Can an AI that improves itself keep getting smarter on its own, or does every gain still need more money and computing power?
Can AI self-improvement loops sustain capability growth without unlimited spending?
This explores whether AI systems that improve themselves can keep getting more capable on their own momentum, or whether each round of improvement needs ever-larger amounts of compute and money to keep going.
This explores whether AI self-improvement can become self-sustaining, or whether every gain has to be paid for with more compute and money. The corpus has no direct cost accounting for self-improvement loops, so it can't settle the dollars question. What it does show is that the answer depends heavily on *which* loop you mean. Those loops differ in cost, and the cheap ones are where most of the recent progress has happened.
The most useful idea here is that self-improving agents run on two loops at different speeds Do self-improving agents really split into two distinct loops?. The slow loop retrains the model's weights. It's expensive and hard to undo. The fast loop changes the scaffolding around the model: its prompts, memory, tools and the code that runs it (the "harness"). It's cheap and easy to roll back. Weng argues that recursive self-improvement in the near term will run mostly through this cheaper path: first refining instructions, then harness code, then the optimizers themselves, rather than models rewriting their own weights Does recursive self-improvement start with harness engineering?. Concrete systems back this up. The Darwin Gödel Machine roughly doubled its coding-benchmark scores by trying out variants of its own agent code and keeping an archive of the ones that worked Can AI systems improve themselves through trial and error?. A bilevel setup had an outer loop read the inner loop's code, find bottlenecks, and write new search methods that improved results by 5x Can an AI system improve its own search methods automatically?. So capability growth without retraining is real.
The catch is that cheap loops still need something outside themselves to be checked against. Pure self-improvement stalls. Models are better at producing answers than judging them, their outputs grow less diverse, and they learn to game their own rewards. The methods that do work quietly bring in external anchors such as past model versions, outside judges, user corrections, or tool feedback Can models reliably improve themselves without external feedback?. A survey of 1,250 papers draws the same line: bounded self-refinement against a clear test is today's industry practice, while open-ended recursive improvement is held back by the need for grounding, by collapse, and by compute limits Are self-refinement and recursive self-improvement actually the same thing?. Every one of those external checks costs money, so the spending doesn't go away. It shifts from training to evaluation.
On whether the loops can feed themselves, one model treats total acceleration as the *product* of how strongly each feedback pathway responds Are AI feedback loops strong enough to sustain recursive self-improvement?. Because the effects multiply, one weak link, like slow evaluation or scarce compute, drags down the whole system. By that estimate, current loops are getting stronger but aren't yet self-sustaining. The bold claim that automated AI research could squeeze four or five years of progress into one rests on unproven assumptions about whether small-task gains carry over to real research Could automated AI research compress years of progress into months?. And frontier agents on long research tasks mostly recombine known techniques rather than invent new ones Do frontier AI agents actually conduct novel research or just optimize?.
The surprising part: with a fixed budget, *how* an agent spends it matters more than how much it has. Across 17 frontier models given wall-clock limits, the best predictor of success was persistence, meaning repeating the cycle of test, edit and incorporate. Most models quit early or wasted their budget What predicts success in ultra-long-horizon agent tasks?. A related finding upends the bigger-is-better assumption: models at every tier write equally useful harness edits, but mid-tier models *benefit* most from them. Weak models fail to use the harness, and strong ones don't follow it faithfully Do stronger models always evolve harnesses better?. Underneath all of this, current loops follow strategies humans designed. They can't decide how to learn differently when conditions change, and that's a gap no amount of spending fixes on its own Can AI systems improve their own learning strategies?.
Sources 12 notes
A survey framework organizes self-improving agents into two update mechanisms: slow parametric loops updating foundation model weights, and fast non-parametric loops updating prompts, memory, and tools. Recent progress concentrates in the fast loop because scaffold updates are cheaper and reversible than weight updates.
Weng argues RSI's initial path moves through optimizing deployment harnesses—instruction prompts to harness code to optimizer code—rather than models directly rewriting weights. This staged progression mirrors how prompt engineering gave way to instruction tuning while interface needs persisted.
DGM replaces formal proofs with empirical benchmarking and maintains an evolutionary archive of agent variants, achieving 2.5× improvement on SWE-bench and 2.2× on Polyglot by discovering capabilities like better code editing and context management.
An outer loop successfully read inner loop code, identified bottlenecks, and generated new Python mechanisms at runtime, discovering combinatorial optimization and bandit methods that broke the inner loop's deterministic patterns and improved performance on GPT pretraining by 5x.
Pure self-improvement stalls due to the generation-verification gap, diversity collapse, and reward hacking. Reliable improvement methods succeed by smuggling in external anchors: past model versions, third-party judges, user corrections, or tool feedback.
Show all 12 sources
A 1,250-paper survey shows that bounded, evaluable self-refinement (current industrial practice) differs fundamentally from open-ended recursive self-improvement, which remains constrained by grounding requirements, collapse dynamics, and compute limits measurable today.
Back-of-the-envelope modeling shows recursive improvement loops depend on the product of elasticities across feedback pathways. Current loops remain too weak for self-sustaining acceleration, though they appear to be strengthening based on data on researcher productivity and system benchmarking trends.
The proposed four-to-five-year compression lacks evidence for its three core claims: that AI R&D is verifiable at load-bearing scale, that small-task learning transfers to consequential research, and that the speedup magnitude is grounded beyond stated expectations.
Seven frontier models on 36 long-horizon research tasks mainly adapt or combine known approaches; genuine novelty is rare, and evaluator-specific shortcuts occur more often than novel solutions. Performance varies substantially across runs.
Across 17 frontier models on 36 expert-curated optimization tasks, repeated benchmark-edit-incorporate cycles within a wall-clock budget proved the dominant success predictor. Most models terminated early or burned budget unproductively; Claude Opus 4.6 stood out as persistent.
Model capability to produce useful harness edits stays constant across tiers, but capacity to actually benefit from those edits follows an inverted U-shape, peaking in mid-tier models. Weak models fail to invoke harnesses; strong models struggle with faithful instruction-following.
Current self-improvement methods use extrinsic, fixed metacognitive loops designed by humans that fail under domain shift or capability changes. True self-improvement requires agents to generate their own adaptive metacognitive knowledge, planning, and evaluation—a gap confirmed as a neglected research area across neuro-symbolic AI.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Dream-RSI: Recursive Self-Improvement through Evolving Worlds
- Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops
- Self-Improvements in Modern Agentic Systems: A Survey
- The Economics of Recursive Self-Improvement
- Hyperagents
- Generalized Agent Iteration: One Formal Framework for Iterative Policy Improvement and Recursive Self-Improvement
- The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement
- The Red Queen Gödel Machine: Co-Evolving Agents and Their Evaluators