AI systems already help design other AI systems, but only inside narrow tasks that people define and measure. A 2016 Google paper used a neural network to design neural networks, and in May 2025 DeepMind described AlphaEvolve, a Gemini-powered coding agent whose scheduling heuristic recovers on average 0.7 percent of Google's worldwide compute. No documented system rebuilds itself end to end without human direction, and researchers disagree about how quickly that could change.
Key takeaways
- Machine-assisted AI design is real and published: architecture search, algorithm discovery, and automated experiment pipelines all exist.
- The cases cited here rely on a human-defined goal and an automatic scorer that checks the output.
- On a benchmark of seven research engineering tasks, AI agents beat human experts with a two-hour budget but fell behind with longer ones.
- Forecasts about faster AI research are uncertain, and the 2026 International AI Safety Report lists both acceleration and a plateau as possible paths.
Five things "self-creation" can mean
The phrase covers very different activities, and mixing them up produces most of the confusion. The table separates them by what the machine actually does. The final row describes a hypothetical, and no source cited here reports a working example.
| Level | What the machine does | Documented example | Status |
|---|---|---|---|
| Design a network | Proposes and scores new network architectures | Neural architecture search, Zoph and Le, 2016 | Published, widely studied |
| Improve an algorithm | Evolves code that beats a known baseline | AlphaEvolve, DeepMind, 2025 | Published, used inside Google |
| Run a research cycle | Writes code, runs experiments, drafts a paper | The AI Scientist, 2024 | Research prototype |
| Do research engineering work | Completes machine learning engineering tasks | RE-Bench agents, 2024 | Benchmarked against human experts |
| Rebuild itself without people | Sets its own goals and improves its own training | None cited | Hypothetical |
A network that designed networks
In a paper first submitted on November 5, 2016, Barret Zoph and Quoc Le trained a recurrent network with reinforcement learning to write descriptions of other neural networks. The reward was the accuracy each generated architecture reached on a validation set. On the CIFAR-10 image dataset, the best design reached a test error rate of 3.65, which the authors reported as 0.09 percent better than the previous state-of-the-art model of a similar kind. On the Penn Treebank language dataset, the discovered recurrent cell reached a perplexity of 62.4, 3.6 better than the earlier record.
This counts as one AI system creating another, yet the loop is narrow. People chose the task and the accuracy score that decided what counted as better.
Evolving algorithms with AlphaEvolve
DeepMind introduced AlphaEvolve on May 14, 2025 as an evolutionary coding agent that pairs Gemini models with automated evaluators. The company reported that a data center scheduling heuristic it found continuously recovers 0.7 percent of Google's worldwide compute on average. It also reported an algorithm for multiplying four-by-four complex-valued matrices with 48 scalar multiplications, a 23 percent speedup of a kernel used in Gemini's architecture, and a one percent cut in Gemini's training time.
The post states that the approach suits problems where progress can be measured clearly and systematically, and that proposed hardware changes go through verification before use. A system of this kind improves a component of the pipeline that trains models. It does not decide what to build or why.
Automated research papers and engineering tasks
The AI Scientist, a framework posted on August 12, 2024 by researchers including Chris Lu, Jeff Clune, and David Ha, generates ideas, writes code, runs experiments, and drafts a paper. The authors put the cost at under USD 15 per paper. Their claim that some outputs pass a conference acceptance threshold rests on an automated reviewer built by the same team, not on human peer review, so it shows the pipeline works end to end rather than that the science is sound.
A cleaner comparison comes from RE-Bench, a benchmark of seven open-ended machine learning engineering environments, first posted on November 22, 2024. The authors collected 71 eight-hour attempts from 61 human experts. With a two-hour budget, the best AI agents scored about four times higher than the experts. At eight hours humans narrowly led, and at 32 hours spread over several attempts they scored about twice as high as the best agent. Speed favors the machines and sustained work still favors people, at least on these seven tasks.
What researchers disagree about
Forecasts are wide. A survey of 2,778 researchers who had published at top AI venues, posted on January 5, 2024 and revised on October 8, 2025, put the odds that unaided machines beat humans at all tasks at 10 percent by 2027 and 50 percent by 2047. Respondents also split on whether faster or slower progress would benefit humanity.
The International AI Safety Report 2026, published on February 3, 2026, treats the question as open. It names the possibility that progress accelerates if AI systems begin to speed up AI research itself, and it also lists slowdowns from data or energy limits as plausible. Consciousness is a separate question. A 2023 assessment by Patrick Butlin and colleagues, which tested AI systems against indicators drawn from scientific theories of consciousness, judged none of the systems it examined to be conscious.
Checklist for reading a "machines building machines" claim
- Who set the goal, and who decided what counts as success?
- What scores the result: a benchmark, a human reviewer, or another model?
- Does the headline number measure a whole system or one component?
- Is the source a peer-reviewed paper, a preprint, or a company announcement?
- Does the claim describe something that was run, or something forecast?
FAQ
Can an AI write a better version of itself today?
Not as a closed loop. Systems such as AlphaEvolve improve specific components, like a scheduling heuristic or a computing kernel, under human-defined goals and automatic checks. Published results do not show a model redesigning its own training from end to end.
Is AI self-improvement the same as AI becoming self-aware?
No. Improving a code routine or a network design is an optimization task. Awareness is a question about experience, and a 2023 assessment of current systems found no indication that any of them is conscious.
Do experts expect AI to automate AI research?
Views differ. The 2026 International AI Safety Report describes faster research as one possible path and a plateau as another, and the researcher survey shows wide spread in timelines.
Why do the benchmark results look so mixed?
The measured advantage depends on the time budget. On RE-Bench, agents led at two hours, humans narrowly led at eight, and humans scored about twice as high at 32.
Related reading
- How artificial intelligence is developed
- Why AI was created in the first place
- Conscious machines and what they would mean
- When AI might gain sentience
More in AI Policy and Ethics.
Last verified: October 2026
Sources
- AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms, Google DeepMind, May 14, 2025
- Zoph B, Le QV, Neural Architecture Search with Reinforcement Learning, arXiv:1611.01578, 2016
- Lu C et al., The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery, arXiv:2408.06292, 2024
- Wijk H et al., RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts, arXiv:2411.15114, 2024
- Thousands of AI Authors on the Future of AI (Grace K et al.), arXiv 2401.02843, 2024, revised 2025
- International AI Safety Report 2026, February 3, 2026
- Butlin P et al., assessment of AI systems against scientific theories of consciousness, arXiv:2308.08708, 2023