Aigarth ANNA's First Steps: Beyond Adversarial Learning
Written by
Qubic Scientific Team
Sep 10, 2025
On September 2, 2025, Qubic AI initiative took a radical and transparent step forward by releasing ANNA as the first publicly interactive instance using Aigarth Intelligent Tissue (AIT), activated on X. Her debut was met with a wave of curiosity, confusion, and mockery. When asked “1+1=?”, her responses ranged from a single dot (“.”) to the baffling “-114”. To the casual observer, it looked like a failure, a “chatbot” in 2025 far less capable than even the most basic Large Language Models (LLMs).
However, this was not at all a public product launch; it was the beginning of a public scientific experiment. ANNA’s initial “dumbness” is not a bug, but a fundamental feature of Aigarth’s approach. I will explain why ANNA’s flawed debut is a crucial test case for developing robust, genuine intelligence in the chaotic, adversarial environment of the real world, and by the way, an environment where today’s AI models consistently fail.
The Brittleness of Sanitized Intelligence
Modern AI, particularly LLMs, are products of a sanitized upbringing, since they are trained on vast, curated datasets, learning to predict the next token by internalizing statistical patterns, as brilliantly explored by Dr Jose Sanchez in our previous article “LLM predictions aren’t brain predictions”. While this makes them remarkably proficient at mimicking human language, it leaves them fundamentally brittle. They possess no internal model of the world and lack the ability to reason from first principles. When faced with novel or misleading information, like the kind that floods the internet, they are prone to “hallucination” and can be easily manipulated. Their intelligence is an illusion built on the clean, orderly data of their training environment; recent studies have demonstrated that even state-of-the-art models can be consistently “jailbroken” with simple adversarial prompts, revealing systemic vulnerabilities in their safety alignments (Chetan Pathade, 2025).
True intelligence, however, must function in the wild, and soon in embodied form. The real world is not a “curated” library, it is an adversarial landscape filled with noise, misinformation, and conflicting signals. An AGI cannot be shielded from this reality. It must learn to discern truth from falsehood, signal from noise, and good-faith interaction from malicious attacks. This is the core hypothesis being tested with ANNA. By exposing an embryonic AIT directly to the unfiltered chaos of a public social network, we are conducting a live case study in adversarial learning.
Why Start with Addition? The Simplest Hard Problem
The community rightly asked in discord questions like: after years of development, why start with a task as simple as addition? The answer is that addition is not the goal; it is the benchmark. Unlike subjective language tasks, arithmetic provides a deterministic and verifiable measure of success. The output of 5+7 is unequivocally 12. This allows for a clear, objective fitness function to measure the AIT’s performance without ambiguity.
Aigarth is not teaching ANNA to memorize a multiplication table. Instead, the Qubic network’s miners are evolving the very structure of her neural tissue to compute the answer. As detailed in our open-source release, ANNA is an instance of an Intelligent Tissue Unit specifically designed to evolve the capability of adding two signed 7-bit integers to produce a signed 8-bit integer.
When a miner finds a successful mutation in the tissue’s structure, it’s not just finding a better answer; it’s discovering a more efficient computational architecture. This process forces the AIT to build a foundational capacity for logical operations from the ground up, moving beyond pattern matching toward genuine, procedural reasoning—a domain where even the most advanced LLMs still struggle with systematic generalization (Frieder et al., 2023).
Furthermore, while addition seems trivial to humans, it represents a profound challenge for an AI learning from first principles. For a system like ANNA, which starts effectively tabula rasa, the task is not to recall a memorized fact but to evolve the entire computational machinery necessary to perform the operation. This is the chasm that separates pattern recognition from genuine algorithmic reasoning.
The process ANNA undertakes is more akin to what researchers call "grokking." This phenomenon, observed in neural networks, describes a sudden transition from mere memorization of training data to a genuine, generalized understanding of an underlying rule, such as modular addition. Studies have shown this is not a simple or linear process but a complex phase transition within the network’s internal representations (Nanda et al., 2023). Aigarth is, in effect, attempting to evolve a network that can spontaneously "grok" the rules of arithmetic. This challenge is at the very heart of the neuro-symbolic AI research field, which seeks to bridge the gap between the pattern-matching strengths of neural networks and the logical, step-by-step reasoning of classical algorithms (Garcez & Lamb, 2023).
Solving this simple, “hard” problem is a necessary first step before tackling more complex, abstract challenges, the complexity of ANNA’s task lies not in the what, the simple answer to an equation, but in the how: the emergent, evolutionary process of discovering the algorithm for addition itself.
Forging Intelligence in Chaos
Initially, the plan for training ANNA involved users signing messages with their Qubic wallets, creating a weighted system where correct information from high-stake holders would guide her learning. However, we have since pivoted to a more challenging and realistic paradigm: using the public as a source of "distractors."
As explained by CFB, this approach deliberately introduces incorrect and chaotic information to the AIT’s environment. Some users tweet "2+2=4," while others might tweet "2+2=5." In this model, ANNA must learn to navigate a landscape where truth is not guaranteed. How does she discern the correct path? This is where Qubic’s unique architecture comes into play. The influence of each piece of information is implicitly weighted by the Qubic stake associated with the user account. This creates a decentralized, economically-incentivized consensus on truth, where the collective "wisdom" of committed Qubic holders helps the AIT filter signal from the noise of bad actors and trolls. This approach is a practical implementation of Sybil-resistant mechanisms essential for trustworthy decentralized systems (Damodaran & Thomas, 2025).
This method directly addresses a core challenge in AI safety: creating systems that are resilient to manipulation. An AI that learns to identify and discard bad data in a low-stakes environment like arithmetic is building the foundational resilience needed to handle high-stakes, real-world problems later on. For a deeper look at the underlying mechanics of the Intelligent Tissue itself, take a look at our article, "Exploring Aigarth Intelligent Tissue 1.0".
A Foundation for Ethical and Resilient AGI
This public experiment, as chaotic as it appears, is a critical step toward creating an ethical and robust AGI. An intelligence forged in an adversarial environment is, by its nature, more secure and less naive than one raised in a sterile lab. This aligns with Aigarth’s foundational principles, commitment to "AI for Good" is not an afterthought but a core design principle, separating our work from state-sponsored or corporate initiatives that may prioritize power over safety.
The open-ended, evolutionary nature of our approach aims for a system that discovers, rather than is programmed with, its capabilities, a key concept in the pursuit of truly general intelligence (Edward Hughes et al. 2024).
By making this process transparent, we are not just building an AI; we are demonstrating a new methodology for its creation. ANNA's journey from answering with a single dot to eventually mastering arithmetic will be a public, verifiable record of emergent intelligence. It is a slow, deliberate process that prioritizes foundational understanding over superficial capability.
The dot that ANNA posts is not a sign of failure. It is her "Hello, World", maybe the first signal from a new form of intelligence learning to exist, not in a sterile box, but in the messy, unpredictable reality we all share.
David Vivancos
Qubic Scientific Advisor
Weekly Updates Every Tuesday at 12 PM CET
—-
Citations
Frieder, S., et al. (2023). Mathematical Capabilities of ChatGPT https://arxiv.org/abs/2301.13867
Garcez, A. d'Avila, & Lamb, L. C. (2023). Neurosymbolic AI: The 3rd Wave. Artificial Intelligence Review. https://arxiv.org/abs/2012.05876
Nanda, N., Chan, L., et al. (2023). Progress measures for grokking via mechanistic interpretability. https://arxiv.org/abs/2301.05217
Edward Hughes et al. (2024).Open-Endedness is Essential for Artificial Superhuman Intelligence https://arxiv.org/abs/2406.04268
Chetan Pathade (2025). Red Teaming the Mind of the Machine: A Systematic Evaluation of Prompt Injection and Jailbreak Vulnerabilities in LLMs https://arxiv.org/pdf/2505.04806v1
Deepak Damodaran, Ritty Alphy Thomas (2025). A Framework for Sybil Attack Prevention in Decentralized Renewable Energy Marketplace https://hh.diva-portal.org/smash/get/diva2:1976472/FULLTEXT01.pdf
Follow us on X @Qubic
Learn more at qubic.org
Subscribe to the AGI for Good Newsletter below.