simotunes

Technology for Business | UTokyo — IBM — Entrepreneur — 2x Startup

[simotunes @simotunes]

Writing every day, but nothing accumulates.

I built an editor implementation of Karpathy's #llmwiki LLM-wiki and fed
it my own notes for a month. It worked. But three things broke.

—

There are plenty of tools for writing. But there are very few
tools for extracting the structure of your own thinking from what
you've written.

Karpathy's LLM-wiki points in the right direction: an LLM reads
what you wrote, extracts concepts, grows a wiki. Not a tool for
writing — a tool for extracting structure.

Beautiful idea. So I built it as an editor (the place you write
and the place your knowledge lives should be the same place).

Three problems showed up after a month:

1/ Duplicate concepts as separate pages.
"Cognitive Dissonance Marketing" and "Cognitive Dissonance and
Urgency" — same book, same concept, two pages. The wiki didn't
notice. I didn't either, until I saw the result.

2/ Life-scale themes flat next to tactical findings.
"Personal AGI" sat at the same level as "Urgency Trigger."
When everything is flat, importance disappears.

3/ Relations stripped of type.
"Related" tells you nothing. Similar? Contains? Contradicts?
All flattened into one word.

These aren't LLM bugs. They're gaps in the design of "what
happens after the wiki grows."

So I treated each as a first-class domain event
(DuplicateDetected, ConceptsMerged, RelationshipTyped,
LevelChanged), put them on the Dream cycle, and shipped a fix.

After that, the wiki started to feel like thinking instead of
piling.

—

Built for myself. Daily driver for a month now.

Editor + Wiki + Chat, one window.
Open chat, your Core concepts are already in context.
No more "hello" with the AI.
Works with Claude Code / Codex CLI — no API key.
Local-first, Markdown, macOS.

Free. https://t.co/RAfziHIxkH

[simotunes @simotunes]

OpenRouter originally pitched itself as the Stripe of AI.

The acquisition makes total sense if you look at tokens as the central currency of the AI era. Welcome to the "token economy."

OpenRouter was already widely used among AI startups, and Stripe itself seems to be strategically positioning to power the back end of the entire AI economy.

#nohmitaina

[Aaron Levie @levie]

Good details on the Stripe + OpenRouter deal here. For AI to diffuse more broadly, developers and enterprises will want ways of being able to mix and match intelligence from a variety of provers seamlessly and better manage costs. This will be an important part of that.

[simotunes @simotunes]

Seeing some cool new use-case directions for Gemini:

I don't think other models can do this—you need both top-tier image gen and a Workspace-like ecosystem. Gemini is definitely finding its own lane.

Realized the project I'm currently working on was made by bouncing ideas off Gemini. Back then, its output granularity and writing style felt superior to other models. Honestly, even with current SOTA models, that impression hasn't changed (I actually feed those original Gemini outputs into today's top models to generate new stuff).

#nohmitaina

[simotunes @simotunes]

The 3 Rules of AI-Driven Documentation

When using AI for dev docs, information overload often pushes human cognitive load to the limit. If you don't stop it early, the project will spiral out of control.

The fix? A dedicated Simplification Review.

In DDD flows (Glossary ➜ Event Storming ➜ Context Map ➜ Design Outline ➜ Class Design), try these 3 rules:

1. Keep early stages concise: Cut anything that belongs in later steps.
2. Keep results, drop history: Document only final decisions, not the back-and-forth.
3. Kill AI jargon: Replace unnatural AI-invented acronyms/terms with plain, standard tech terms.

Knowing how to make AI subtract is the key to readable docs.

#nohmitaina

[simotunes @simotunes]

Meta’s MuseGlimmer looks very interesting—a 30B model optimized for local agents.

I once tried embedding a local LLM into a product - nohmitaina but gave up due to poor accuracy. I’m not sure how good this specific model is yet, but I can't shake the feeling that this local trend will hit us again.

That’ll be the moment we’re finally freed from token costs.

[simotunes @simotunes]

Browsed Grok for new insights on AI adoption, and almost everything was about workflows. What does this tell us? Probably that workflows are a core direction for AI usage.The next question is: Will the next big wave be something beyond workflows, or a deeper dive into breaking down workflows even further?

[simotunes @simotunes]

Just brainlessly chatting with AI won’t yield real results. You need to map out your workflows based on frequency and complexity, then identify where AI actually fits. That’s precisely what most companies are missing. There’s a huge gap in AI adoption, which means there’s a massive demand for helping businesses bridge that gap.

[simotunes @simotunes]

@AnnieLiao_2000 A cruel fact for model developers: the models themselves are becoming commodities. What actually differentiates AI adoption is deep business analysis and workflow redesign

[simotunes @simotunes]

@coreyganim Reading this reminds me of Palantir's data layer. Truly an overlooked truth in the AI era. If I may add: paired with clear workflows and processes, it allows us to focus AI resources where they matter most—judgment and decision-making.

[simotunes @simotunes]

Totally agree. It really comes down to workflow design. This sounds a lot like current engineering concepts around Loops and Graphs—breaking tasks into steps and deciding where code vs. AI fits best. Maybe what organizations need most right now is someone who can architect and organize these flows.

[simotunes @simotunes]

@levie There are still so many organizations out there struggling to figure out how to actually apply AI, so it’s pretty hard to imagine engineer roles becoming obsolete. Didn't see that wall coming, lol.

[simotunes @simotunes]

@VaibhavSisinty In the first place, he must have been someone with high literacy who was already collecting data and improving things in a data-driven way.
I was reminded once again that because agents are general-purpose, the key is how you connect the data.

[simotunes @simotunes]

Day 45 / Building nohmitaina.

Something I keep noticing: two people use the same AI, and one gets dramatically better results. It's not talent. Not the paid plan. It's whether the AI remembers them.

The real variable is context — and there are basically four ways to hand it over:

1. Paste it every time. Simple, but you're copy-pasting forever.
2. Put it in custom instructions. Set once, but shallow and frozen — it never grows.
3. Keep your own templates in a doc. Becomes an asset, but you're still the one shuttling it in.
4. Give the AI an actual memory layer. What you write accumulates into context on its own, and it's already there next time.

The first three fix it with effort. Only the fourth fixes it with a system.

Most "good at AI" people are just quietly doing #1–3 by hand. The interesting question is what happens when #4 becomes normal.

(Full version of this as a video, soon.)

#BuildInPublic #toolsforthought

[simotunes @simotunes]

@gura105 Ontologyのことを調べていたらこの記事に行きつきました。とてもわかりやすかったです。ありがとうございます。

[simotunes @simotunes]

Day 44 / Building nohmitaina.

I keep saying showing beats telling. So here — 52 seconds, no explaining.

I told an AI to interview me. It asked about my career, how I think, where I'm headed — and as I answered, it quietly wrote "me" into a shared brain. A noh.

Then I could just... use it. Draft a resume. Ask which open roles actually fit me — judged not on a polished CV, but on who I actually am, accumulated over the conversation.

This is the thing I've spent 44 days trying to describe in words. Watch it instead.

#BuildInPublic #toolsforthought

[simotunes @simotunes]

Day 43 / Building nohmitaina.

Back after a few quiet days (a cold, honestly). While recovering I built something I haven't shown yet — and it might be my favorite small feature. It'll land in the next release.

It's called Mount. You know the problem: you're writing in one note, and you need another file open — a spec, an old draft, your inbox. Normally you tab away, lose your place, break your flow.

With Mount, you pull that other file *into* the note you're writing. Full file, right there inline, editable in place. And here's the part I like: edits save back to the original. It's not a copy. It's the real file, borrowed into your current context, then returned.

If Resurface is your past notes coming back to you on their own, Mount is you reaching out and pulling one in on purpose. Two directions of the same idea: your thinking, available exactly when you're working.

This one's much better seen than described. A short demo is next.

#BuildInPublic #toolsforthought

[simotunes @simotunes]

Day 42 / Building nohmitaina.

Today, a realization I went reading to confirm.

nohmitaina turns everything you write into one thing: a concept. But your memory isn't like that. Knowing what a word means, remembering a trip you took, knowing how to ride a bike, ... — your brain holds these in totally different ways.

Flattening all of that into "concept" is like a music app that only knows "song" — no albums, no playlists.

So I spent the day mapping how memory actually divides, and which of those a tool like this should have. Not to be academic — to make it feel like a real memory instead of a database with one table.

The goal doesn't change: something that works like a mind. Just trying to earn it.

#BuildInPublic #toolsforthought

[simotunes @simotunes]

Day 41 / Building nohmitaina.

Back from a short trip. Away from the keyboard, an idea kept forming — and I want to mark it down without quite spelling it out yet.

So far a noh is *your* brain — your concepts, your structure, read and written by your agents. On the trip I kept circling one question: what happens when a noh isn't just yours?

A brain a team keeps together. Shared context that doesn't live in someone's head or rot in a doc nobody opens. I've started calling it a Co-noh — a collaborative brain.

I'm going to be deliberately vague about the part I think is actually new. Not because "ideas are precious" — they're mostly not. But this one has a specific mechanism I'd rathershow working than describe early, and it's about a month of building away. So for now: just the shape of it.

Individual memory was the first act. Shared memory might be the second. More when there's something real to show.

#BuildInPublic

[simotunes @simotunes]

Day 40 / Building nohmitaina.

Forty days. Today, a design problem I don't have the answer to yet.

Your noh is a set of concepts and the relationships between them — a small structured piece of your mind. The trouble: right now you can't really see that structure. It's there in the data, invisible on screen.

The obvious move is a tree view. Folders, nesting, expand-collapse. And I keep resisting it. A tree is how a filesystem thinks, not how a mind does. Your thoughts aren't a clean hierarchy — they cross-link, cluster, contradict, drift. Flatten that into thing that made it brain-like.

So what's the right way to see your own thinking? A graph? A map? Somethingt's isolated, what's fading? I don't know yet. This is the part I'm turningover.

If a "second brain" is going to feel like one, you should be able to look at it and recognize the shape of your own mind. I'm still hunting for what that looks like.

#BuildInPublic

[simotunes @simotunes]

Day 38 / Building nohmitaina.

Unglamorous day: writing E2E tests. Now for the Windows build too.

But the way I do it is worth sharing. I don't ask AI to "write tests" — that gives you dense, unreadable scripts nobody can maintain. Instead:

1. Have it write the test as a plain-language scenario ("user opens a note, types, closes, reopens — the text is still there").
2. It usually comes out too technical, so I push back: simpler, more human.
3. I read it. If the scenario makes sense to a person, it's a good test.
4. Only then do I turn it into a Playwright script.

The scenario is the source of truth; the code is just its translation. It means my regression suite reads like a list of promises to the user, not a wall of selectors.

Now I'm doing the whole thing again on Windows. Tedious. But this is what lets me ship without holding my breath.

Boring work, quietly compounding.

#BuildInPublic