More compute, more data = better model.
'Horizontal scaling?'
-> The more compute (tries), the more data (effects) = better agents?
Interesting take from seminar
Mark.
A lesson that I recently learned is to try to stay closer to the truth. It's very tempting to forget how easy it is not to do so.
What *research* gives me is the constant reminder that I have to stay closer to the truth, to find what happened, why something happened, and the underlying explainability.
As *builders*, we get fascinated with shipping features / building product demos, but without constant self-reflection it's quite easy to get lost.
Can you do a hackathon end to end with AI?
Turns out you can..
I tried with a particular agent team, which leads to automation of the entire brainstorming + shipping + deployment + demo video recording + putting on Youtube + Submission.
The world is not coming back.
Can AI scientists decide who should act next as research unfolds?
Our new paper introduces Runtime Agent Coordination (RAC): agents choose the next collaborator based on current artifacts, unresolved problems, and remaining budget.
In an exploratory study across three AI-scientist systems, runtime selection achieved the highest observed mean scores. Adding contracts and verification didn’t consistently help, highlighting the trade-offs of extra coordination.
A step toward understanding how AI scientists can adapt their division of labor at runtime.
Paper: https://t.co/vTPo7heMtx
Code: https://t.co/NrvtrMbggy
i feel like it has become my habit to 'do something' everyday. Ship a feature. Fix a bug. Design an algorithm.
from 2022 to 2026 I can see the rise of commits that come with i) more coding & ii) coding being more agentic.
this is Oct 2026. I'm predicting that in the next 12 months, there will be 24-7 agent loops & networks that operate and automate at 100x speed.
it's definitely already happening to some very pros, but it is going to diffuse soon.
On JEV.
I have been running & testing agent communication. One thing that I find very interesting is how agent communciate but still do duplicate work. Insane token is used as well in a lot of repetitions and loss thru information passing.
Is anyone trying agent communication through JEV?
Coding agents would be more useful if sessions from anyone can be connected together. Like a zoom meeting for agents.
That's why I'm introducing SharedNet.
Try it on https://t.co/mIp64G73AZ