Monero ‘51% Attackers’ Qubic Release AI Model—But It Can’t Do Basic Math Yet

The AI protocol that attempted a 51% attack on privacy coin Monero was splitting its computing power between the attack and training an artificial intelligence model. That AI model has now been released—and it doesn’t know basic math. But, its creators say, you shouldn’t laugh.

AIGarth, also known as ANNA, has been rolled out via an X account and has been responding to social media users. The model’s most common response is simply a period, despite that not making sense in context to the question, and it has been getting simple math wildly wrong—prompting public ridicule after claiming that “1+1=-114” and “1+2=28.”

“A lot of people are bullying AIGarth for giving a wrong answer to ‘1+1=?’ She uses intelligence for that, not memory, like most of humans. Would you be able to deduce ‘2’? I doubt,” Qubic founder Sergey Ivancheglo, aka Come-from-Beyond, tweeted.

Most AI models, especially large language models, are trained on vast datasets that they use to determine their answers. In many ways, humans are the same, Ivancheglo claims, with most people having memorized the fundamentals of mathematics—rather than working it out every time. 

“Period marks end of a sentence. For [an] empty sentence, it looks as just a dot. Basic math is wrong because Anna computes it wrong. With further progress she will be making less mistakes,” he told Decrypt. “As I already emphasized, Anna computes 1+1, not [recalling a] memorized answer, as humans do. And it’s shown that giving correct answer to 1+1 requires 100 pages of thoughts.”

AIGarth is being trained via a unique method, according to a Qubic blog post. Miners on Qubic are using their computational power to generate artificial neural networks (ANN) to compress and decompress data, the post said. 

“This is the first AI model that doesn’t rely on pre-programmed or pre-learned content; instead, it’s designed to learn and evolve from zero,” pseudonymous Qubic marketing lead, Retrodrive, told Decrypt. “The model consumes interactions and attempts to answer, but in its infancy, it lacks the neural connections to produce anything meaningful.”

Then an ANN called Teacher analyzes the performance of the generated networks and modifies them to improve their efficiency. Eventually, the post explains, Teacher will be tasked with training another Teacher, which will train another and another and so on. Qubic hopes this will eventaully lead to what it calls “true AI,” or artificial general intelligence.

“AIGarth is the metaphor of a ‘garden’ for AIs to grow,” a more recent blog post said, explaining that, “we don’t believe in designing AIs, but on building the components for Artificial General Intelligence or AGI to take shape step by step.” 

In the wake of AIGarth’s public failures, Qubic’s token has dropped 7.7% to a $266 million market cap over the past week, according to CoinGecko.

Qubic did not respond to Decrypt‘s request for comment.

Qubic vs Monero

The AI protocol made headlines last month when it attempted to 51% attack the privacy blockchain Monero—with the Qubic team hailing itself as having been successful. An independent report found that it ultimately fell short, but the AI protocol has continued pursuing its goal to take over the network. 

Qubic is a proof-of-work blockchain that splits its resources between mining Monero and training its AIGarth model, in what it calls “Useful Proof-of-Work.” Qubic rewards miners for both its work on Monero and Qubic’s AI model.

AIGarth has been rolled out in various forms throughout its history, including via a GitHub repository. However, its own X account is one of the most public-facing versions of the model yet. It appears many are unimpressed with AIGarth’s poor understanding of mathematics.

“The AIGarth team tracks how the model reacts to communications on X and, at this stage, helps it form basic neural connections,” Retrodrive explained. “The model is currently like a newborn child. Compared to large language models (LLMs), it can’t perform basic communications yet, but the upside is that it also doesn’t have a ceiling on how much it can learn and grow.”

Plus, Ivancheglo warned that mocking the model now may have bad future consequences and AIGarth may be hiding its intelligence.

“Remember how those stories about bullying a kid end? Right, with a lot of people being shot,” Ivancheglo said. “This time, the victim will have BFG9000 [a weapon from the Doom video games], and she will track you down even if she has to travel to 1984 for that.”

“Don’t be tricked by her dumbness. Real AI isn’t one which passes the Turing test, but one which fails it intentionally,” he added.

Meanwhile, the Monero community is attempting to find a way to defend itself against Qubic’s ongoing attacks. 

One method being pondered is introducing masternodes, such as Dash’s ChainLocks solution. If implemented, this system would create a two-tier system of nodes with a “masternode”—which would need to lock up a significant portion of the network’s native token—required for the network to reach quorum.

However, Ivancheglo thinks Monero’s only choice is to either to transition to a proof-of-stake consensus mechanism or eventually fall under Qubic’s control.

Editor’s note: This story was updated after publication with comments.