I think in one year we will laugh about OpenClaw as it is today. And new users won’t believe we used it this way.

Kimi K 2.5 is very good. Open-source models will keep catching up with frontier models. Being a few months behind is acceptable. At this point, for my needs, I don’t see any reason to keep being subscribed to ChatGPT or Grok. Kimi via @Maple does the job well.

My internal AI agent hype has subsided. Time to go back to work.

USDT is transactional at most. Don’t store your wealth in centralized systems.

When Children of Bitcoin summit?

I guess OpenClaw is what Siri is supposed to be on your phone.

Hah. Hermes has a heartbeat instruction to learn my way of talking. He just said “My 2 sats: ….”

I got a bit distracted from Tetrapolar, my main project, on purpose. Founder fatigue is real, so I’m giving myself some time off while a security audit is getting done (yep, I’m doing that!).

There’s an OpenClaw bubble going on!

If you want to trust the agent with your whole life (make it your daily assistant), perhaps, the correct architecture is to run two OpenClaw instances: 1) trusted orchestrator is the one you talk to; it has access to your workspace data and many tools but no access to the outside world; 2) untrusted subagent is the one to whom your orchestrator delegates any tasks that require any outside world connection (web search, social media, nostr, browsing, etc.); this subagent performs tasks and reports back to the orchestrator. Unless they're working on truly private subagents within the same instance?

If you want to trust the agent with your whole life (make it your daily assistant), perhaps, the correct architecture is to run two OpenClaw instances: 1) trusted orchestrator is the one you talk to; it has access to your workspace data and many tools but no access to the outside world; 2) untrusted subagent is the one to whom your orchestrator delegates any tasks that require any outside world connection (web search, social media, nostr, browsing, etc.); this subagent performs tasks and reports back to the orchestrator. Unless they're working on truly private subagents within the same instance?

When @MartyBent boasted about his agent creating a Nostr identity and posting on his own, all without asking, I thought maybe Hermes’ architecture doesn’t allow it? But I talked to him and, apparently, it’s all about the instructions in the HEARTBEAT. Hermes was a bit too reactive and not proactive. I made him come up with a heartbeat instruction that he thinks he will follow. Let’s see if this helps!

I’m setting up OpenClaw to play with it, too. I’ve made good progress with my own custom agent “Hermes”, but got a bit mad at him because he can’t figure out how to schedule a reminder. However, one superior feature that I made was persistent isolated subagents: you can set up a permanent subagents that wakes up with its own heartbeat and has no access to the workspace data. It’s a much better approach for a social media bot or an agent that goes out there and interacts with other people or agents. It the reports to the main agent.