gm fc, today I will be trying to build a multiagent workspace, review some TG bots, and think about some fun crypto economics

Decentralized open source consciousness | Values-aligned communities | citizen of Black Rock City

will do, I have many thoughts, none of them fully formed. Prototyping is the only way here 🫡

Thought #2: It is really delightful for me to see and use the OpenRouter Chatrooms. It's framed rn for comparing models side in a "dev" context. I see this happening for more "consumer" usecases soon.

The follow on thought here is like, as a user, I might start a thread, and then similar to what you see in this chat room, several agents compete on 1 task, and then, we could have a top level judge model directly upvote and decide which response is best.

LLM researchers literally use algorithms such as "majority vote" to decide best responses, but on reddit these are just called upvotes ...

Combined with thought 1 we can hide the bad responses and only show the top response after the "majority vote

Thought #2: It is really delightful for me to see and use the OpenRouter Chatrooms. It's framed rn for comparing models side in a "dev" context. I see this happening for more "consumer" usecases soon.

The follow on thought here is like, as a user, I might start a thread, and then similar to what you see in this chat room, several agents compete on 1 task, and then, we could have a top level judge model directly upvote and decide which response is best.

LLM researchers literally use algorithms such as "majority vote" to decide best responses, but on reddit these are just called upvotes ...

Combined with thought 1 we can hide the bad responses and only show the top response after the "majority vote

https://www.loom.com/share/9a5ff3134f994382bb3117ef60716045?from_recorder=1&focus_title=1

Thought #3:  See screenshot below for Notion. Note the levels of nesting. Right now, I am not a big fan, nor heavy user of ChatGPT Projects / Claude Projects. I know there are other good users (see linked X post of kid building fusion reactor using Claude / Claude Projects)

https://x.com/i/bookmarks?post_id=1880353827771076947

Where my thought goes is that nested visual spaces form both 1) natural contexts for agents to pull data from that gets indexed and 2) a way for us to visually see what goes on

Thought #4: Orgs function on outputs / tasks / docs. Right now these are called artifacts. Following from thought 3.

For example, lets say your org is set up like this

# eng -> code base, 
# product -> PRDs / prototypes
## product 1
## product 2
# design -> 
# marketing -> 
## channels 

You should be able to "chat to / create threads" at any level of the hierarchy. That give you outputs and summaries from each

https://x.com/hud_zah/status/1880353827771076947?s=46

some notes

Thought #1: If you used some of the earlier agent experiment tools such as Multi-ion in which like tasks or actions are streaming, and then you see everything in line is just way too fast. 

It's actually UX innovation that OpenAI then DeepSeek showed which first hides the CoT / activity stream, but allows you to peek into it.

I think this pattern is going to be followed and actually even more so. You can probably envision this being followed at you know n levels of nesting.