Aigarth, BPP-9000 and Neuraxon: how it works
A network grown by Qubic miners learns to call bitcoin's next hour. From 18 November 2026 it calibrates on live hours it has never seen; the calls go public at the next stage. This page explains the three names behind it, how a call is made, how the live calibration works, who the scientists are, and what it would mean to join their brain model to the network. Every quote is word for word, with its date and place. Where we interpret, we say so.
1. Aigarth: a search for a new kind of AI
Aigarth is Come-from-Beyond's research programme inside Qubic. Its bet is that general intelligence is not written by hand or trained into one giant model, but grown: many small networks are mutated, scored and selected, and the good structure survives. Qubic's miners supply the compute for that search.
“Aigarth's main goal is to discover a new paradigm allowing to create true AI, I have just seen the wireframe of this paradigm. And I have realized that Artificial Intelligence will not be created, it will emerge. With help of Qubic miners.”
The engineering form is the Intelligent Tissue: small ternary units, each in one of three states and wired to a few neighbours, released as open source in July 2025. The first public showcase was Anna, and in June 2026 Anna.exe let anyone race a network down to zero errors on its test. The author's own claims grew over 2026, and we keep them in his words:
“Current paradigm developed by Aigarth, while not finished, already provides a hypothetical solution to the most important problem of Artificial Neural Networks — the generalization problem.”
“Finished means that new paradigm is created and now we need to polish some details”
“Aigarth is accomplished, now need to show it to others with BPP9000”
Our reading: “accomplished” is a claim about the paradigm, not a demonstration of general intelligence. The author himself names BPP-9000 as the way to show it. That is why a public record of every live call matters.
2. BPP-9000: the task that tests it
BPP-9000 is the task Qubic mining runs today. It asks one narrow, hard, checkable question every hour: will bitcoin close at or above the hour's open, or below it, counted in whole dollars as the core counts it? It is narrow on purpose. A single number per hour is easy to verify and impossible to argue with afterwards.
- 1
The exam
A real year of bitcoin, BTC/USDT on Binance, hour by hour: about 8,760 hours, each turned into one trit, up or down. Every network is graded on how many of those hours it gets wrong.
- 2
The genome
A network is tiny and ternary: up to 1,024 neurons, each in one of three states, each reading 3 others through a table of 27 rules. No weights trained by gradient descent, no backpropagation.
- 3
The search
Miners mutate networks and score them. Millions are tried every week; this is the useful work Qubic mining does instead of burning hashes.
- 4
The colony
Ant colony mining: a miner may publish only a network that strictly beats the one it grew from. Good structure is inherited instead of thrown away.
- 5
The audit
Scores are recomputed independently by the network, and the top 676 become the computors that run Qubic. Anyone can re-run any solution: the Jetski ANN Explorer archives them.
- 6
The live hour
From epoch 240 the computors agree on the real price of the past hour as a quorum, the Oracle Machine, and the network calls the next one: up, down, or not sure.
Three outputs, not two: the network may say it does not know. At the All-Hands of 17 September 2026 the core team described the live step in the same terms: the colony predicts hourly candles “with three possible outputs: up, down, or I don't know”, and the oracle replaces “a past year's frozen data” with live prices. The device itself, step by step, is on the BPP-9000 page; the live hours are in the terminal.
3. Why it is a test, not a promise
A year of past prices can be memorised. A network that learned the year by heart scores perfectly on it and still knows nothing about the next hour. Machine learning calls this the gap between training and generalisation, and it is the exact problem Aigarth claims to address.
- The homework and the exam. Brain Watch re-runs the epoch's best network with the core's rule and grades it twice: on the year it trained on, and on hours it never saw. Only the second number says anything about the future.
- The coin line. Up or down is close to a coin flip, so 50% is the floor, not a score. The record shows how likely a fair coin would be to do as well over the same number of hours.
- Nothing afterwards. Each call is sealed into a hash chain while its hour is still open, and checked against the close. Nothing is back-filled, and anyone can recompute every hash.
Our reading: if the paradigm generalises, it shows here as accuracy above the coin line that holds over many live hours, not over a lucky day. If it does not, the record shows that too. Either outcome is worth knowing.
4. Live calibration: from a year of history to hours never seen
Every network is graded on a window of one year: 8,760 hours of bitcoin, each up or down by the core's rule. Until epoch 240 that year was frozen in the past. From epoch 240 (18 November 2026, 12:00 UTC) the core takes the live price: five minutes after each hour the Price oracle brings the hour's BTC/USDT price from Binance, the computors agree on it, and the hour joins the window. The window slides forward, up to a week (168 hours), and every network is graded on hours it has never seen.
- The year, from the core itself. Epoch 240 starts from the bootstrap file in qubic/core v1.313.0: 8,929 hourly BTC/USDT prices in a row, 11 Nov 2025 08:00 to 18 Nov 2026 08:00 UTC, none filled. Its KangarooTwelve hash matches BTC_PRICE_BOOTSTRAP_HASH in the code, so every node starts from the same year. In the first frame 4,370 hours went up and 4,390 down: close to a coin flip, so the bar below is a real one. From then on the oracle adds every live hour.
- The bar. A network counts only if it gets no more than 3,504 of the 8,760 hours wrong: right on at least 60%. From epoch 240 the bar moves from 3,942 to 3,504.
- A missing hour. If an hour's price cannot be fetched within 24 hours, the previous price is used and the hour is flagged. Nothing is invented.
- No score yet, on purpose. In this stage the network calibrates; it makes no public call. Accuracy appears with the first public calls, and even then the first weeks of any live score say little. What is shown now is what is true now: how many live hours there are, and which way each went.
- Trained on historyA frozen year of past hours
- Calibrating liveFrom 18 Nov 2026: hours never seen
- Public callsOne per computor, every hour
- In the fieldThe calls, used
Our reading: this is the step that makes the test honest. A year of the past can be learned by heart; an hour that has not happened yet cannot. Watch the live hours arrive in the terminal.
5. Neuraxon: the scientists' brain model
David Vivancos
Qubic scientific advisor since 2024; AI researcher and entrepreneur since 1995; founder of MindBigData, an open dataset of brain signals for machine learning; author of several books on AI. @VivancosDavid
Jose Sánchez García, PhD
Researcher and lecturer in cognitive and affective neuroscience at Universidad Internacional de La Rioja (UNIR); peer-reviewed work on EEG and social emotions. @josesanchezhb
Neuraxon is their model of a neuron that behaves more like a living one. Each unit answers in three ways: +1 excite, −1 inhibit, or 0, a neutral state that keeps integrating its input without firing. It runs in continuous time, its synapses change while it runs, and slow channels that play the role of neuromodulators shift its thresholds. The work is open: code, demos and datasets.
- 2024: Human and Artificial Intelligence: Toward an AGI with Aigarth, the first Qubic science paper.
- November 2025: the Neuraxon blueprint; February 2026: Neuraxon 2.0.
- 2026 venues: ICMLT in Berlin (best paper of its session), AGI-26 in San Francisco (Springer LNAI), Artificial Evolution 2026 in Nice. Awards are not replication, and the authors present the results as early.
- ARC-AGI-3, the interactive reasoning benchmark: a final Kaggle score of 0.41 on a 0–100 scale, with no language model and no GPU. Small in absolute terms; the point is the constraint. 155,192 evolved brains from that work are public.
6. Two tracks, not one
Aigarth and Neuraxon are often mixed up. They are separate lines of work: Aigarth is the engineering track Come-from-Beyond leads; Neuraxon is the academic track of Vivancos and Sánchez. On 13 November 2026 Sánchez published a short video that sets them side by side, and its slides say it best:
“64 neurons. Each one is 0, 1 or unknown. Each neuron reads 3 others. Fixed network. It improves by search.”
“Events: +1, −1 or nothing. A live signal flows in. It predicts, then learns from the error.”
What they already share, in the same slides: “Ternary code”, “Local wiring: no unit sees the whole network”, “Results anyone can re-run and verify”. And the summary: “Aigarth searches for networks that work. Neuraxon 3.0 aims for a network that learns by itself.” (Jose Sánchez, X, 13 November 2026)
7. The merge: what is on record, and what we think
What has been said in public about bringing Neuraxon into the network:
- Come-from-Beyond, asked “Only start with p3 or other applications as well?”: “P3 and Neuraxon after polishing”. (Qubic Discord, 22 August 2026)
- The core team, All-Hands of 17 September 2026: “the dev responsible for algorithm coding is translating the NeurAxon research into something the network can run as useful proof of work, alongside the existing BPP 9000”.
- Qubic Science's Q4 2026 proposal lists “python code to teach and mix Neuraxon Aigarth in the Qubic Network, integrated in the ant-hill state in sync with CFB”.
- Sánchez's video ends with the next step: “smaller networks, closer to Aigarth's size.” (X, 13 November 2026)
Our reflection
The two tracks answer different halves of one old question. Evolution is slow and blind but keeps what works across generations; learning is fast but lives only inside one lifetime. In 1987 Hinton and Nowlan showed, in a small and now classic model, that the two together find solutions that neither finds alone: learning smooths the landscape that evolution has to search. This is often called the Baldwin effect.
Read that way, a merge has a natural shape:
- Mining finds the brain. Aigarth's search keeps doing what it does well: structure, wiring, a genome that beats its parent.
- Neuraxon lets it learn on the live hour. The found network keeps adjusting to the stream of real hours instead of staying frozen until the next epoch.
- The record keeps it honest. A network that changes while it runs is harder to audit than a fixed one. For BPP to stay verifiable, its state would have to be sealed with every call, so anyone can replay how it learned, hour by hour.
The risks are as real as the promise. A learner on a noisy live price can chase noise as easily as signal; a fixed network can at least be tested once and for all. Neuraxon's own results so far are early and modest, and its authors say so. None of this is announced as a plan by Come-from-Beyond or the core team; it is our reading of what each side has said.
8. What would count as evidence
- Live accuracy above the coin line, held over hundreds of hours, with a small chance that a fair coin would match it.
- A shrinking gap between the homework and the exam in Brain Watch, epoch after epoch.
- Independent re-runs: anyone downloading a brain and getting the same score.
- For Neuraxon: results replicated by people outside the project, and gains that survive on tasks it was not tuned for.
Until then the honest summary is short: a new way to grow AI, a narrow public test, and a record that writes down what happens. Watch it live in the terminal.
Sources
- qubic/core: BPP-9000 scoring (score_bpp9000.h) and the Oracle Machine price interface
- qubic/Qiner PR #29: the BPP-9000 algorithm update (mutation modes, rolling-frame scoring)
- Aigarth Intelligent Tissue, open source since July 2025
- Jetski ANN Explorer: every mined solution, re-runnable
- Sánchez & Vivancos (2024), Human and Artificial Intelligence: Toward an AGI with Aigarth, ResearchGate
- Neuraxon: code, papers and the Aigarth hybrid licence (MIT + no military use)
- Neuraxon interactive demos and datasets on Hugging Face
- Nxon2ArcAGI3Brains: 155,192 evolved brains for the ARC-AGI-3 games, CC BY-SA 4.0
- ARC Prize 2026, ARC-AGI-3
- Qubic Science Q4 2026 proposal
- Hinton & Nowlan (1987), How Learning Can Guide Evolution, Complex Systems 1: 495–502
- Come-from-Beyond on X
- David Vivancos on X
- Jose Sánchez on X
Discord quotes are from the public Qubic Discord, with their date and channel. Times on this site are UTC.