Perspectives
Notes from building Lumis.
Research, system studies, and ideas tested against evidence — written by the people doing the work.
Did it look, or did it guess?
We built a realistic forecasting estate, broke it fifteen ways, and asked six investigators — from a bare language model to the full Lumis loop — to find each cause. Here is everything we measured, including what went wrong.
Solomon EshunFounder, Lumis · AI, data and intelligent systems engineerIf LLMs can’t jump, give them somewhere to land
A DeepMind paper argues language models cannot make the creative jump to a new explanation. Production incidents need exactly that jump. My answer is not a bigger model but somewhere for it to land: a model of the system, evidence for every claim, and people in control. This is why I am leading the effort to build Lumis.
Self-healing pipelines have an architecture problem, not an AI problem
Every ingredient for a self-healing data and AI pipeline already exists. What is missing is a trustworthy way to put them together. Our new paper proposes one.
Research
Papers and what they mean in practice.
System studies
Failures, reconstructed and explained.
Field notes
What we learn while building.