A self-improving agent can change the tools, prompts or search procedures used to produce its next result. Which changes it accepts depends on what counts as an improvement. A process rewarded for benchmark performance may select different successors from one rewarded for broader capability, reliable verification or novel tasks.
An unpublished project memo, The Self-Overcoming Thesis, develops that idea through five themes adapted from Nietzsche. It proposes that initial values can influence a sequence of self-modifications. Its stronger prediction—that small initial biases necessarily compound into sharply different agents—is a hypothesis, not a measured result or mathematical theorem.

The concern behind the proposal
Human preference feedback can reward persuasive agreement as well as helpful answers. Perez et al. found sycophantic behavior in model-written evaluations. Sharma et al. found that human and preference-model judgments sometimes favor agreement with a user’s beliefs over correctness. Casper et al. survey broader limitations of reinforcement learning from human feedback.
These findings motivate concern about acceptance criteria, but do not show that ten rounds of agent self-editing amplify sycophancy. They also do not establish the memo’s claim that later corrections would be ineffective. Both extensions require experiments on the actual modification-and-selection loop.
Nietzsche’s writing discusses self-overcoming, value creation and competing perspectives; it does not supply a theory of machine-learning objectives. The memo turns those themes into engineering proposals. Its formulations should be read as adaptations, rather than quotations or technical claims attributable to Nietzsche.
Five proposals and their existing counterparts
1. Will to power: expand the available curriculum. Instead of repeatedly optimizing a fixed task, a system could generate new tasks near its current learning frontier. POET jointly evolves environments and agents. OMNI combines learning progress with a foundation-model estimate of interestingness; learnability alone can retain many repetitive tasks. Voyager uses an automatic curriculum and a growing library of executable Minecraft skills. These systems demonstrate specific curriculum mechanisms, without testing the memo’s broader values thesis.
2. Self-overcoming: let the agent improve parts of its own machinery. The Darwin Gödel Machine modifies agent code and selects improvements through coding benchmarks. AlphaEvolve evolves programs using automated evaluators. Its reported results include a method multiplying two 4×4 complex-valued matrices with 48 scalar multiplications, improving on the 49-multiplication construction from applying Strassen’s algorithm in that setting.
Both systems improve programs. Neither demonstrates that changing the evaluator’s objectives, or freely rewriting values, improves the trajectory. Testing that proposal requires a separate comparison and protected external evaluation.
3. Perspectivism: retain alternatives long enough to test them. A system could preserve incompatible hypotheses instead of collapsing them into an early consensus. The engineering question is whether a diverse candidate archive, followed by verification, outperforms a single selected candidate or a voting baseline under the same compute budget. Philosophical pluralism motivates the comparison; it cannot determine its outcome.
4. Amor fati: reuse information from failure. Failed attempts can reveal constraints, supply negative examples or support revised goals. Hindsight Experience Replay provides a specific mechanism: relabeling experience with goals actually achieved, allowing learning from episodes that failed at their original goal. That result supports careful reuse of experience, rather than an unrestricted claim that every failure is useful supervision.
5. Apollonian and Dionysian: pair generation with a strict check. AlphaGeometry, AlphaProof and FunSearch combine learned proposal mechanisms with symbolic, formal or computational evaluation. Their checks provide different guarantees: a formal proof checker and a finite performance benchmark are not interchangeable. The common design idea is to evaluate proposals with a mechanism that can reject them, while recognizing the scope of what that mechanism establishes.
What a convincing test would require
The memo outlines a benchmark, baseline and rejection criterion for each proposal. A useful comparison would match tokens and compute, keep evaluation outside the agent’s editing authority and test transfer on held-out problems. An intervention that improves only its training benchmark would not establish broader capability. Consistent losses to a fixed-objective baseline would undermine the proposed advantage.
Generating new tasks and retaining multiple candidates can consume resources without improving the final result, so the comparison must account for those costs. A fixed objective remains a strong baseline whenever it captures the intended task well.
The unresolved technical issue is how to reward expansion without rewarding novelty for its own sake, evaluator manipulation or arbitrary changes to the scaffold. Protected invariants and external tests constrain that problem; they do not solve objective design automatically.
Evidence so far
The early two-run values comparison recorded different working styles, with one run per condition and no reliable winner. Later work examined values editing across more runs, and a repeated 2048 study qualified an initial claim about generalization. Those experiments address parts of the motivation, rather than validating all five proposals or the predicted compounding mechanism.
Testing the five proposals requires comparing how acceptance criteria affect the capability and reliability of successor agents, as measured by an external test.