Two Paths to AI Improving AI: From the Gödel Machine to HarnessX

In 1965, statistician I. J. Good wrote: a machine that can design even smarter machines—once the first one is built, human invention is finished. This was the earliest version of the "intelligence explosion," more than half a century before neural networks became practical.

What turned it from conjecture into an engineering problem is a wave of systems from the past two years. Some let models keep training themselves at test time; others let a shell of surrounding code repeatedly rewrite itself. None of them are smart yet, but all of them are actually running and actually gaining points. This article lays these systems on one map: where each path has gotten to, and where the potholes are.

First, how the map is drawn. There are only two places to intervene when a system improves itself: the model itself (the weights), or the ring of code around it (tools, memory, workflows—jargon: the harness). A 2025 survey calls these the θ axis and the Σ axis. The names sound intimidating; the idea is simple: either keep training the model, or upgrade its equipment.

The weights path: judging yourself

TTRL (Test-Time Reinforcement Learning, April 2025) does something that looks impossible at first glance: reinforcement learning on completely unlabeled data. Where does the judge come from? The model's own majority vote. Sample the same question a hundred times; the majority answer counts as correct. This "majority consensus as reward" trick lifted Qwen-2.5-Math-7B's pass@1 on AIME 2024 by 211%—using only unlabeled test questions.

One detail deserves a closer look. The paper reports that TTRL stably surpasses the model's own maj@n ceiling—the upper bound achievable by majority-voting a hundred times, exceeded by a model that answers only once. The mechanism isn't mysterious: RL amplifies outputs whose answers land near the majority, effectively distilling the luck of thinking a question over multiple times back into a single attempt. But it also plants a self-confirming loop: the reward is the model's own consensus; as the model grows more confident, consensus gets easier to reach and the reward easier to satisfy. Judge and contestant are the same person.

Agent0 (November 2025, UNC Chapel Hill and Salesforce) turns self-generated tasks into a two-actor drama. Two agents start from the same base model: one is the question-setter, tasked with producing problems "slightly above the solver's current level"; the other is the solver, tackling them with tools. As the solver strengthens, the setter is forced to escalate—each lifting the other. With zero external data throughout: +18% on mathematical reasoning, +24% on general reasoning. An arms race between asking and solving replaced the human annotator.

The equipment path: from Gödel to Darwin

The other path has an older ancestor. Schmidhuber's 2003 Gödel Machine is a program that can rewrite any of its own code, but under a harsh constraint: it may only act when it can *prove* the rewrite is strictly better within a formal system. Elegant, rigorous, and nearly unable to run—the cost of proof is too high.

The Darwin Gödel Machine (DGM, Sakana AI, May 2025) answers it in its name: replace "prove" with "measure." It rewrites its own code, runs benchmarks, keeps the variants that improve, discards those that don't. SWE-bench rose from 20% to 50%—two years earlier, this line of work was still just an idea.

After DGM, the line accelerates, with a clear direction: going deeper into the act of "modifying the equipment" itself.

ADAS (August 2024) lets agents write agents in code, archiving discovered designs for later retrieval. Hyperagents (March 2026) notices a hidden restriction in DGM: it improves coding agents, evaluates improvements with coding tasks—what it modifies and what it measures are in the same domain. This self-reference happens to close within coding, and falls apart outside it. Hyperagents' move is to make the "meta-level modification program" itself an editable object: the agent modifies both the task code and the code that modifies the task code.

Meta-Harness (Stanford et al., March 2026) tackles how feedback is delivered. Prior automated optimizers compressed historical experience into scalar scores or a few summary lines—compressed so hard that all detail was lost. It inverts this: the proposer browses the full source, scores, and execution traces of every historical candidate through a filesystem. On online text classification it beat the previous strongest context-management system. Treat history as files to browse, not summaries to read.

AHE (Fudan et al., April 2026) is the most thorough engineering effort on this line so far. Three pillars: component observability—the harness is split into seven file-level components, every change goes into git, inspectable and revertible; experience observability—tens of millions of tokens of raw traces are distilled into hierarchical reports, reading summaries by default with drill-down to raw records anytime; decision observability—every change must carry a stated "expected impact," automatically verified against the next round's task results, with failures exposed. Terminal-Bench 2 pass@1 rose from 69.7% to 77.0%.

HarnessX (June 2026) turns the parts into algebra. Harness components become typed, composable primitives, assembled with a substitution algebra, with an engine called AEGIS driving evolution from execution traces. Average +14.5% across five benchmarks, up to +44%—and the largest gains land where baselines are worst. It also closes the loop: traces feed back not only into the harness but into model training signal as well—here the equipment path and the weights path converge.

Scores went up—so what?

Every system on this map reports gains. But in 2026, two groups of researchers started picking apart the gains themselves.

A review paper from AI2 and the University of Washington points out a structural flaw in harness-evolution evaluation: the evolutionary process uses test cases as search feedback, then reports results on the same public benchmark, while compute-matched naive baselines (simply sampling more times on the task) are often not seriously ruled out. How much of the gain is "better equipment," and how much is just "grinding the same exam paper repeatedly"? This protocol can't tell them apart.

Another benchmark, called HarnessDev, lists four forms of harness overfitting: overfitting to the model that wrote it, memorizing dev examples, gaining one capability while quietly losing another, and improving on the feedback set while failing on a different task.

Interestingly, AHE ran into the same class of problems in its own analysis, and put it more bluntly. Two limitations: first, components are non-additive—individually effective changes stacked together hit a ceiling on total benefit; the harness is not LEGO. Second, self-attribution is reliable for fixes and blind to regressions—the system is very good at explaining what it fixed, and nearly blind to what it broke. The predictions attached to each change are accurate for "it will get better" and fail for "it will get worse."

This second point deserves a pause. Self-attribution's blindness to regression is not an AHE bug; it is a property of all self-referential systems: when judge and contestant share an origin, the contestant's progress is visible, but the contestant's fouls are not. AHE's response is to let predictions be falsified by the next round's real task results—turning "I said it would get better" into an assertion that can be refuted. The Gödel Machine tried to solve this with formal proof, at the cost of being unable to run; now people solve it with held-out replay, at the cost of regressions slipping through. The tension between rigor and runnability has not changed since 2003.

Back to the explosion

Good's condition, translated into today's language: can improvement itself be improved? Current evidence says "somewhat, and very early." TTRL surpassed its own majority-vote ceiling; Agent0's question-setter escalates automatically; HarnessX feeds traces back into training signals; and last month a team had a model search architectures, write kernels, and build an entire model. The improvement of improvement exists in embryo.

But the evaluation papers from the same period say something else: a substantial share of the "progress" runs inside the evaluation's blind spots. The rate of improvement is hard to fake, but the direction of improvement can be chosen to pick low-hanging fruit—where scores rise easily is not necessarily where intelligence lives. The most fragile link in the explosion thesis may not be the machine's speed of improvement but the human side: setting questions, grading, distinguishing real progress from overfitting—these jobs are still done by humans. Machines improve themselves on timescales of days; humans change evaluation habits on timescales of years.

The DGM took 80 iterations to lift SWE-bench from 20% to 50%. AHE went from 69.7% to 77.0% in ten rounds. Both numbers are real. The question separating them—scores went up, but did intelligence follow?—still has no judge.

---

*Sources: TTRL (arXiv:2504.16084), Agent0 (arXiv:2511.16043), STOP (arXiv:2310.02304), ADAS (arXiv:2408.08435), Darwin Gödel Machine (arXiv:2505.22954), Hyperagents (arXiv:2603.19461), Meta-Harness (arXiv:2603.28052), AHE (arXiv:2604.25850), HarnessX (arXiv:2606.14249), Rethinking the Evaluation of Harness Evolution (arXiv:2607.12227), HarnessDev (arXiv:2609.01437); I. J. Good (1965), Schmidhuber Gödel Machine (arXiv:cs/0309048). All system figures are taken from the corresponding papers' abstracts.*