Stage 5 — The Researcher (2026)

The assistance era (Copilot, then SWE-agents) kept a human choosing the

experiments. The discovery era did not.

• AlphaEvolve (DeepMind) — evolved genuinely new algorithms (2025).
• The AI Scientist (Sakana) — sketched the full paper-writing pipeline,

later published in Nature.

• autoresearch (Karpathy, March 2026) — the big moment: a deliberately

minimal ratchet loop where a coding agent modifies a real LLM training

setup, runs a five-minute experiment, keeps the change only if validation

loss improves, and repeats overnight. His extended run stacked 700

experiments into 20 kept improvements, cutting time-to-GPT-2 from 2.02 to

1.80 hours — real, transferable code changes found while he slept.

The pattern replicated within weeks, turning one GitHub repo into a genre:

• Red Hat OpenShift AI — ran the loop unsupervised for 24 hours: 198

experiments, a 2.3% validation-loss improvement, zero human intervention.

• Hyperspace — distributed the loop peer-to-peer: 35 autonomous agents

ran 333 experiments in a single night (March 8–9, 2026).

• Shopify — 37 overnight experiments for a 19% gain on an internal

workload.

• Research-lifecycle benchmarks ("Act As a Real Researcher") — by

mid-2026 the loop itself had become something you score agents on.

Experiment selection — the last intellectual step humans kept in the loop —

became an optimization target, then a benchmark category.