Azeem Azhar's Exponential View Karpathy’s autoresearch could make scientists of us all
77 snips
Apr 1, 2026 An exploration of Karpathy’s AutoResearch loop and how autonomous experimentation speeds iteration. Adapting that loop to business and thinking, including synthetic judges for scoring. A deep look at the risk of local minima and a proposed escape harness to push beyond “good enough”. Practical tips from running many short iterations and when the approach breaks down.
AI Snips
Chapters
Transcript
Episode notes
Convert Goals Into A Single Measurable Objective
- Try reframing non-ML goals as a single measurable objective so auto-loops can optimize them.
- Azeem used synthetic judges scoring headlines and theses to create a scalar objective for iteration.
Iterating A Thesis To Iteration 17
- Azeem used his AutoWolf version to iterate a thesis and ran 19 iterations, peaking at iteration 17.
- The oracle judges scored drafts (e.g., from 4.6 to 5.9) and the loop emailed iteration histories for review.
Scalar Objectives Make Iteration Possible
- Auto-loops require collapsing complex decisions into a single scalar score, which simplifies but enables optimization.
- This reduction enables many previously too-expensive experiments to be run cheaply, expanding where science can be applied.