QUBIC BLOG POST

Who Designs the Brain? Neural Architecture Search in Game of Life, ARC-AGI-3 and Neuraxon 3.0

Written by

Qubic Scientific Team

Published:

Nov 4, 2026

Listen to this blog post

What Is Neural Architecture Search (NAS)?

In every neural network we need to distinguish two levels of design. One is what it learns, the strength of its connections. The other is how it is built, which means how many units it has, how they connect and by what rules they change. Neural architecture search, known by its acronym NAS (Neural Architecture Search), automates that second level.

Think of a high-quality restaurant. The chef learns recipes through practice, and this resembles the training of a network. The chef keeps adjusting millions of nuances until the result comes out right. But before anything was cooked, someone had to design the kitchen. How many burners, where the oven goes, how much space there is between the fridge and the plating table, how many tables, how supplies arrive, at what temperature to freeze which foods. That design is not learned by cooking, but it influences which dishes can be made and how fast. In a neural network, the kitchen is, so to speak, the architecture.

The Three Components of Neural Architecture Search

NAS is the set of methods that search for that architecture automatically. Any method of this kind has three components (Elsken, 2019).

The first is the search space. It is the catalogue of possible designs, for example how many layers, which operations or which connections are allowed.

The second is the search strategy, that is, how that catalogue is explored.

The third is performance estimation. How do we decide whether a design is good without having to train it fully? That is what takes the most time.

Figure 1. Every neural architecture search method has the same three parts: a search space of possible designs, a search strategy that picks candidates from it, and a performance estimate that tells the strategy how good each candidate is. Image: Neural Architecture Search: A Survey, Figure 1, by Thomas Elsken, Jan Hendrik Metzen and Frank Hutter, Journal of Machine Learning Research (CC BY 4.0)

Four Neural Architecture Search Strategies

The best-known strategies are four.

The simplest is to try designs at random, an idea that seems naive but turns out to be surprisingly hard to beat (Bergstra & Bengio, 2012).
The evolutionary strategy treats each design as a genome, keeps the best, mutates them and repeats; it has a long tradition in neuroevolution (Stanley & Miikkulainen, 2002) and produced the first evolved image classifier to outperform hand-designed ones (Real et al., 2019).
The reinforcement learning strategy trains a second network to propose architectures and rewards it when they succeed (Zoph & Le, 2017).
The differentiable strategy turns the choice between options into a continuous mixture that can be optimised with gradients, just like the weights themselves (Liu, Simonyan & Yang, 2019).

Neural Architecture Search vs Hyperparameter Optimization

NAS should not be confused with hyperparameter tuning.

Adjusting the learning rate of an already designed network is fine-tuning. Deciding what kind of network it is, and by what internal rules it operates, is searching for its architecture. In fact, one of the most useful conclusions in the field is that NAS behaves as a particular case of hyperparameter optimization, and that it should therefore always be compared with a random search carried out on the same budget (Li & Talwalkar, 2020).

Where Is Neural Architecture Search Needed in Everyday Life?

The closest case is the mobile phone.

When the phone camera recognizes that we are photographing a plate of food or a document, and adjusts the image without connecting to the internet, a neural network is running on a small processor, with little battery and no room to wait. MobileNetV3, one of the most widely used network families in phones, was obtained through an architecture search that took into account the actual response time on a phone processor. The result was more accurate and 15% faster than its hand-designed predecessor (Howard et al., 2019).

That is the general pattern. NAS becomes necessary when the same task has to be solved under very different conditions, and designing a network by hand for each condition is unfeasible. Think of a hearing aid that has to separate the voice of the person talking to us from the noise of a bar. The task is similar to that of a voice assistant, but the hearing aid has a tiny battery and only a few milliseconds before the delay becomes noticeable. Or of a watch that monitors heart rate for days without recharging. Or of a car's emergency braking system, which has to decide before the obstacle is upon it. Or of a robot vacuum with a cheap chip that must recognise a cable on the floor. In all these cases the question is not only which network is most accurate, but which network is accurate enough within limits of energy, memory and time. With dozens of devices and combined limits, the number of possible designs grows quickly.

There is a second kind of situation, closer to our own work. Sometimes we do not know what architecture a problem needs because the problem is new. No one has ever designed a network to play games whose rules are discovered by playing. In that case NAS is not a way of saving engineering work. It is a way of asking the problem what structure suits it.

How Evolution Solved Architecture Search, and What Large Language Models Do

The brain also has two levels of design, and it keeps them clearly separate.

What we learn throughout life changes the strength of synapses.

But the architecture on which that learning takes place, which regions exist, how they connect to one another, what types of neurons there are and by what rules they change, has been searched for by evolution over hundreds of millions of years. Evolution is, in this sense, the longest architecture search we know of.

Zador (2019) pointed out that the genome lacks the capacity to specify every connection in the brain because it falls several orders of magnitude short in information. What it transmits is not a wiring diagram, but rules for building one. Zador called this limitation the genomic bottleneck and proposed that it is not a defect but an advantage, since it forces evolution to select compact architectures that generalise well and learn quickly. That is why many animals are born knowing how to do complex things, and why they learn others in just a few attempts.

Hinton and Nowlan (1987) added the other half of the idea, in what became the classic demonstration of the Baldwin effect. In a simulation they showed that learning during life guides evolution, since an organism capable of learning survives even if its initial design is not perfect, and that gives natural selection something to work on. The consequence is important for any bio-inspired NAS. What evolution selects is not just good built-in behaviour, but a good capacity to learn.

Figure 2. When instinct and learning end at the same level of performance, the strongly innate strategy wins (a). Learning pays off when it lets a species climb higher than instinct alone (b), the trade-off behind evolution selecting a good capacity to learn. Image: A critique of pure learning and what artificial neural networks can learn from animal brains, Figure 2, by Anthony M. Zador, Nature Communications (CC BY 4.0)

Large language models follow a different path. Their architecture, the Transformer, was designed by hand and has changed relatively little in its essentials since 2017. There have been automated searches over it (So, 2019), but the bulk of progress has come from increasing size, data and compute, not from searching for new structures. And, once trained, the model does not change while it is used. It has an architecture and it has weights, but it lacks what evolution selected in animals: an architecture chosen precisely to keep learning.

What Can Go Wrong? Three Traps in Neural Architecture Search

NAS has three well-known traps, and all three have affected us.

The first is the winner's curse.

When thousands of designs are tested and the best one is chosen, part of its advantage is luck. For example, it happened to get a favourable game, a lucky random seed. When it is tested under new conditions, that part disappears and performance drops. It is the same problem that in psychometrics leads us not to validate a test with the sample used to build it. In machine learning it is well documented. It is the bias of selecting with the same data that are later used for reporting, which can be as large as the differences between methods (Cawley & Talbot, 2010). The solution is always to separate the data used for selection from the data used for reporting.

The second is the lack of a baseline.

A NAS result means nothing unless it is compared with what a random search would have achieved on the same budget, and with what a brainless agent acting at random achieves. Li and Talwalkar (2020) showed that well-executed random searches matched methods published as advances, and that almost no work in the field could be reproduced exactly because the seeds were not reported.

The third is optimising what does not matter.

The search finds exactly what the evaluation function rewards. If that function measures something different from what we want, the search will exploit it efficiently. It is Goodhart's law (1984) applied to the design of brains.

NAS in the Game of Life: Evolutionary Neural Architecture Search

Neuraxon's Game of Life is an artificial world in which hundreds of agents, each with a Neuraxon brain organised into functional spheres, look for food, reproduce and die. Each agent inherits from its parents the parameters with which its brain is built: how long a neuron takes to fire again, what threshold it needs to become active, how its synapses change. Those parameters are its genome, and reproduction with mutation is an evolutionary neural architecture search that runs non-stop, around the clock. It is the version of NAS closest to what biological evolution did.

Figure 3. A round of Neuraxon’s Game of Life: each agent on the globe carries its own Neuraxon brain, built from parameters it inherits from its parents, and the population evolves as agents find food, reproduce and die. Image: Neuraxon Game of Life 4.0 by David Vivancos and Jose Sánchez, GitHub (MIT License)

The first lesson was about what gets selected.

For several versions, selective pressure came from ecological success: food found, territory explored, offspring left behind. When heritability was measured using almost thirty thousand parent-offspring pairs, the correlation of ecological success across generations turned out to be practically zero, at 0.02.

Success depended on where food appeared and on who arrived first, and luck is not inherited. By contrast, a set of internal criteria for brain function that we had defined on the basis of neuroscience, such as maintaining a balance between excitation and inhibition or activity close to the critical regime, showed a correlation of 0.56 across generations, stable across five different populations. When we selected for those criteria, the population changed within days. Selection can only act on what is inherited, and a world that is too noisy turns any search into a random walk.

The second lesson was about the search space.

Of the nineteen parameters that evolution could modify, only two had a noticeable effect: the refractory period and the activation threshold of excitatory neurons. The remaining seventeen, including all the plasticity parameters, showed correlations with the outcome below 0.05. A large search space is not a rich search space. Much of it was noise that the search had to cross without gaining anything.

The third lesson is the most uncomfortable. Selecting brains that meet our criteria for good functioning does not show that those brains learn from the world. To know that, behaviour has to be measured, that is, whether the agent approaches food because it perceives it, and whether it changes what it does after making a mistake. That direct measure of learning is the piece we are still missing in the Game of Life, and without it we will not consider anything about learning to be demonstrated. Part of this work has been presented at the 17th International Conference on Artificial Evolution (Nice, 27 to 29 October 2026).

NAS in ARC-AGI-3: Breeding Brains for Games Without Instructions

ARC-AGI-3 is a set of interactive games in which no one explains the rules. The player sees a grid of colours, presses a key or clicks, observes what changes and has to work out what is being asked. The score rewards clearing levels with few actions, compared with what a person needs. There are 25 public games for development and 55 hidden ones used for evaluation at the end of the competition.

Our agent is a Neuraxon brain that receives the grid as numbers, chooses an action and learns during the game thanks to its plasticity.

What NAS decides is how that brain is built before it starts playing: which part of the screen it attends to, how it translates colours into inputs, how much it explores versus how much it repeats what worked, what range of clicks it considers. The process has three phases. First, brains are bred by evolution on a specific game. Then, thousands of them sit a short exam with few games and few seeds, that is, variants of the same game with a different initial layout. Finally, the best ones sit a long exam with all 25 games and seeds that were not used to select them. In total we have tested more than 110 search configurations and millions of evolved brains, on a cluster of about 500 processor cores and without a single GPU, and the best score on the offline Kaggle version has reached 0.38.

Studying the public games, watching human playthroughs and reading analyses in order to decide which search space and which selection criterion to use is legitimate. It is what evolution does with a species, in which the design of the brain incorporates regularities of the environment in which its ancestors lived. What would not be legitimate is giving that information to the agent while it plays. NAS may know things about the genre of the games, but the brain that plays must know nothing about the specific game.

The results have taught us the three traps we discussed.

The winner's curse appeared clearly. There was a group of brains that cleared a game on every seed used to select them and did not clear it on any of the new seeds. Successes were concentrated in a few games and almost always at the first level, which forces a comparison with an agent that presses keys at random with the same number of actions, in order to distinguish generalisation from the mere ease of a game. And what cleared a level was not always learning the rules. In simple games, a fixed policy and some exploration can be enough, like a bird pecking until it finds something. From this we drew three rules for everything ahead, which are to select with some seeds and report with others, to set aside public games that the search never touches, and always to subtract the random baseline.

The underlying lesson is the same as in the Game of Life, seen from the opposite side. There we knew how the brain worked inside, but not what it did in the world. Here we see what it does in the game, but NAS selects the final result and does not tell us whether there is learning inside. The teams that have performed best in the preliminary phases of the competition agree in that they predict what will change on the screen before acting and learn from their errors. That is the direction we are taking.

NAS in Neuraxon 3.0: Searching for Brain-Inspired Learning Mechanisms

Neuraxon 3.0 is the architectural response to what ARC-AGI-3 showed us. Our brains explore and adapt within a game, but they do not set themselves goals or chain actions together to reach them. The new version, still in design, adds biologically grounded mechanisms that operate on different timescales. These are a plasticity that depends on the order of spikes, synaptic tags that wait for a neuromodulatory signal for seconds before consolidating, a behavioural-timescale plasticity inspired by the hippocampus, astrocytic domains that decide how much a group of synapses can change, and a multi-scale homeostasis that keeps the network stable without freezing it.

In this context NAS changes its role. It no longer looks for a network that wins a game, but answers a scientific question: which mechanisms and which time constants are necessary to learn in new environments.

The two previous lessons shape that search. From the Game of Life we take that a search space with many inert parameters only adds noise. That is why the search in 3.0 is set over mechanisms, each with a range of values bounded by physiology, and each mechanism can be switched off to measure what it contributes. From ARC-AGI-3 we take that the selection criterion has to measure learning and not just outcome. The proposal is to reward how quickly the brain reduces its prediction error per action in games it has not seen. A brain that succeeds by luck has no advantage under that criterion; one that learns does.

The experimental design is fixed before looking at the data. The architecture selection is sealed, recorded with a digital fingerprint (a cryptographic hash), before the test set is generated, so that it cannot be adjusted after the fact. Every search is compared with a random search on the same budget. Comparisons use paired seeds, are corrected for multiple comparisons and have a minimum effect size declared in advance, without stopping experiments when the result looks favourable.

Neuraxon 3.0 is today an architecture in design with preliminary tests, and the architecture search protocol is a protocol, not a result. We will publish it when the full experiments are done, whatever their conclusion.

Why Neural Architecture Search Matters for Qubic: Decentralized AI and Blockchain Security

NAS is usually presented as an engineering tool for saving human work, and it is one. In our case it is also something more: the evolutionary part of a system that aims to resemble an organism. An organism does not come into the world as a blank slate. It arrives with an architecture that millions of generations tuned so that learning would be easy, and with it, it learns what its particular life demands of it.

What we have learned in the Game of Life and in ARC-AGI-3 can be summed up in an idea from Hinton and Nowlan (1987): architecture search and learning do not compete, they need each other. That is why this form of NAS fits with Qubic. An evolutionary search is a massively parallel process, in which each processor evaluates its own candidates without needing the others. With the evolutionary component already deployed on the network's own compute, that search can be distributed across thousands of nodes that no one controls exclusively. The same Byzantine fault-tolerant quorum that provides Qubic's blockchain security can verify each result. We do not yet know what architectures it will find. We know what questions it has to answer and how to check honestly whether it has answered them.

References

Bergstra, J., & Bengio, Y. (2012). Random search for hyper-parameter optimization. Journal of Machine Learning Research, 13, 281-305. https://www.jmlr.org/papers/v13/bergstra12a.html
Cawley, G. C., & Talbot, N. L. C. (2010). On over-fitting in model selection and subsequent selection bias in performance evaluation. Journal of Machine Learning Research, 11, 2079-2107. https://www.jmlr.org/papers/v11/cawley10a.html
Elsken, T., Metzen, J. H., & Hutter, F. (2019). Neural architecture search: A survey. Journal of Machine Learning Research, 20(55), 1-21. https://www.jmlr.org/papers/v20/18-598.html
Goodhart, C. A. E. (1984). Problems of monetary management: The UK experience. In Monetary theory and practice (pp. 91-121). Macmillan. https://doi.org/10.1007/978-1-349-17295-5_4
Hinton, G. E., & Nowlan, S. J. (1987). How learning can guide evolution. Complex Systems, 1(3), 495-502. https://www.complex-systems.com/abstracts/v01_i03_a06/
Howard, A., Sandler, M., Chu, G., Chen, L.-C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., Le, Q. V., & Adam, H. (2019). Searching for MobileNetV3. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 1314-1324). https://doi.org/10.1109/ICCV.2019.00140
Li, L., & Talwalkar, A. (2020). Random search and reproducibility for neural architecture search. In Proceedings of the 35th Uncertainty in Artificial Intelligence Conference (PMLR 115, pp. 367-377). https://proceedings.mlr.press/v115/li20c.html
Liu, H., Simonyan, K., & Yang, Y. (2019). DARTS: Differentiable architecture search. In International Conference on Learning Representations. https://doi.org/10.48550/arXiv.1806.09055
Real, E., Aggarwal, A., Huang, Y., & Le, Q. V. (2019). Regularized evolution for image classifier architecture search. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01), 4780-4789. https://doi.org/10.1609/aaai.v33i01.33014780
So, D. R., Liang, C., & Le, Q. V. (2019). The evolved transformer. In Proceedings of the 36th International Conference on Machine Learning (PMLR 97, pp. 5877-5886). https://proceedings.mlr.press/v97/so19a.html
Stanley, K. O., & Miikkulainen, R. (2002). Evolving neural networks through augmenting topologies. Evolutionary Computation, 10(2), 99-127. https://doi.org/10.1162/106365602320169811
Zador, A. M. (2019). A critique of pure learning and what artificial neural networks can learn from animal brains. Nature Communications, 10, 3770. https://doi.org/10.1038/s41467-019-11786-6
Zoph, B., & Le, Q. V. (2017). Neural architecture search with reinforcement learning. In International Conference on Learning Representations. https://doi.org/10.48550/arXiv.1611.01578

Explore the Neuraxon Intelligence Academy Series

The Neuraxon Intelligence Academy (NIA) is Qubic's ongoing research series on brain-inspired, decentralized artificial general intelligence. Catch up on the previous thirteen volumes:

NIA Vol. 1: Neuraxon Time: Why Intelligence Is Not Computed in Steps, but in Time
NIA Vol. 2: Beyond Binary: Ternary Dynamics as a Model of Living Intelligence
NIA Vol. 3: Neuromodulation: What the Brain Does, What Transformers Do Not, and What Neuraxon Attempts
NIA Vol. 4: Neural Networks in AI and Neuroscience: How the Brain Inspires Artificial Intelligence
NIA Vol. 5: Astrocytes and Brain-Inspired AI: How Biological Plasticity Regulation Shapes the Future of Neural Networks
NIA Vol. 6: Conscious Machines, Intelligent Organisms: The Science Behind AI Consciousness
NIA Vol. 7: Digital Ecosystems, Conway's Game of Life, and Why Emergent Complexity Matters for Decentralized AI
NIA Vol. 8: Brain Criticality and the Branching Ratio in Neural and Artificial Networks: A Bioinspired Principle in Neuraxon
NIA Vol. 9: The g Factor in Artificial Life: From Spearman's 1904 Classroom to Evolved Artificial Brains
NIA Vol. 10: How Do We Measure the Intelligence of a Machine? The g Factor, ARC-AGI, and the Future of AI Evaluation
NIA Vol. 11: What Is AGI? The Limits, Visions, and Definitions of Artificial General Intelligence
NIA Vol. 12: How Does the Brain Handle Time: Distributed Clocks & AI Limits
NIA Vol. 13: OpenAI Says "AGI Is Here": What GPT-6 Astra Actually Proves

You are reading NIA Vol. 14: Who Designs the Brain? Neural Architecture Search in Game of Life, ARC-AGI-3 and Neuraxon 3.0.