AI · NOBORu · Lab
One conversation at a time,
it becomes NOBORU.
AI NOBORu is the personal AI environment that Noboru Ikuta, author of AI of One's Own, built for himself. It rents its intelligence from the cloud and keeps its memory and RAG on hardware he owns. Then it goes one step further: it learns his judgment and reasoning from everyday conversations, and uses them to grow an AI clone of NOBORU.
AI NOBORu Lab · Noboru IkutaTHE WHOLE PICTURE
AI NOBORu at a glance
All NOBORU does is talk. The intelligence is rented from the cloud, while his memory and the record of his judgment keep building up inside AI NOBORu.
Codex 14 · Claude Code 20 · Gemini 10
DeepSeek · Qwen · Mistral · PLaMo
THE CLONE
Growing a NOBORU clone, one conversation at a time
“What would NOBORU do?”The goal: an AI that can answer that question for him.
A judgment-ownership module reads through NOBORU's everyday conversations and picks out what he chose, what he turned down, and why. From these it builds up a picture of his character, the way he makes decisions, and the quality of his judgment. Everything it captures is added to the RAG, where any AI can draw on it in the very next conversation.
What it clones is NOBORU's judgment, his way of thinking, and the principles behind both.
WHAT CAN BE CLONED
Judgment and principles
Faced with the same situation, which way would NOBORU go? What does he value, and what can't he stand? The more conversations pile up, the sharper and more faithful the copy becomes.
- His criteria and priorities
- What he values, and what he won't tolerate
- How he thinks, and what he believes
WHAT RESISTS CLONING
Inspiration
A move nobody has thought of yet can't be learned from a record. Genuinely new ideas are still NOBORU's own job.
- Truly original ideas
- Wild leaps out of nowhere
- Sparks of inspiration in the moment
No extra effort required.
Every conversation today brings tomorrow's clone a little closer to NOBORU.
HOW IT GROWS
Everyday conversation is the fuel
01
Talk
NOBORU works as usual, with whichever AI suits the task: Claude, GPT, or Gemini.
02
Extract
The decisions he made along the way, and the reasons behind them, are pulled out of the conversation.
03
Accumulate
The judgment record builds up in the RAG, and the map of how NOBORU thinks keeps growing.
04
Apply
Whichever AI he talks to next answers with all of that accumulated judgment in mind.
And then he talks again.
Every turn of the loop makes the clone a little more precise.
ONE MEMORY
Talk to any AI. It's the same memory.
The interface is nothing more exotic than the Claude or ChatGPT desktop app, or Claude, ChatGPT, and Gemini in a browser. The whole idea is to keep using inexpensive, capable cloud AI exactly as he always has.
The memory and the RAG don't live inside any vendor's AI. They live in AI NOBORu. So whether NOBORU is talking to Claude, GPT, or Gemini, he gets the same memory, the same knowledge, and the same cloned judgment.
- Claude Desktop
- ChatGPT Desktop
- Claude, ChatGPT, and Gemini in the browser
44 LANES
Just ask, and 44 lanes get to work
AI NOBORu has 14 Codex lanes, 20 Claude Code lanes, and 10 Gemini lanes: 44 in all. Each lane gets its own AI model, and development moves forward on many fronts at once.
The line-up includes Claude Fable, Opus, and Sonnet 5.5; GPT-6 Astra and GPT-6 Luna; Gemini 3.8 Flash; and locally run Qwen3.6:35B-A3B, Gemma4:12B, and Gemma4:26B-A4B. A Codex or Claude Code lane can just as easily run Gemini or Qwen.
All NOBORU does is ask Opus in the Claude Desktop chat. Opus breaks the work down, hands it out to the lanes, and reports progress and results right there in the chat.
Orchestrated by Opus, from the chat
Codex14
Claude Code20
Gemini10
Running now: 23Idle: 21
Get the DB rebuild and the search speed sorted today. Once it's fixed, have a different AI review it.
23 lanes are running. 11 of them are on the DB.
| LANE | MODEL | STATUS |
|---|---|---|
| Claude Code K3DB | Opus | Reviewing the DB design and drafting fixes |
| Claude Code K12DB | Sonnet 5.5 | Implementing the DB migration, tests running |
| Claude Code K15DB | Qwen3.6:35B-A3B | Rebuilding the indexes |
| Claude Code K7 | Fable | Final review before the milestone |
| Codex C11 | GPT-6 Astra | Speeding up search |
| Codex C13DB | GPT-6 Luna | Hunting for bugs in the diff |
| Codex C4 | Gemini 3.8 Flash | Reading through the related docs |
| Gemini G05DB | Gemma4:26B-A4B | Checking test data locally |
BLIND SPOTS
Planting flags on AI's blind spots
Here, an “AI blind spot” means something NOBORU or AI NOBORu knows that none of the world's leading AIs do.
AI NOBORu is connected to eight AIs, built by different makers in different countries.
- ClaudeUnited States
- GPTUnited States
- GeminiUnited States
- GrokUnited States
- DeepSeekChina
- QwenChina
- MistralFrance
- PLaMoJapan
- 01Take something NOBORU knowsStart with a fact that NOBORU or AI NOBORu knows.
- 02Ask all eight AIsHave each of the eight AIs look into the same fact.
- 03If none of them know it, plant a flagFile it as an AI blind spot in AI NOBORu's blind-spot store.
If it is NOBORU or AI NOBORu that turns out to be wrong, nothing is flagged and nothing is stored. Only facts that all eight AIs have looked into and genuinely didn't know become AI blind spots.
AI gets smarter with every new version, so yesterday's blind spot may no longer be one today. That's why the collected blind spots are regularly rechecked against all eight AIs, and the blind-spot store is audited and pruned.
PRIVATE RAG
All the power of cloud AI, with none of the data leaving
Until now, private AI has come in two flavors.
- Cloud-based: you get powerful cloud AI, but accept that your information goes out the door.
- Fully local: your information stays safe, but you make do with open models running on your own machine.
Brains or privacy: one had to go. AI NOBORu refuses that trade-off.
The parts that go into the RAG, such as research findings, results from academic papers, pieces of finished work, and summaries of all of these, are produced by high-performance cloud AI like Claude and GPT-6 Astra. The cloud only ever sees material that's safe to share.
Think of how a smartphone is made. Circuit boards, microphones, memory, CPUs, even the screws come from specialist suppliers, and the phone maker puts them together in its own factory. The suppliers never touch the finished phone, or the data of the person who uses it.
The finished parts move into the private RAG in one direction only. From there, assembling those parts and using them day to day are handled solely by open models such as Qwen and Gemma, running on NOBORU's own computers and NAS, and only through an interface kept separate from any cloud AI. Cloud AI has no way to read the private RAG at all.
The private RAG isn't protected by politely asking an AI, or by some protocol. It's sealed off by where the RAG data lives and the only paths it's allowed to travel.
DEEPER EVERY DIP
Deeper with every dip
Cloth dyed with indigo grows a shade deeper each time it goes back into the vat. AI NOBORu works the same way. Day to day, NOBORU simply uses inexpensive, high-performance cloud AI. Out of those conversations, judgments accumulate, blind spots get flagged, and the memory grows. Whether he's using Claude, GPT, or Gemini, every result feeds AI NOBORu.
- 甕覗kamenozoki
- 浅葱asagi
- 縹hanada
- 藍ai
- 褐kachi
Five traditional names for indigo, from the faintest “peek into the vat” to near-black.