Aigarth, shown: a year of Qubic science and BPP-9000

On 1 November 2026, Come-from-Beyond @c___f___b

Intelligence can be grown by evolution on ordinary processors, from small ternary networks that answer yes, no or not sure. Over the past year both teams put that bet on the table, in public.

The scientists: four peer-reviewed papers in a year

Qubic Science is led by two researchers: David Vivancos (Artificiology Research) @VivancosDavid and Dr. José Sánchez-García (UNIR) @josesanchezhb. They build Neuraxon, a neural network modeled on the brain. Its neurons have three states: excite, inhibit, and a neutral state where the neuron keeps listening but doesn’t fire. The network grows, learns and evolves all at once, with no separate training phase.

In 2026 they took it to four peer-reviewed conferences on three continents:

• 21 May, ICMLT 2026, Berlin: Neuraxon v2.0: A New Neural Growth & Computation Blueprint. Best paper presentation of the evening.

• 22 July, AMLDS 2026, Osaka: The Neutral Buffer State: Trinary Logic Advantage in Branching Ratio Stability for Continuous-Time Networks. Best paper presentation of the evening.

• 27–30 July, AGI-26, San Francisco: Multi-Neuraxon: Emergent Specialization, Modular, Frequency-Gated Neural Dynamics for Context-Dependent Behaviour. Published by Springer, LNCS / LNAI.

• 29 October, EA 2026, Nice: From Synapse to Society: Multi-level Continuity Through Neuromodulation in a Minimal Artificial Ecology. Oral presentation.

The Osaka paper lays out the formal basis of the ternary logic that Aigarth and Neuraxon share. AGI-26 is the main conference in artificial general intelligence. Code is open on GitHub, data on Hugging Face.

No LLM, no GPU: brains bred on ordinary processors

Papers are only half of it. All year the team ran evolution out in the open.

A world that doesn’t stop. Neuraxon Life is a Game of Life where creatures with Neuraxon brains see, smell, hear, eat, breed and die. Since summer it’s been running 24/7 on nxon.online. Its runs get published as datasets: from a million grown networks in December 2025 to the NeuraxonLive2M release in August.

ARC-AGI-3. The ARC Prize is one of the hardest public tests of machine reasoning: new games the agent has never seen. The team went in with evolved Neuraxon brains and no language model at all. Their score climbed step by step over summer and autumn: 0.13, 0.21, 0.25, 0.28, 0.38. On the last day, 2 November, it reached 0.41 on a 0 to 100 scale. 5,073 teams took part, and a lot of the leaders are built on large language models. Final places come after the private half of the test is checked.

155,192 brains, open to everyone. On 3 November they put every brain the search bred on Hugging Face: the Nxon2ArcAGI3Brains dataset. That’s 603,054 files, 100 GB, from 114 architecture search runs over 25 public games, under CC BY-SA 4.0. Anyone can download them and look inside.

What they found in Nice. Their EA 2026 paper asked if one brain-like agent can do well on three levels at once: inside its own head, in its world, and across generations. In Sánchez’s words: “Our answer at EA 2026 in Nice: partly.” Freeze the neuromodulation and the balance of fast and slow activity shifts fourfold, from 2.04 to 7.99. Those agents live a bit longer but find less food. “Living longest isnt adapting best.” An honest result.

Come-from-Beyond: from "we know how" to "accomplished"

Aigarth is Come-from-Beyond’s project inside Qubic: the search for a new way to make AI, where the AI is grown by the network’s miners. Qubic mining is useful work (uPoW!). The computation that secures the network also grows Aigarth’s networks, and every node checks the result.

He laid out this design back in 2023: “Aigarth uses Qubic’s UPoW, also Qubic has everything that an AI would need: Smart contracts as brain, Oracles as sensors to outer world, Outsourced computations as Manipulators to change that outer world.”

The last twelve months read like a run of milestones, in his own words:

• 6 December 2025: "And ETA for Aigarth was set as April 13, 2027" That’s five years after Qubic’s first block.

• 19 February 2026: "Aigarth already found how to get AI, now trying to find out how to get AI without 1000 nuclear plants"

• 21 May 2026: "We solved AI and now need practical proof which is just matter of getting enough power"

• Late August 2026: ant colony mining starts. A miner can publish a network only if it’s strictly better than the one it grew from. The miner is ranked by its single best network.

• 28 August 2026: "Every computor/candidate constructs a separate model predicting next hour price movement direction. So with 676+ models we may assess accuracy of the whole colony"

• 20 October 2026: "In a month historical BTC price feed will be replaced with live one" After one epoch, he wrote, qualification will depend on prediction accuracy: "It means that those who predict price correctly will be getting more computor seats"

• 1 November 2026: "Aigarth is accomplished, now need to show it to others with BPP9000"

The line goes from claim to proof. And the proof he picked is public, hourly and easy to check.

BPP-9000: Aigarth's first product and how it works

BPP-9000 (Bitcoin Price Predictor 9000) is Aigarth’s first product on Qubic. Its job fits in one line: before each hour closes, say whether bitcoin closes that hour UP or DOWN. A hard question, a clear answer and a score anyone can keep.

1. A tiny brain. Every network has 32 neurons. Each neuron has 3 wires and a table of 27 rules. On every tick a neuron reads its wires (up, down or unknown), finds its rule and outputs “up”, “down” or “not sure”. When the clock neuron fires, the answer neuron calls the hour.

2. Grown by miners. Qubic miners mutate these networks and score them on a real year of hourly BTC/USDT candles from Binance, about 8,760 hours. The score is the number of hours the network got wrong. Lower is better.

3. Only the better children survive. In ant colony mining a miner can only publish a network if it’s strictly better than the one it grew from. Each miner is ranked by its single best network.

4. Everyone checks. Every Qubic node recomputes every score with the same scorer from qubic/core (score_bpp9000.h). The top 676 miners become the computors that run Qubic for the next week. Jetski ANN Explorer archives and re-checks every solution. And QC runs the best fresh network right in your browser, by the same rule.

5. Live from 18 November. From epoch 240 the Qubic core asks for the BTC price of the past hour. Computors fetch the real data, agree on it by quorum and feed it in. This is Qubic’s Oracle Machine. Then the network calls the next hour, and the call is checked against the real close.

6. Seats follow accuracy. After one epoch under the old rule, qualification depends on how well a miner’s model predicts. Computor seats go to the networks that call the market right. Mining itself becomes the test.

What to watch. Until the live feed, the match rate on the page is a fit to last year’s history, not a forecast. A coin is right half the time. What matters is how far above 50% the calls hold live, over many hours, made before each hour closes and checked against the real close.

18 November: the crown of the year

Everything above comes together on one date:

• Neuraxon showed in public, at four conferences and in 155,192 published brains, that small ternary networks bred on processors can learn hard things.

• Aigarth runs the same kind of networks inside Qubic itself. There hundreds of miners breed them every week, and every node checks the work.

• BPP-9000 is where that work stops being a claim and starts becoming a track record.

From 18 November every call is public and timestamped. It’s made before the hour closes and checked against the market an hour later. Nothing can be fitted after the fact. The longer it runs, the clearer the answer gets, whichever way it goes.

That’s what “show it to others” means.

Join the live experiment

Come-from-Beyond asked the community to put together a group of people who follow the hourly calls and report what happens. That way the test has enough witnesses to mean something. Qubic Church built a place for it.

• Watch: QC (https://qubic.church/bpp9000) runs the network live in your browser and counts down to the first live call.

• Join: sign in with a Qubic wallet, or just a name and a secret code. Set your starting capital and keep a journal of your calls and trades.

• Check: BPP-9000 is scored only on its own calls against the market, separately from anyone’s journal. Its mining scores are recomputed by every Qubic node.

First live call: 18 November 2026. Be there when it starts.

QC doesn’t hold any personal data.

This is an experiment, not financial advice. If you trade with real money, the responsibility is yours.