AI creating AI: can artificial intelligence create another AI?

AutoML designed a better network than humans did

In November 2017 Google published the result that made headlines. AutoML, its approach to automating the design of machine learning models, had created a new vision algorithm, NASNet, that beat the best models published until then. It was not built for faces, as the first version of this post said, but for general image classification and object detection. To test NASNet, Google used the ImageNet image classification set and the COCO object detection data set, which it described as “two of the most respected large scale academic datasets in computer vision.”

NASNet reached 82.7 percent prediction accuracy on the ImageNet validation set, 1.2% better than all previously published results. According to the paper, it did so with 9 billion fewer floating-point operations than the best human-invented design, a 28% cut in computational demand. A small version for mobile devices reached 74%, 3.1% better than state-of-the-art models of the same size, and on COCO object detection the largest model reached 43.1% mAP, 4% better than the previous best published result.

How neural architecture search works

The method behind NASNet is neural architecture search (NAS). One network, the controller, proposes designs for another network. Each proposed design is trained and tested, its accuracy goes back to the controller as a reward, and over thousands of rounds the controller learns which designs work. Barret Zoph and Quoc Le introduced the approach with reinforcement learning in 2016, and it was expensive: the original search ran on 800 GPUs for 28 days.

NASNet made it cheaper with a trick. Instead of a whole network, it searched for a small building block, a cell, on the small CIFAR-10 image set, then stacked copies of that cell into a large network for ImageNet. Reusing what was learned on one problem for a bigger one is transfer learning applied to design itself. Even so, the search used 500 GPUs for four days. In 2018 ENAS, which let candidate networks share weights instead of training each one from scratch, cut the cost about 1,000 times. In 2019 Google’s EfficientNet family started from a base network found by architecture search, and its largest model reached 84.4% top-1 accuracy on ImageNet.

What is AutoML?

AutoML, automated machine learning, is the broader idea: let software do the parts of building a model that used to need a specialist. Architecture search is only one of them. A typical AutoML tool automates:

• data preparation, such as filling gaps and encoding categories;
• feature engineering, choosing and building the inputs that matter;
• model selection, trying many algorithms on the same data;
• hyperparameter tuning, often with Bayesian optimization;
• architecture search for deep networks;
• ensembling, combining the best models, and validation on data the model has not seen.

What is AutoML used for? Mostly ordinary business problems: classification such as fraud or churn detection, regression and forecasting such as demand or prices, image recognition and text classification. Google brought it to customers on 17 January 2018, three days after the first version of this post, with Cloud AutoML Vision, which built custom image models using transfer learning and neural architecture search. Microsoft’s Azure Machine Learning, Amazon’s SageMaker Autopilot and open-source libraries such as Auto-sklearn, Auto-PyTorch and AutoKeras offer the same idea.

Is AutoML free? The open-source libraries are: you pay only for the computers they run on. Google’s managed AutoML is a paid cloud service, billed by the hour of training and of serving predictions.

Is AutoML deprecated? The idea is not, but the products change. Google retired its legacy AutoML Vision, Natural Language and Tables services in 2024 after moving them into Vertex AI, then shut down AutoML Text in June 2025 and AutoML Video in July 2025, pointing those tasks to prompting and tuning its Gemini models. Image and tabular AutoML are not on its deprecation list. In August 2026 the groups in Freiburg, Hannover and Tübingen behind Auto-sklearn and SMAC also closed automl.org, writing that the field had spread beyond any single lab.

How AI creates AI in 2026

Generative AI, systems that produce new text, images, code or sound from patterns in their training data, changed the question. OpenAI’s ChatGPT, launched in November 2022, made that kind of AI an everyday tool, and the interesting part is no longer one network designing another, but AI taking part in almost every step of building the next AI:

• Writing the code. In October 2024 Google’s CEO said that more than a quarter of all new code at Google was generated by AI, then reviewed and accepted by engineers. Coding agents now write training scripts, data pipelines and experiments.
• Improving the training itself. In May 2025 Google DeepMind presented AlphaEvolve, a coding agent built on Gemini. It found a better way to split a large matrix multiplication, which sped up a key kernel in Gemini’s architecture by 23% and cut Gemini’s training time by 1%. DeepMind says this improved the training of the very models that AlphaEvolve runs on.
• Making the training data. Microsoft’s phi-1, a small model for writing code released in June 2023, was trained partly on textbooks and exercises generated by GPT-3.5. Large models are also used to label data and to teach smaller models their skills, a method called distillation.
• Giving the feedback. Anthropic’s Constitutional AI, published in December 2022, trained an assistant to be harmless without human labels of harmful answers: a model compared answers against a short list of principles, and that AI feedback served as the reward.
• Playing against itself. AlphaGo Zero, the successor to AlphaGo, learned Go in 2017 without any human games, as its own teacher, by playing against itself.
• Doing the research. In March 2025 Sakana AI reported that a paper written end to end by its AI Scientist-v2, from hypothesis to experiments to text, passed peer review at an ICLR 2025 workshop. It had been agreed in advance that the paper would be withdrawn.

Where people still matter

None of this takes people out of the loop yet. Someone has to decide which problem is worth solving, which data may be used, what counts as success and whether a model is safe to ship. AutoML can also be expensive, can overfit to its test data and tends to produce models that are hard to explain.

Training AI on AI output has a limit too. A study published in Nature in July 2024 showed that models trained, generation after generation, on data produced by earlier models suffer model collapse: the rare cases in the original data disappear first, and the output gets narrower and stranger. The authors’ point was not that synthetic data is useless, but that it must be filtered carefully and that real human data becomes more valuable as generated text spreads online.

So how do you create an AI today? The steps are the same as before, but AI helps with most of them: define the problem and the measure of success, collect and clean the data, start from a pre-trained model where you can, train or fine-tune it, test it on data it has not seen, deploy it and monitor it after launch. AutoML and coding assistants speed up the middle steps. The first and the last remain human work.

ML and AI: threat or help for humanity?

Machine learning gives many artificial intelligence systems the ability to perform specific tasks. Although the concept behind it is quite simple, an algorithm that learns from a lot of data, the process requires a huge amount of time and effort, which is exactly what AutoML tries to reduce.

When it comes to NASNet, accurate and efficient computer vision algorithms are highly sought after because of the number of potential applications. They can be used to build advanced robots with artificial intelligence support, tools that describe the world to visually impaired people, and the vision systems of self-driving vehicles. The faster an autonomous vehicle recognizes objects in its path, the faster it can react to them, which makes such vehicles safer. Google noted that NASNet’s learned features may be reused for many computer vision applications.

The system is open source: Google released NASNet for image classification and object detection in its TensorFlow repositories and wrote: “We hope that the larger machine learning community will be able to build on these models to address multitudes of computer vision problems we have not yet imagined.”

However, the creation of artificial intelligence that can build artificial intelligence raises some concerns. For example, what prevents a parent from transmitting unwanted prejudices to its child? What if AutoML creates systems so fast that society cannot keep up? The worry is old: in 1965 the mathematician I. J. Good argued that a machine able to design better machines would set off an “intelligence explosion”.

In 2018 it was easy to imagine NASNet-style vision in automated surveillance systems before regulations could control them. One of the people who called for the regulation of artificial intelligence was Tesla’s CEO, Elon Musk, who told US governors in July 2017 that AI is “a fundamental risk to the existence of human civilization.” In March 2023 he was among the signatories of an open letter asking AI labs to pause, for at least six months, the training of systems more powerful than GPT-4. Rules have since arrived: since 2 February 2025 the EU AI Act has banned building facial recognition databases by untargeted scraping of images from the internet or CCTV, and it allows real-time remote biometric identification in public spaces for law enforcement only in narrow cases.