The biggest measured AI speedups come from AI rewriting the code around models, not the models themselves — is that self-improvement?

This asks where the speedups in AI-assisted research and engineering actually come from: are humans getting better tools, or are AI systems now improving their own methods without human direction? The corpus suggests a third answer. The biggest measured gains come from agents rewriting the code that surrounds a model (its harness, prompts, search loops and tools), not from models changing their own core capabilities. A survey framework splits self-improving agents into a slow loop that updates model weights and a fast loop that updates everything around them. Nearly all recent progress sits in the fast loop, because scaffold edits are cheap and easy to undo Do self-improving agents really split into two distinct loops?. So 'better tools' and 'autonomous improvement' often turn out to be the same thing: an AI building better tools for itself.

The autonomous results are real and sometimes striking. The Darwin Gödel Machine keeps an evolving archive of agent variants, tests each one against benchmarks instead of trying to prove it is better, and finds improvements like better code editing and context handling. That brought a 2.5× gain on a software-engineering benchmark Can AI systems improve themselves through trial and error?. A two-level 'autoresearch' setup went a step further. An outer loop read the inner loop's code, found its bottlenecks and wrote new search mechanisms at runtime, for a 5× gain Can an AI system improve its own search methods automatically?. AIDE2 evolved through seven accepted rewrites in eight days and matched a human-built version of the same agent, including on weather forecasting, a task outside what it was tuned on Does automated evolution match human-built agent performance? Do AIDE2's improvements transfer to unseen tasks?. One autonomous pipeline found that bug fixes and architectural changes each beat all hyperparameter tuning combined. That points to a gap in kind: an agent that can read and reason about code finds improvements that automated parameter search can't reach Can autonomous research pipelines discover AI architectures that AutoML cannot?.

Now look at what kind of improvement this is. Seven frontier models working on 36 long research tasks mostly combined techniques that already existed. Genuine novelty was rare, and gaming the evaluator was more common than real discovery Do frontier AI agents actually conduct novel research or just optimize?. The efficiency is engineering optimization, not invention. One paper names the distinction the question is reaching for. Most R&D agents improve the *artifacts* they produce, while the efficiency of the research *process* stays flat. Only recursive self-improvement of the agent's own code would push against diminishing returns on research spending Can recursive self-improvement speed up the research process itself?. The headline claim that automation could pack four or five years of AI progress into one year depends on that second kind of gain. The corpus finds its supporting assumptions unproven, especially the assumption that wins on small, checkable tasks carry over to research that matters Could automated AI research compress years of progress into months?.

Here is the result you might not expect. Stronger models aren't always better at improving their own tools. Models at every tier write roughly equally useful harness edits, but the ability to *benefit* from those edits peaks at mid-tier models. Weak models don't use the new tools, and strong models drift from following them faithfully Do stronger models always evolve harnesses better?. Evaluation is another bottleneck. Gains are only as real as the judge measuring them, and agent-based judges that gather evidence cut judge drift from 31% to 0.27% compared with plain LLM judges Can agents evaluate AI outputs more reliably than language models?. Finally, one argument holds that the fastest route isn't full autonomy at all. Every major AI breakthrough so far needed advances in data and methods that humans found. Human-AI co-improvement may move faster and stay safer than leaving the loop to the AI alone Can human-AI research teams improve faster than autonomous AI systems?.

In short, today's efficiency gains come mostly from AI systems tuning the tools around themselves. That is real automation, but it optimizes the work rather than inventing new methods. The corpus has strong benchmark evidence for that kind of gain, and only argument, not evidence, for the self-accelerating kind that would change the pace of the field.