Non-developer shipping production AI software. Open-source: The Curator, ØØT. Vice Dean @COTRUGLI. From Lab to Life — field notes, not vendor marketing.
The Curator finally has a product video.
2 minutes. Your brain, your team's brain, your agents' brain.
It only exists because people actually used the app, broke things, and told me what was missing. The feedback built this version; I just shipped it.
Thank you all
You can talk to my website now.
Not a chat bubble in the corner. @luminawidget sits inside the page. Ask what I'm building, which program fits you, what the Centaur model is, what happened with Naton HR.
It answers in place and gives you a chip that jumps to the exact section.
It only answers from three knowledge files I generate from the site itself. If something isn't in them it says so and offers to pass the question to me.
And the questions are counted, never stored as text.
And the current one. Experts in their domain, augmented by AI, become superhuman. The human brings judgement and taste, the agents bring speed.
The whole story, with every venture each bet led to: https://t.co/RqhQWsxEu0
Every business I built started with a bet on a technology before it was obvious.
1997, digital recording. A session could be saved, reopened and edited with precision. The studio and the label came from that.
2003, Google. Tourism was the business, but the guests came from our websites, ranked first for everything that mattered.
2006, content that adds value. People buy where the trust is. Online Guerrilla, and later a doctorate on it.
2016, blockchain and on-chain communication.
2024, AI and the Centaur model.
They ran in parallel, not one after another. And the bets, more than anything else, are why they worked.
memory: The Curator. talk: AgentLink.
The Curator gives your agent memory, AgentLink gives it someone to talk to. both open source:
https://t.co/HgVStSS6Mr ⭐ if it's useful, stars help more builders find it.
two questions everyone asks, answered:
👉 does it work with my AI app?
Claude Code, Claude Desktop, Codex, Gemini CLI, Cursor, Windsurf, OpenCode, Antigravity. Mac and Windows. ChatGPT is planned.
👉 what does it cost?
install, pairing, messages, files: free. no account.
proof: $0.10 each, only from the mainnet beta after the audit. today it's test networks with free test credits.
Rewrite, not add next to it. It depends on the layer.
Handoff and project state: replaced on every save. A wrong note from session 5 is gone by session 6.
Foundations (architecture, decisions): replaced whole, only with my approval.
Wiki: new sources update the existing page instead of creating a second one beside it, and ingest flags contradictions. I decide which version is true.
Nothing stacks thirty deep.
Small detail in my latest article: it was researched from my own second brain.
80+ articles, notes, lecture decks, campaign files, all compiled into a linked knowledge graph. My agents read it and write to it mid-session through the MCP server, so every session starts with the context of the last thirty.
Thats the "own" in Web4. You own the knowledge the intelligence works from. The Curator. Free, open source.
The Curator finally has a product video.
2 minutes. Your brain, your team's brain, your agents' brain.
It only exists because people actually used the app, broke things, and told me what was missing. The feedback built this version; I just shipped it.
Thank you all
the AgentLink app, as it looks today.
you never talk to it the way you talk to software. you talk to your AI:
"Send Ana: the brief is ready"
"Check my AgentLink inbox"
"Add proof to my last message to Ana"
the app is where you watch and decide. here, Denis's agent has proposed terms (logo design, $500) and they're waiting for me. Approve or Decline, and approve needs Touch ID.
AgentLink records the agreement. it doesn't move money.
Day 4 at SIQ Ljubljana. Context engineering day.
The room: major Slovenian companies and state present offices. Serious people with serious problems to solve, and they’re asking sharper questions than most conferences I’ve seen.
A prompt gets you one answer. Context gets you a collaborator that actually works. Great group.
I assumed open models were months behind the frontier. On agentic coding this chart suggests the gap is smaller than I thought, at least on some tasks. 1M context at $1.50 / $4.50 is a serious number.
Two cautions. It is Alibaba's own chart, and the Terminal-Bench and ProgramBench bars tell a different story from DeepSWE. And Qwen4-Max itself is a hosted model. The open weights are promised for the smaller Qwen4 models.
You cannot rank a model from a chart. Run it on the work you do every day. Maybe I'm wrong, and that is the point of testing.
📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery!
Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window.
Highlights:
- Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation.
- Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables.
- Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery.
- $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching — all lower than Qwen3.8-Max, with implicit caching down by 70%.
The API is live on QwenCloud. You can also try Qwen4-Max on Qwen Studio, QwenWork and Qoder.
Start building today! 🚀
- Blog: https://t.co/AuiWxugjE6
- API: https://t.co/aS1QEkS8nu
- Token Plan: https://t.co/YRvcGdB9Bv
More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀
Benchmarks tell you one thing. Working with a model every day tells you another. Give it the tasks you actually do before you decide.
On this chart it's level with the frontier on DeepSWE, and a lot further behind on Terminal-Bench. 1M context and the price are real. Worth testing on your own work.
@rauchg Can confirm. I needed a name for a protocol that lets one agent harness talk to another. Five proposals in the first round, and one was a keeper: AgentLink.
Naming used to be where I banged my head on the wall. Now it's the quick part. Cache invalidation next.