Exploring Aigarth Intelligent Tissue 1.0
Written by
Qubic Scientific Team
Aug 19, 2025
Last week the Aigarth team released, in a pivotal moment for Qubic history, the first open-source code of a set of key components in our True-AI architecture, the ternary neural mechanism, the generation of evolutionary building blocks, the growth rules, and the basic arithmetic operations to test its performance.
If you are new to our jargon Aigarth is the metaphor of a “garden” for AIs to grow, we don’t believe in designing AIs, but on building the components for Artificial General Intelligence or AGI to take shape step by step, creating the nurturing environment where it will do it.
When Jose Sanchez and myself wrote back in December 2024 the Qubic AGI Journey paper (1), we had the ideas and blueprints from the Aigarth team, but now it is coming to light, with the three pillars already visible:
The Aigarth Intelligent Tissue (AIT) as the discrete mathematical substrate, think of the soil and plot layout, like a garden needs good soil as its foundation, we need a special mathematical "soil" where artificial "organisms" can live and grow.
A ternary computing approach that adds an “unknown” state to true and false. In the previous example plants that can be healthy, sick, or "not sure yet", normal computers think in binary: Yes/No, On/Off, Healthy/Sick, Aigarth “thinks” in three states: Yes/No/Maybe, like a plant that could be healthy, sick, or "we're not sure yet, let's keep watching"
And evolutionary dynamics for open‑ended improvement, it will be the natural selection in our garden, just like in nature, the "plants" (AI components) that work well survive and reproduce, the ones that don't work well die off, over time, our garden gets better and smarter without you having to redesign everything.
This is just a first step in a broader roadmap toward general intelligence (see the discussion in our paper (1) and “Fig. 4” on pages 24,25).
Aigarth Intelligent Tissue (AIT) source code is best understood as opening the box on a working prototype of a new kind of AI basic “material”, rather than shipping a finished, monolithic model, like the ones that we see in Hugging Face (a repository of open AI models), it is like a programmable “Petri dish”: a simple substrate where tiny computational organisms, or Intelligent Tissue Units, or ITUs, can self‑modify, compete, and get selected for doing useful work.
Although the brain ultimately works in a continuous and analog manner, ternary captures its functional diversity better than binary. Neurons can be at rest (0), excited (+1), or inhibited (–1), which is closer to balanced ternary logic than simple binary. Synapses can excite, inhibit, or have no effect, and the brain integrates millions of such signals in parallel.
At the code level, an Intelligent Tissue Unit (ITU) is built as a ring of simple neurons that exchange small signals, make a prediction, and then update themselves, these “neurons” think in balanced ternary values or "trits" of −1, 0, or +1, that way information naturally carries degrees like “negative,” “absent/unknown,” or “positive”.
For the technical folks, several functions convert integers into bitstrings and then into trits, feed those to the tissue, and convert the output back to familiar numbers. That same three‑state logic is the base of ternary computing, where the extra “unknown” state is not a quirk but a fundamental signal for uncertainty and incompleteness, this is a design choice that can make learning more robust when data are noisy or partial, (1, 2).
What turns this into a research platform, now open for everyone, is that the tissue’s built-in capacity to change itself, the mutate routine tweaks synaptic weights and, when a tweak would push a connection beyond its allowed range, it spawns a new neuron by cloning the most relevant neighbor. Over time, the circle can grow where growth helps and prune where pruning helps. The feed-forward cycle runs until either everything stabilizes, or the outputs are non-zero, or a tick limit is reached; the run is timed and the end condition recorded, promising modifications stick, unhelpful ones are discarded, and complexity emerges step by step.
Because AIT is designed for real experimentation, it includes versioning and data provenance, each ITU carries a cryptographically organized set with training seasons labeled, episodes incremented, datasets are hashed so anyone can verify which data produced which behavior, and micro-versions tick upward with each successful mutation. The object can be saved to or loaded from a tiny database (SQLite store), complete with metadata, so experiments can be paused, shared, and reproduced.
Beyond Qubic’s “Anna”
As the first task, if you remember what initially was named “Anna”, the GitHub repository (where the code is shared) includes detailed sources for the tissue to solve mathematical operations, again for the technical folks: we have an eight-bit integer addition, where two small, signed integers go in, the tissue, there it “reflects” on them until it settles on a result; the output is also an eight-bit signed sum. Under the hood, the addition unit encodes the two operands into ternary, runs the feed-forward dynamics, and decodes the output, along the way tracking how many output bits match the ground truth, how many remain “unknown,” and how many ticks it took to converge.
Evidently the idea is not that AIT replaces a “calculator”, but it provides a laboratory for evolution, since the code reports the “hit” rates and “unknowns” across many examples, it lets you identify mutations that are genuinely better rather than just different, that echoes the broader body of work in neuro-evolution, where researchers compare variants across many trials to separate signal from noise (3).
The tissue runs without specialized accelerators, favoring CPU-first computation and a distributed, deterministic workflow that many people can participate in, Aigarth aims to democratize experimentation, encourage independent verification, and reduce the cost of iterating on ideas. The argument is not that GPUs or NPUs are not needed, but that accessible CPU clusters with even simple machines, stitched together, can drive different kinds of innovation under realistic constraints.
Opening Qubic’s AI to a broader community
The release also opens a playground for ideas that have gathered momentum over the last decade, the tissue metaphor connects for example with the work on neural cellular automata, where small, local rules in a “living” grid produce complex, resilient patterns (4). (Check the authors demo here )
The decision to embrace a third, “unknown” value resonates with ternary quantization research showing that three-level signals can cut memory and energy while preserving useful precision in deep networks (2). And because AIT’s parameters and topology can shift over time, it offers a tractable testbed for architecture search, from gradient-based methods like differentiable architecture search to evolutionary and open-ended strategies such as quality-diversity and novelty search, which were designed to discover steppingstones rather than optimize a single score.
Our AI approach isn't ready to replace current systems yet (stay tuned), but it gives scientists and enthusiasts a testing ground to try out different methods - like systems that can remember things, make decisions based on local conditions, and handle uncertainty - instead of just turning everything into simple math calculations like today's AI does. Regarding ethics it has been embedded in Qubic since the beginning, when CFB started building it was already there (5), the license itself explicitly prohibits military use and flags the risks of misuse and bias.
Releasing this code it it’s the first step in an open call for the community to test, measure, and argue in the open, sharing Qubic AGI vision with a broader community.
Let’s see what you can build with it!
PD: For the technical folks (again) check the DeepWiki exploration of the code
David Vivancos
Qubic Scientific Advisor
Weekly Updates Every Tuesday at 12 PM CET
—-
Citations
1 Sanchez, J., & Vivancos, D. (2024). Qubic AGI Journey: Human and Artificial Intelligence: Toward an AGI with Aigarth (12/23/24). Qubic. https://www.researchgate.net/publication/387364505_Qubic_AGI_Journey_Human_and_Artificial_Intelligence_Toward_an_AGI_with_Aigarth
2 Zhu, C., Han, S., Mao, H., & Dally, W. J. (2017). Trained Ternary Quantization. arXiv:1612.01064.
https://arxiv.org/abs/1612.01064
3 Salimans, T., Ho, J., Chen, X., & Sutskever, I. (2017). Evolution Strategies as a Scalable Alternative to Reinforcement Learning. arXiv:1703.03864. https://arxiv.org/abs/1703.03864
4 Mordvintsev, A., Randazzo, E., Niklasson, E., & Levin, M. (2020). Growing Neural Cellular Automata. Distill, 5(2), e23. https://doi.org/10.23915/distill.00023
5 Come-from-Beyond (2024). Ethics Manifesto of an Artificial Intelligence Mentor https://medium.com/@comefrombeyond/ethics-manifesto-of-an-artificial-intelligence-mentor-847ad8ff8c23
Follow us on X @Qubic
Learn more at qubic.org
Subscribe to the AGI for Good Newsletter below.