My Lab Diary
A personal record of my homelab journey — successes, failures, and everything in between.
Dev Toolbox
Research
My take on papers I read once or twice and played around with. No scientific methodology, no systematic comparison. Just one person's reaction after sitting with it for a bit. Feel free to disagree.
Autogenesis: the self-evolution paper that is actually about the substrate
A self-evolving agent paper where the real contribution is the substrate underneath. Versioned resources for prompts, tools, agents, memory.
April 30, 2026
Diversity Collapse: when more agents and tighter coordination produce fewer ideas
An NUS-led team measures why multi-agent ideation collapses to similar ideas across three layers: model, cognition, system. The fix is wiring, not capacity.
April 29, 2026
Attention Residuals: the paper I mostly did not understand but completely believe
The Kimi team replaces fixed residual connections with attention over preceding layers. I understood maybe 60% of it. The 60% that landed is some of the most compelling infrastructure work I have read.
April 01, 2026
HyperAgents: self-improving agents, or evolutionary prompt engineering?
A Meta/UBC paper claims agents can improve not just their task performance but their own improvement mechanism. The framing is exciting. I am not sure the mechanism lives up to it.
March 31, 2026
SkillNet: npm for agent skills? Reading a paper that scratches a real itch
A paper out of Zhejiang University treats AI skills like composable packages with dependency graphs. Not a rigorous critique, just my reaction after reading it a couple times and playing around with the ideas.
March 23, 2026
About Me© 2026 Victor Coimbra. All rights reserved.