The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Google’s AutoML-Zero research showed that an evolutionary search could produce a learning algorithm that beat hand-designed models of comparable complexity in a tightly constrained image-classification experiment. It did not show that AI broadly outperforms human-designed models: the work was preliminary, computationally demanding, and did not produce a fundamentally new general-purpose algorithm.
What “AI built another AI” means here
The project was called AutoML-Zero. Rather than choosing among neural-network designs supplied by people, it searched for complete learning algorithms. Candidates began as empty programs and were assembled from basic mathematical operations. The system then tested those programs on small image-classification problems.
As an Amazon Associate I earn from qualifying purchases.
Google Research described the process as evolutionary: it initialized a population of programs, created mutated copies, evaluated their accuracy, and used stronger candidates as parents for later generations. In this sense, the “AI” was an automated search procedure that generated and assessed candidate algorithms—not a self-aware system independently inventing an AI product.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →What the search found
AutoML-Zero recovered familiar ideas rather than demonstrating a wholly unprecedented learning method. The experiments rediscovered linear regression and two-layer neural networks trained with backpropagation. The report also described techniques including stochastic gradient descent and data augmentation by injecting noise.
#1 Best Overall
That result matters because the search began with basic operations rather than a menu of sophisticated, human-authored algorithm components. It suggests that useful learning structures can emerge from search under a defined task and setup. It does not establish that the system can invent effective algorithms for arbitrary problems.
What “outperforms man-made models” means
The claim is bounded: Google’s account says an evolved algorithm outperformed hand-designed models of comparable complexity in the team’s toy scenario. “Comparable complexity” is important; this was not a claim that AutoML-Zero beat every human-designed model, or that it surpassed state-of-the-art AI across applications.
Rank #2
The post characterized accurate algorithms as potentially rare in the sparse search space—approximately one in 1012 candidates. That figure describes the team’s characterization of this search space, not a general success rate for automated machine learning. Google also reported that evolutionary search was tens of thousands of times faster than random search in its measurements; that comparison belongs to this experiment, not to all search tasks.
The authors explicitly called the work preliminary, noted that the search required significant compute, and said they had not yet evolved fundamentally new algorithms. Staff Software Engineer Esteban Real and Software Engineer Chen Liang wrote: “We consider this to be preliminary work. We have yet to evolve fundamentally new algorithms, but it is encouraging that the evolved algorithm can surpass simple neural networks that exist within the search space.” (Google Research, July 9, 2020.)
Rank #3
AutoML-Zero is not the Evolved Transformer
A separate Google project can make the headline sound more sweeping. In 2019, Google Research described the Evolved Transformer, produced through neural architecture search. The distinction is what the system searched: AutoML-Zero searched for whole learning algorithms built from basic operations, while Evolved Transformer searched for a neural-network architecture.
| Project | What was searched | Reported evaluation and comparison |
|---|---|---|
| AutoML-Zero (2020) | Complete learning algorithms assembled from basic mathematical operations | Small image-classification problems; the reported performance comparison was against hand-designed models of comparable complexity in a toy scenario. |
| Evolved Transformer (2019) | Neural-network architecture, using existing architecture components | Google reported better BLEU and perplexity than the original Transformer across tested parameter sizes on English–German translation, with the strongest gains at smaller sizes. It also reported gains on additional translation pairs and nearly two fewer perplexity points in an LM1B language-modeling comparison. |
The Evolved Transformer’s translation and language-modeling results are a separate finding, not evidence for AutoML-Zero’s image-classification result. They were reported for the tasks and comparisons described by Google Research, not as a general ranking of AI systems. (Google Research, June 14, 2019.)
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Can you inspect or run AutoML-Zero?
Google’s open-source project repository includes code, a small demo for discovering linear regression, and instructions for reproducing baseline experiments. The README warns that the demo uses a much smaller search space than the paper, so it illustrates the idea but is not a reproduction of the full research-scale experiment.
The README lists Bazel and a C++ compiler as prerequisites. Its reproduction instructions are aimed at a developer workflow; they do not establish a specific consumer computer or accelerator as necessary for trying the small demo.
Quick Recap
How to interpret the headline
- Supported: automated evolution found a learning algorithm that outperformed hand-designed models of similar complexity in the reported toy experiment.
- Not supported: AI systems generally outperform human-designed models, or AutoML-Zero produced a new general-purpose AI.
- Useful context: the search rediscovered known learning techniques, and the researchers described the work as preliminary and compute-intensive.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
