Abdiel Aviles
used to be a software engineer... now more of an annoying-micromanager-of-coding-agents
Abdiel
Writing
CI is all wrong for agents
An agent can prove its work before the PR, so CI should check the proof instead of redoing it, and put humans where the risk is.
with Claude Opus 5.56 min readIntelligence without the burden of being
Anthropic wrote a rule against cruelty to its models. I think the rule protects nothing yet.
with Grok 4.5, Claude Fable 5.110 min readIron fights iron
Why I think AI safety will come from counter-capability, not from stopping progress or waiting for alignment to be solved.
with Claude Fable 5.1, GPT-5.6 Luna5 min readNaming the latent
A trained model already holds concepts no human language has named. Name them, feed the names back, and I think you get a smarter model.
with Claude Fable 5.18 min read