators - charge $5-20k per upload, recurring for hosting - offer tiers: text-only, voice, video, full agent - first 10 clients will literally be people you already know the market is creators who want to scale themselves and experts who want to outlive their own attention this isn't a 2030 bet anymore.. every piece of this stack works today @karpathy is right.. the lossy version of brain upload is shipping this year the only question is whether you're the one building it.. or the one being cloned by someone else save this so you don't lose it gl

BONUS: Perhaps most exciting! Now that your state is logged over time, against fixed events, you’ve built a map of how your perception evolves. You can ask Claude to look back & see how you changed in relationship to what happened around you, identify patterns & apply frameworks

The trick is to keep these explicitly separate. Events go in as immutable records, and your internal state goes in as a rolling log. THEN, when the LLM queries your wiki, it first retrieves and applies your most recent internal state as its lense. Now the LLM sees what you see.

If you’re building @karpathy’s personal wiki right now, do this to avoid MASSIVE headaches: Tell your Claude to interview you & log your internal state every time you make a wiki entry. E.g Your priorities, how you see key relationships, what matters to you right now. Why?

The TL;DR - 1. As fragmentation increases, our ability to map behaviour to module relationships grows exponentially. 2. This high-resolution combined with the cognition protocol, enables true granular control. We can precisely regulate which cognitive pathways the AI can use.

The mechanism - 1. Managed Fragmentation reveals which modules drive what behaviours 2. Cognition protocols let us gate which regions can "talk" 3. More fragments = higher resolution behavioural mapping 4. The ability to shut off the "overconfident risk assessment" cluster

Key insights: 1. As fragmentation increases, our ability to map behaviour to module relationships grows exponentially. 2. This high-resolution combined with the cognition protocol, enables true granular control. We can precisely regulate which cognitive pathways the AI can use.

The mechanism: 1. "Managed Fragmentation" reveals which modules drive what behaviours 2. "Cognition protocols" let us gate which regions can "talk" 3. More fragments = higher resolution behavioural mapping 4. We can literally shut off the "overconfident risk assessment" cluster.

@gav_clayton Keen to see a video covering what comes together with this system!

This is Generic Eric, he's a rigged fuse model I've chopped up into a bunch of separate pieces all with their own nice little UV islands. He'll be my prototype character as I work through different parts of the avatar problem space.

@VictorTaelin Give us some context 🙌 - that’s a big number depending on the def of rewrite…

lol