Generative AI Fundamentals Map
Map/Loops/reinforcing

Autoresearch

reinforcing loop · 5 steps · Trace on the map →

The cycle

Coding agents writeresearch codeAgents automate AIresearchAlgorithms get moreefficientStronger modelsBetter coding agentsR
  1. 01Coding agents write research codeCoding agents as products
  2. 02Agents automate AI researchSuperhuman AI researcher
  3. 03Algorithms get more efficientAlgorithmic efficiency
  4. 04Stronger modelsReasoning models
  5. 05Better coding agentsAI writes most new code

What drives it

The software lane feeds straight back into the algorithms lane: AI writing the code for AI research is the shortest feedback path on the map.

What breaks it

Forethought puts returns to software research at roughly 0.5 to 2 once hardware is fixed; below 1 the loop fizzles. Compute for experiments is the brake.

What to watch

Share of frontier-lab research code written by agents. Epoch’s algorithmic efficiency halving time (about eight months).

Also feeds it

Sources

  1. 01Forethought: Will AI R&D automation cause a software intelligence explosion?
  2. 02Semafor: Google CEO says 75% of new code is AI-generated
  3. 03Epoch AI: Algorithmic progress in language models