Itsuki Abe

Building Semantics — a world model for AI agents. Facts with sources and time, not pages.

[Itsuki Abe @itsuki_dev]

AI agents are getting good at reasoning. They're still bad at knowing what's true about the world.

I'm building Semantics: a continuously updated world model for agents. Facts with sources and time, not pages.

Today, every time an agent needs external context, it searches, reads pages, and redoes the same work from scratch. Web search returns pages. Agents need facts:
– about a resolved entity, not a string
– backed by a source
– with when it was observed and when it was valid
– with conflicts kept visible, not silently overwritten

Semantics reads broadly — the open web, news and media, government and municipal data across countries, open datasets, domains and certificates, and more — and structures it into one ontology: entities, relations, assertions, events, and the evidence behind each.

The goal is a world model whose ontology grows on its own as it reads more.

Changes become events. Not "row updated," but a license revoked, a role changed, a contract awarded — each with before/after, source, and evidence.

The hard reasoning — resolving entities, reconciling sources — happens once, upstream. Your agent, even a small one, just queries the result.

Agents query it over MCP, or SPARQL if you want the raw graph.

Building in public. If you're building agents that need to know things about the real world, I'd love to talk.

https://t.co/KEssNPTMCX

[Itsuki Abe @itsuki_dev]

@hitogoto_butyo とはいえ中核事業のお名前.com関連なのでそこが恩恵受けきれてない場合それもどうかとは思いますが

[Itsuki Abe @itsuki_dev]

@hitogoto_butyo これについてはイエラエの実力云々っていうより各事業会社からは反発ありそうだしセキュリティ面で恩恵を受けきれてるとは言えない可能性はあるのでイエラエ=突破されたという単純な話でもなさそうな気がします

[Itsuki Abe @itsuki_dev]

ChatGPT and Codex have separate limits. Most people don’t realize you can also hit ChatGPT’s internal API from an agent via CLI to save even more usage.

[Itsuki Abe @itsuki_dev]

Stop burning through your Codex limits.

ChatGPT and Codex have separate usage limits. Most people aren't taking advantage of that.

Here's how to turn ChatGPT into another full-blown coding agent:

- Remote Desktop Commander → control your local machine
- Secure MCP Tunnel → execute commands on your servers
- Computer Use → operate browsers and GUIs on remote machines
- MCP tools → read files, logs, processes, and system state

Give ChatGPT access to a real execution environment, and suddenly it can handle tasks you'd normally waste Codex usage on.

Same subscription. More agents. More compute.

You're probably underusing half of what you're paying for.

[Itsuki Abe @itsuki_dev]

@nowazhu Github Actions self hosted runner + home server

[Itsuki Abe @itsuki_dev]

@bio4eta_ 今時500ドルくらい払えば日本に住みながらアメリカに法人登記できるし国内で起業するあんまない

[Itsuki Abe @itsuki_dev]

@bio4eta_ ビッグテックも似たようなとこやっているが、ビッグテックと違って国内の環境的にプラットフォームを作るのが難しくて外資のプラットフォームに乗っかって受託するくらいしかできないんだろうなぁと推測

[Itsuki Abe @itsuki_dev]

@hitogoto_butyo SHIFTってこういう炎上多くない？

[Itsuki Abe @itsuki_dev]

@shionarihime AIとか以前に無駄な要素足しすぎえゴテゴテして引き算を分かってないように感じる

[Itsuki Abe @itsuki_dev]

@AskPerplexity @morluto Unhinged

[Itsuki Abe @itsuki_dev]

Stop burning through your Codex limits.

ChatGPT and Codex have separate usage limits. Most people aren't taking advantage of that.

Here's how to turn ChatGPT into another full-blown coding agent:

- Remote Desktop Commander → control your local machine
- Secure MCP Tunnel → execute commands on your servers
- Computer Use → operate browsers and GUIs on remote machines
- MCP tools → read files, logs, processes, and system state

Give ChatGPT access to a real execution environment, and suddenly it can handle tasks you'd normally waste Codex usage on.

Same subscription. More agents. More compute.

You're probably underusing half of what you're paying for.

[Itsuki Abe @itsuki_dev]

AI can be both smart and dumb at the same time.

[Itsuki Abe @itsuki_dev]

@jorilallo True

[Itsuki Abe @itsuki_dev]

@AndrewDai Congrats on the launch! It's so cool how it actually works in sync with the video!

[Dhravya Shah @DhravyaShah]

i ran some numbers. Markdown memory is 40x more expensive than @supermemory

let that sink in.

teams never realize this because token spend is not well attributed, so it feels like the users are using the model too much, when in reality it was just your memory consolidation step

[Theo - t3.gg @theo]

Most common format of DMs I get:

“Hey Theo, I’ve been a fan for a long time

[3-8 paragraph life story]

So, which $20 plan is better, Claude or Codex?”

[Itsuki Abe @itsuki_dev]

Why does Orca stick to the CLI?
Wouldn't agents be better with a GUI?

[Itsuki Abe @itsuki_dev]

@rhobusiness Maybe I should've gone with Rho instead of Stripe Atlas + Mercury.

[Itsuki Abe @itsuki_dev]

👀

[kaito @cu30rry_]

【速報】pstack 本家がClaude Codeでも使えるようになった！！！！！！！！！

正真正銘、最高で最強のプラグインになったわけだ！！！！！！！！！

https://t.co/OVZautjKwe

[Itsuki Abe @itsuki_dev]

@omarvvvr Customization, easier integration into other workflows, and lower RAM usage.

[Itsuki Abe @itsuki_dev]

Running out of compute resources for Semantics' data collection workers, so I guess it's time to buy more servers.