Qubic Open Science AMA Recap: August 20, 2026

TLDR

• David Vivancos and Dr. Jose Sanchez hosted the Qubic Open Science AMA on August 20, 2026, covering three accepted conference papers plus a fourth announced live on the call.

• Osaka put numbers behind the trinary neural network design. Across 5,582 experiments, trinary networks held the healthy operating range 88.8% of the time against 37.2% for binary, and the paper took a Best Oral Evening Presentation Award.

• Berlin introduced Neuraxon v2.0. San Francisco brought Multi-Neuraxon to AGI-26, the 19th Conference on Artificial General Intelligence, now in Springer's proceedings.

• A fourth paper has been accepted at the 17th International Conference on Artificial Evolution in Nice, France, 27-29/10 2026.

• Neuraxon brains score 0.25 on the offline Kaggle version of ARC-AGI-3, evolved on roughly 500 CPU cores with no GPUs involved.

• Neuraxon 3.0 is in development, and the architecture reaches the live Qubic network once the ant colony algorithm is proven.

The August 20, 2026 AMA was the first time the Qubic Open Science team laid out its full research year in one sitting. Vivancos and Sanchez traced the work from a 2024 preprint through three accepted papers, then announced a fourth on air. Here is what changed, in plain language.

What Is Aigarth? The Foundation Behind Qubic Open Science

Qubic's AI research starts from a claim borrowed from biology: intelligence grows rather than gets designed. Aigarth applies that to "intelligent tissue," a blank web of artificial neurons where small AIs sprout, the ones that solve problems survive, and the rest are discarded.

The 2024 paper by Sanchez and Vivancos set the groundwork. A century of psychology keeps rediscovering one underlying ability behind almost every mental skill, the factor researchers call g. Separately, the brain constantly predicts what comes next and corrects itself when the guess turns out wrong, an idea associated with Carl Friston and his free-energy principle. The team attended his lecture in San Francisco.

That leads to a clean break from how mainstream AI learns. Backpropagation pushes a global error signal backwards through the whole network and splits life into a training phase and a use phase. A real synapse has access to neither. It knows only what is happening around it, and it learns while the brain is busy doing everything else. Neuraxon took that as its design constraint: every synapse uses local information only.

The Three Neuraxon Papers Accepted in 2026

Osaka: Why a Trinary Neural Network Beats Binary

Ordinary computing thinks in two values. A trinary neural network adds a third, so a neuron can fire, actively inhibit, or rest at zero. Think of that zero as a shock absorber: a binary neuron has to commit on every tick, so a small disturbance immediately becomes a decision that spreads.

The paper measured branching ratio, written σ, a standard tool in the study of brain criticality. Picture a rumour in a crowd. If each person tells slightly fewer than one other person, it dies out. Slightly more than one, and it becomes a shouting match. Right around one, it keeps travelling without exploding, which is roughly where brains sit.

The trinary architecture climbed into that band and stayed, with no resets and no supervision. Binary settled below it at σ ≈ 0.82. Three states cost 1.58 bits per neuron (log₂3), the same figure now driving 1.58-bit language model research at Microsoft and elsewhere, which arrived in similar territory from a completely different direction. The paper won Best Oral evening Presentation at IEEE AMLDS 2026.

Berlin: Neuraxon v2.0 Makes the Synapse the Computer

In most artificial networks a synapse is a single number. In Neuraxon it carries three weights running in parallel at three speeds, from a fast ionotropic channel measured in milliseconds through an NMDA-like channel to a metabotropic channel operating across miliseconds, seconds and minutes.

Timing carries the rest, through a local rule known as spike-timing-dependent plasticity. If neuron A fires just before neuron B, A might have caused B, so the connection strengthens; if A arrives late, it weakens. That rule creates a problem it cannot solve alone, since every reinforcement makes the same coincidence more likely next time. Anyone who has held a microphone too close to a speaker knows how this ends.

Most of what v2.0 adds exists to stop that runaway. The team also reported the uncomfortable finding: only three of the seven added mechanisms proved necessary under the tested protocol, because runs lasted around 80 steps and the slower regulatory loops never got a chance to act. Jose named this as a target for the next round of testing. David noted on the call that the Berlin presentation drew recognition in its evening session. Reference code is open under MIT as neuraxon2.py.

San Francisco: Multi-Neuraxon and Emergent Specialization

The Multi-Neuraxon model connects four Spheres: VIS for vision, AUD for hearing, ASC as the associative hub, and MTR for motor output. Each is a full Neuraxon v2.0 network, and they communicate through frequency gating, so a channel opens only when the rhythms of sender and receiver line up. Two radios tuned to the same station.

Across fourteen tests against a matched Nengo baseline, specialization appeared without anyone assigning it, and the associative hub showed a 2.88× multimodal gain. With the cue removed, Multi-Neuraxon sustained information flow at 1.96 while the baseline collapsed to 0.42. The lesion result is the one worth remembering: lose three quarters of an orchestra and the melody still comes through.

One finding cut against intuition. Randomizing the phase raised raw information throughput while motor stability fell, which is why the team treats coordination as a product of structured communication.

A Fourth Paper Accepted at Artificial Evolution 2026

Jose delivered the news live. A fourth paper has been accepted at the 17th International Conference on Artificial Evolution, a biennial event in Nice, France on 27-29/10 2026. It comes out of the Neuraxon Game of Life work, where brains live in a simulated world and their traits pass to the next generation. Four accepted papers in a single year gives the project something crypto research rarely has: outside scientists engaging directly with the claims.

Neuraxon on ARC-AGI-3: A 0.25 Score on CPUs Alone

ARC-AGI-3 is the hardest interactive reasoning benchmark currently running, and no frontier model has beaten it. The Kaggle version is strict. Brains play offline with no internet access and no prior knowledge of the games, which rules out leaning on an LLM.

Qubic sits at 0.25, up from 0.18 in early July. The leading score across the field moved to 3% in the week before the AMA, among close to 2,300 contenders, placing the team a little past the middle of the pack. Reaching that took over 110 search architectures and millions of evolved brains tested offline against the 25 training games. The current generation plays at least 15 of them, with strong performance on around eight.

All of it runs on roughly 500 CPU cores over a couple of months. GPUs are permitted and the team has them, along with a CUDA library for Neuraxon, but they held to CPU-only to keep the scalability argument intact. Far more CPUs exist in the world than GPUs, and a laptop can join.

Holding a score is harder than posting one. A Neuraxon brain never stops changing, and the games shift slightly on each restart, so a brain that performs well across twenty runs may stumble on the next six. When the competition closes, 55 unseen games get released, which will separate real generalization from anything tuned to the public set. The evolved brains and their datasets go on GitHub once that happens.

Neuraxon 3.0 and the Qubic Open Science Roadmap

Neuraxon 3.0 addresses the problem that keeps recurring: a brain has to learn on several clocks at once. Milliseconds for which signal arrived first. Seconds for a reward reaching back to flag what mattered. Minutes for rebalancing how loud each input is. Hours for growing or pruning connections. Most artificial neurons handle only the fastest of these.

The design gathers proposals from five separate rules and hands the decision to a single criticality dial. The unit measures whether its own activity is fading or running away. Near the balance point, learning gets through. The further it drifts, the more learning is damped and the excitation/inhibition mix nudged back. Every proposal is summed and then clipped, so no single connection can escape. Astrocytes are under evaluation as well.

From the Q&A: Aigarth Integration and the CPU Question

Is Neuraxon already in Aigarth? The evolutionary component is embedded in the architecture. Live deployment waits on the ant colony algorithm being proven, which David expected online roughly a week after the AMA, with input constraints and network capacity still being worked out.

Can discoveries spread across the network? Yes, per David. Rewards can be embedded and successful configurations can seed later generations, which fits Aigarth's survival-of-the-fittest core. Jose added that the miner producing a winning solution gets rewarded, and that the mechanism still needs testing.

Why CPUs when everyone else uses GPUs? GPUs exist for tensor math, a capability the gaming industry paid for, and they hit limits on VRAM cost and parallelization. CPUs remain far more numerous, and some operations benefit from their clock rates. Qubic is not anti-GPU, and the CUDA library lets Neuraxon run on GPU where that makes sense.

Closing: Where Qubic Open Science Goes Next

A year ago this was a set of ideas in a preprint. It is now four papers accepted at international conferences, an open codebase under MIT, a live simulated world running around the clock, and a competitive entry in the hardest reasoning benchmark in AI. None of it required a frontier data center.

The next markers are already dated. Nice in October. The ARC-AGI-3 close. Neuraxon 3.0. The ant colony going live, which opens the door to running this work on Qubic's own compute. Everything stays open, which is the point of Qubic Open Science.

Links and resources

• Neuraxon code and papers: github.com/DavidVivancos/Neuraxon

• ARC-AGI-3 benchmark: arcprize.org/arc-agi/3

• Artificial Evolution 2026, Nice: ea2026.inria.fr

• AGI-26 proceedings: Springer, Artificial General Intelligence, 19th International Conference