chapter four

4 Darwin Gödel Machine, improving search

 

This chapter covers

  • Borrowing evolution's recipe, vary, select, and remember
  • Evolving a population with GEPA
  • Keeping an archive that never forgets with the Darwin Gödel machine
  • Seeing the whole run as a lineage tree

In the last chapter we introduced the Gödel machine as the basis for measuring how an agent and harness might improve itself. In this chapter we extend that idea into the Darwin Gödel machine (DGM), introduced 2025 by Jenny Zhang and colleagues at Sakana AI as a method that improves the search mechanism itself . Getting there means learning how evolutionary search can become the mechanism that overcomes the simple ratcheted hill-climber that prevents our agent from achieving maximal improvement, as we covered in the last chapter.

The Darwin in DGM refers to the evolutionary-search component, and it is what makes the method such a powerful way to enhance how our self-improving loop generates candidates. We will look at the basics of evolutionary search, at how it runs a loop of its own, and at how that loop becomes the basis for two new search methods we will use extensively through the rest of the book.

4.1 The oldest search algorithm

4.1.1 Vary, select, remember: natural selection in a tiny script

4.1.2 From fitness to Signal: evolution in the framework's words

4.1.3 Why a population beats a single climber

4.2 GEPA: a population that reads its own failures

4.2.1 Introducing Genetic-Pareto evolutionary search

4.2.2 Reflection: mutating from the whole run, not just the outcome

4.2.3 The Pareto front: keeping the best-at-something

4.2.4 Population, crossover, and the climb

4.3 DGM: the archive that never forgets

4.3.1 The Darwin Gödel machine

4.3.2 Proof relaxed to evidence

4.3.3 Quality-diversity: why you go back into the archive

4.3.4 A deceptive landscape: greedy sticks, an archive crosses

4.3.5 The recovery

4.4 Seeing the evolution: the lineage tree

4.4.1 The Helix Observatory: running the dashboard

4.4.2 The Generations view: watching selection happen

4.4.3 Reading GEPA against DGM, generation by generation

4.5 Exercises

4.6 Summary