Image

QUBIC BLOG POST

Decentralized AGI on Qubic: Highlights From Qubic AGI Lab Basel 2026

Written by

The Qubic Team

The Qubic Team

Published:

Listen to this blog post

Image

Qubic AGI Lab held its first public event, Decentralized AGI on Qubic: Live in Basel, on 7 November 2026 at Escape Bar in Basel. The free, day-long program was open to the community, developers, researchers and anyone curious, with talks in English and German. One question ran through the day: who builds an intelligence, who grades it and who can switch it off?

Dr. José Sánchez García opened with a keynote on Aigarth and Neuraxon. joetom, Qubic's Core Tech Lead, explained what decentralized AGI means for the network. Computor operator frog-rabbit then showed how Useful Proof of Work keeps Qubic secure, including what a 51% attack would actually take. This recap covers the welcome and those three talks, which all have published slides.

Qubic AGI Lab: A Swiss Association for Decentralized AI

The welcome slides introduced the host. Qubic AGI Lab was founded in August 2026 and entered the Swiss commercial register in September 2026. It works on decentralized AI on Qubic and is independent from the Qubic core team. Its slides put the line plainly: no tokens, no investment, no listings and nothing price-related.

The association's work rests on five pillars:

  • Events: Basel was the kickoff, followed by workshops and about one bigger event a year.

  • Research: building bridges to universities, so Qubic, Useful Proof of Work and Aigarth get serious scientific attention.

  • Education: neutral, clear information on decentralized AI, including for people outside the crypto bubble.

  • Open source: open tooling and docs that connect people in Switzerland and abroad.

  • Legal home: a Swiss non-profit that universities and partners can work with directly.

For 2027, the Lab plans four workshops online and in person, a research consortium and proposal under Horizon Europe, an ambassador network for partner introductions, and one or two docs or tooling projects to grow its GitHub organization. It also plans explainers on Qubic, Useful Proof of Work and Aigarth, starting with putting the Basel talks online, and one bigger event for the whole community. Active membership costs CHF 100 a year and includes a vote, and supporting membership costs CHF 50.

[IMAGE PLACEHOLDER: Basel event photo] Alt text: Qubic AGI Lab decentralized AI event at Escape Bar, Basel, Switzerland, 7 November 2026

Keynote: Aigarth, Growing Intelligence Instead of Training It

Dr. José Sánchez García is a neuroscientist, Qubic Scientific Advisor and co-inventor of Neuraxon. He said that meeting people in person matters to him, because email and Discord are not the same, and that he would come back if Basel ran again next year. His keynote worked through five questions: what AI, AGI and ASI mean, how a large language model works, why he fears superintelligence less than concentration, where Neuraxon really stands, and why intelligence should be decentralized.

AI vs AGI vs ASI: Three Words That Mean Less the More They Are Used

José started with working definitions:

  • AI performs a task we used to think needed a mind, such as chess, translation, protein folding or writing code. It is narrow by construction, and each skill is bought separately, with data.

  • AGI is the ability to acquire any skill a human can, as efficiently as a human can. This is François Chollet's definition from On the Measure of Intelligence, which the ARC Prize uses. José stressed the word efficiently: if it took him three years to learn to lift a glass to his mouth, that would not be intelligence.

  • ASI is a hypothetical system that beats all humans at almost everything. It is a forecast, not a measurement. José finds it hard to imagine, because much of what people know is local and private, down to how much salt someone likes in their macaroni.

He then compared the three ways AI is measured today:

  • Knowledge and exams, such as MMLU, bar exams and maths olympiads, reward things already written down. The weakness: the answers are on the internet the model read.

  • Skill acquisition, such as ARC-AGI-1, 2 and 3, rewards learning from little experience. The weakness: these are closed, deterministic 2D worlds, far from the open world.

  • Autonomy over time, such as the METR time horizon, rewards carrying a plan across hours. The weakness: METR says its suite no longer measures the longest horizons well.

From psychometrics, he reminded the room that a test is a sample of behaviour, not the trait itself. Goodhart's law applies: once a measure becomes the target, it stops measuring.

How AI Benchmarks Get Oversold: GPT-6 Astra on ARC-AGI-3

OpenAI launched GPT-6 Astra on 3 September 2026 and framed it as the start of an "AGI era." The same day, the ARC Prize Foundation published its own results on ARC-AGI-3, a set of game-like puzzles where the player has to discover the rules by playing. Under the foundation's standard harness, Astra scored 62.7% at a cost of $26,098. With OpenAI's own adapter, which keeps the model's reasoning state between moves, the best configuration reached 99.9%. The foundation's verdict: "we are not claiming that it is AGI."

"The model did not change. The plumbing around it did," José's slide read. In its replays, Astra writes notes to itself, invents a shorthand for the board and keeps a list of inferred rules. José called this an exoskeleton: the learning happens in the notes and the code, outside the network.

He named four ways AI progress gets exaggerated:

  1. Report the best harness. The launch slides used 99.9%. The neutral protocol gave 62.7%. Both are true, but only one made the headline.

  2. Move the definition. His slides traced OpenAI's AGI definition from its 2018 charter, based on economic work, to the September "AGI era" claim.

  3. Hide the cost. One evaluation run cost between $17,000 and $50,000. Human participants were paid about $13 per game, and a human brain runs on about 20 watts.

  4. Let the seller grade the exam. A claim nobody can falsify is not a scientific claim.

He also listed what the people selling AI have said. Elon Musk predicted in April 2024 that AI would "probably be smarter than any single human next year." In January 2025, Sam Altman wrote that OpenAI was "now confident we know how to build AGI." On 7 September 2026, Jensen Huang said "AGI has arrived". Each of them, José noted, sells the thing being declared, which is a reason to ask for the control, not an accusation.

He was careful not to dismiss the progress. In March 2026, frontier models scored under 1% on ARC-AGI-3, and by September, 62.7%. "The progress is real. The vocabulary is not." Our earlier breakdown of what GPT-6 Astra actually proves covers the same results in more depth.

How Large Language Models Work, and Why a Frozen Network Is Not a Brain

A large language model does one job billions of times: it guesses the next word. It reads the internet, adjusts hundreds of billions of weights after every wrong guess, once, in a data centre, and is then frozen and shipped. Words become points in a space with thousands of dimensions, where "king" relates to "queen" much as "man" relates to "woman." Attention, the mechanism behind the 2017 Transformer paper, lets each word weigh every other word in the sentence. José called it astonishing, and also a model of how words follow words, not a model of the world.

A brain does two things a frozen network cannot:

  • It changes while it acts. Synapses strengthen or weaken with every signal. Practice makes a movement better, and years without practice mean relearning it.

  • It lives in time. A neuron is never off. A Transformer has no clock: it sees a sequence of tokens and resets between calls.

Scale fixed many things, he said, but not these two, because they are about architecture, not size.

Five Assumptions Behind the Superintelligence Story

José presented this part as his own view and said that David Vivancos, who is more computational, might not agree. He took on five assumptions that he said are repeated as if obvious, but none of which has been shown:

  • "AI is getting smarter than us, exponentially." Intelligence has many dimensions. A squirrel remembers thousands of hiding sites for years, and an elephant signals to others kilometres away.

  • "We will build a general-purpose mind." Humans are not general-purpose either. We are a bundle of specific, evolved skills.

  • "Human thought can be copied into silicon." In the brain, hardware and software are the same thing. Equivalent thinking needs an equivalent substrate.

  • "Intelligence can grow without limit." No physical quantity is infinite. Cells divide at their own pace, and digging a mine still takes time.

  • "A superintelligence will solve most problems." It may help find a protein or test a drug, but moral problems between people have no single right answer.

He pointed to biologist Michael Levin, whose keynote the team heard at AGI-26 in San Francisco, as evidence that some kind of intelligence appears in cells, not only in human brains. The forecast José finds more likely is not one mind above us but many different minds beside us.

He compared waiting for a saviour superintelligence to the cargo cults of Melanesia. During the Pacific war, islanders watched planes land and unload cargo, and some later built wooden runways and control towers and waited for the planes to come back.

The Real Risk: AI Centralization and AI Cybersecurity

"I don't fear a superintelligence," José said. What he fears is a monopoly: one data centre with over 100,000 GPUs in one place, and one company deciding who gets access, in what tier and at what price, while grading its own exam.

His clearest example was AI cybersecurity. OpenAI gave Astra its own "critical" cybersecurity rating, because the model can find and exploit unknown vulnerabilities on its own. Under OpenAI's Preparedness Framework, critical is the highest capability threshold. For José, the problem is that the same company is both producer and judge. A tool this capable, owned and evaluated by one organization, is a concentration of power whatever we call its intelligence. Qubic's case for decentralised AGI starts from the same worry.

Neuraxon 3.0: A Neuron That Is Never Off

The second half of the keynote turned to the Qubic approach: growing intelligence instead of training it. Training adjusts a frozen design. Growing lets a living design change itself while it acts, and lets evolution pick the designs that learn best.

A Neuraxon has four core properties:

  • Three states, not two. It can excite (+1), inhibit (-1) or stay neutral (0). José compared the neutral state to a boxer who isn't punching yet but is reading the opponent: very active, preparing the next move.

  • Continuous time. Its state decays and carries forward, and nothing resets between inputs. There is no separate training phase.

  • Always plastic. Synapses strengthen, weaken, form and die while the unit works.

  • Self-organized criticality. The network settles at a branching ratio of about 1, where each active unit triggers about one more. Too high and activity explodes. Too low and it dies out.

In the team's best-performing network, the hidden layer was 63.5% neutral, 18.3% excite and 18.2% inhibit. With fixed thresholds, the layer went 91% silent and the error doubled.

Neuraxon 3.0 also learns on four clocks at once: spike timing over milliseconds, a synaptic tag that waits seconds for a reward or surprise, plateau plasticity from six seconds before an event to four seconds after, and astrocytes setting a budget for change over minutes. Criticality control runs throughout. Each mechanism can be switched off on its own, so its contribution is measured rather than assumed.

What Neuraxon Has Shown, With Controls

José listed four results, each against controls:

  • Self-organized criticality. Trinary networks settle at the critical point without tuning, and binary controls do not. This result was peer-reviewed and presented at AMLDS in Osaka in 2026.

  • 22% lower error on a time-series task. 128 always-plastic Neuraxons beat every echo-state network of the same size on 10 of 10 seeds, with 768 synapses against about 7,100.

  • Three to eight times the short-term memory. A unit holds 30 to 40 frames of input, against 5 to 10 for matched reservoir controls.

  • 98.7% error reduction through synapses alone, in a controlled circuit. He called this a mechanism test, not a task result.

The key experiment freezes the reader that turns the network's activity into a prediction, then changes the target. If the error still falls, the learning happened in the synapses inside. As of 2 November:

What the synapse must do

Result

Respond to error, reader frozen

Shown: cancels a shift in 10 seconds

Credit the right synapse

Shown: 5 times better than random

Stay changed after the error stops

Shown: 91% kept after five minutes

Not store wrong credit

Shown: 20 of 20 runs

Correct a shift

Shown: 9 of 10 held-out seeds

Correct a rescaling

Partly: 71% to 83%, needs 91%

Learn and keep a flip

Partly: 78% to 80%, needs 98%

Do all of this in one synapse

Open

What Neuraxon Hasn't Shown Yet

"A science that only reports what worked is marketing," José's slide read, and he went through the gaps:

  • Lasting memory. The one-second memory turned out to be supplied by the incoming outcomes, not stored. When the outcomes stopped, recall dropped within five seconds. The team found this in October and is fixing it.

  • Learning agents. In the team's Game of Life, evolved brains controlled behaviour only through innate wiring. José admitted the setup is a little circular, because whoever designs the world also designs what gets rewarded.

  • ARC-AGI-3. The best offline Kaggle score is 0.41 out of 100, up from 0.13. Evolved brains explore and adapt within a game, but they do not infer goals or plan.

  • Speech recognition. Frozen and plastic networks tie at about 25% to 47% on a small spoken-word task, while a matched reservoir reaches 59%.

The Bottleneck: One Reader That Sees Everything

José explained the current limit with an orchestra. The hidden Neuraxons are the musicians, and the reader is the conductor who listens to all of them and turns their activity into one prediction. Today that reader watches all 1,024 units at once. It works, but it uses 83% of all computation and degrades beyond 1,024 units.

A decentralized network has nothing like it, because no single node can see every neuron. So the next step is a local read-out, more like a jazz session where small groups follow each other without a conductor. That decides whether Neuraxon can run spread across a network like Qubic. After the read-out come memory that holds, credit assignment across seconds, and finally goals and plans, the missing piece for ARC.

The team has also tightened its research rules: select with some seeds and report with others, compare every search with a random search of the same budget, and seal each protocol with a hash before the test data exist.

How Neuraxon Could Fit Into Aigarth and Useful Proof of Work

Today, Qubic miners search small ternary networks under BPP-9000. The networks have no inputs and don't change while they run, so they improve only when a miner mutates one and the child scores better. The quorum then re-scores each one bit for bit.

José offered an idea, not a plan. Keep the same ternary code and the same bit-for-bit verification, but add a stream of input and synapses that learn during the run. The same seed and the same stream would always give the same result, so any node could still check any other. Aigarth would have to accept inputs and networks that change while running. Neuraxon would have to drop its central reader. He was clear that none of this has been demonstrated.

Evolution already works this way on the network. Since epoch 228, ant colony mining lets a miner who improves a published result publish it, so the next miner starts from there instead of from zero.

What Neuraxon Could Be Used For: A Reflex That Learns

If it all works, José does not see Neuraxon becoming a chatbot. He sees a small system that watches one signal, predicts what comes next and adapts on the device when the signal changes. He gave examples:

  • a sports watch that adapts to the wearer's own heart rate and sleep, instead of comparing them to an average for their age

  • a factory motor whose vibration starts to drift as a part wears

  • smart energy meters, skin conductance under stress and brain rhythms during sleep

Language and images need far more capacity than this small, energy-saving design has. Its competitor, he said, is the neuromorphic chip, not the language model. None of these uses has been built yet.

Why Intelligence Should Be Decentralized

José closed with three arguments from before the computer age:

  • Knowledge is dispersed (Hayek, 1945). What a society knows lives in millions of heads, much of it local. Qubic's answer: compute from thousands of independent miners, so no one owns the training run.

  • Requisite variety (Ashby, 1956). A controller can only regulate a system as varied as itself. Qubic's answer: many small brains, not one giant network.

  • Goodhart's law (1984). Whoever builds and grades a system will optimize for the grade. Qubic's answer: open code, peer-reviewed papers, public benchmarks and results checked by a quorum of 676 Computors.

"Decentralization is not the absence of power," his slide said. It is power that is divided, visible and accountable, and he added that the standard he applied to AI labs applies to Qubic too. To switch off Qubic, he noted, 451 Computors would have to vote for it. His takeaway for the room: ask of every intelligence who built it, who graded it and who can switch it off.

[IMAGE PLACEHOLDER: keynote slide on self-organized criticality] Alt text: Neuraxon trinary neurons holding self-organized criticality at a branching ratio of 1, from the Aigarth keynote at Qubic AGI Lab Basel

What Decentralized AGI Means for Qubic

joetom (Roman Thommen), Qubic's Core Tech Lead, picked up the three questions José left the room with. His slides put it in one line: "José grew an AI that nobody owns. This is the network that makes 'nobody' possible."

  • Who built it? The network.

  • Who graded it? The quorum.

  • Who can switch it off? No one.

What Makes AGI Decentralized?

joetom set three conditions. A decentralized AGI has no single owner, so anyone can contribute compute without permission. It is verifiable by anyone, because results are public and can be re-checked. And it is grown by the network's own work, not one firm's data centre. Centralized AI is the opposite on all three: one company, closed weights, its own compute.

He was blunt about the rest of the field: most "decentralized AI" is centralized AI plus a token. Training is expensive, work has to be verified or it is just trust again, and contributors need incentives, not donations. Unless decentralization reaches the compute, the verification and the incentives, his slide said, "it's theatre."

One Engine That Secures the Chain and Grows the AI

Qubic's answer is that the proof of work securing the blockchain is the same work that grows the intelligence. joetom described the loop:

  1. Anyone mines by optimizing neural networks for Aigarth.

  2. The best scores win, and the top 676 earn Computor seats.

  3. Those 676 Computors form the quorum that finalizes the chain.

  4. The trained models come out as a public by-product that anyone can re-score from open data.

"Security and training are the same act," his slide read: selection by useful work, and security by a 676-member quorum. Our explainer on Qubic consensus, Useful Proof of Work and the 676 Computors covers the mechanics, and the Qubic docs on uPoW explain how Computors are ranked.

He listed what this gives that centralized AI structurally can't: it is open to anyone's compute, verifiable from public data, owned by the network rather than a company, funded by protocol issuance instead of venture subsidy, and unstoppable, with no data centre to shut down.

What Decentralized AGI Demands of the Protocol

joetom called this the engineering frontier and his team's job:

  • Deterministic verification. Results must be bit-exact on any hardware, with no floating point in consensus.

  • Scale without fees. AI agents need to transact on-chain without gas friction.

  • A task that evolves. The task has to get harder as intelligence grows, which makes it a matter for governance.

  • Verification in the tick. Useful work has to be re-checked within about one second.

He framed the stakes as the early internet's question again: a few owners, or a commons? With corporate AGI, one firm owns the model, the data and the price, and everyone else rents access. With network AGI, intelligence is grown, owned and verified by an open network, and no one can revoke it. Looking ahead, he pointed to tasks that get harder each epoch, AI agents that transact on-chain without fees, and the network as a single home for compute, consensus and intelligence.

Useful Proof of Work and Blockchain Security: How Qubic Elects Its Computors

frog-rabbit is an electrical and hardware engineer who has run a Computor since 2022 with QLI Pool. His talk, "Mining that learns something," started from a simple contrast. Bitcoin miners guess trillions of times per second and forget everything, and the only output is a hash. Can the work also be useful, and still secure the network? His motto for the answer: "In a decentralized network it doesn't matter who you are, only what you contribute."

Proof of Work vs Proof of Stake vs Useful Proof of Work

He compared who gets to decide on three networks, and what an attacker would need:

Network

An attack needs

The work also gives

Bitcoin (PoW)

Over 50% of hashrate

Nothing beyond security

Ethereum (PoS)

1/3 or 2/3 of stake

Locked capital

Qubic (uPoW)

1/3 or 2/3 of Computor spots, every week

AI research data

On Bitcoin, a lottery decides each block. On Ethereum, capital decides, as stake. On Qubic, 676 Computor spots are won by compute in a weekly election.

Bitcoin's security has rested on an honest majority of computing power since the Bitcoin whitepaper. On Qubic, the same compute also produces AI research data.

How Qubic Mining Works: Miner, Pool, Computor, Tick

  1. Miners run CPUs and GPUs that search for better AI networks. Any number can take part.

  2. Pools bundle miners under one identity and pay them out.

  3. Computors are the top 676 identities by best score. They run bare-metal nodes that validate transactions and run smart contracts, and only Computors receive protocol rewards.

  4. Ticks become final once 451 Computors agree, which is two-thirds plus one.

Each epoch lasts seven days and switches on Wednesday at 12:00 UTC. frog-rabbit said that switch used to mean panic and now means coffee.

Four Years of Useful Work: From Monero to Ant Colony Mining

Qubic has experimented with what counts as useful:

  • 2022 to 2025: random networks for Aigarth, ranked by the number of solutions found.

  • 2025 to April 2026: CPUs alternated between AI work and Monero. At times Qubic held more than 51% of Monero's hashrate.

  • April to September 2026: ASICs mined Dogecoin while CPUs and GPUs trained in parallel. This has since been wound down.

  • July 2026, epoch 224: BPP-9000, where only the best solution counts, not the quantity.

  • September 2026, epoch 228: ant colony mining, where solutions build on each other.

The lesson he took from Monero: compute is mobile, and that applies to attackers too.

How BPP-9000 Scores Mining Work

Under BPP-9000, miners look for a tiny network that "hums along" with the Bitcoin price. Since epoch 236, each network has 32 neurons (64 before), and each neuron reads three others. Neurons are ternary (0, 1 or unknown), and the network has no inputs: it runs freely, tick by tick. One neuron acts as a clock and one says up or down. The design builds on Stuart Kauffman's random Boolean networks from 1969, with a third state added.

The target is one year of hourly Bitcoin price direction, 8,760 labels from 1 July 2025. A network passes with 4,350 errors or fewer, where pure chance would be about 4,380. At 2,920 errors or fewer, two out of three hours right, the window shifts forward by up to 168 hours, and more shift scores better. An "unknown" at the wrong moment counts as an error.

In ant colony mining, each identity has a tree of solutions rooted in its public key and the epoch seed. Miners first wander, then climb, and a new network must strictly beat its parent. In a pool, all workers grow the same tree. Ant colony mining is a structure, not an algorithm, so BPP-9000 could be swapped out later.

Trust, but Verify: Every Node Re-computes

Scores are bit-exact and deterministic. Replaying a solution is cheap, and finding one is expensive, so every node re-checks. A submission carries the identity, the parent solution, a 32-byte nonce, an anchor tick and the claimed score. It also carries a deposit of 1 million QU, which is lost if the score doesn't hold up, and a separate rule catches attempts to reuse old solutions.

What Would a 51% Attack Take on Qubic?

frog-rabbit's slide asked what an attacker would need, assuming a share of mining compute wins the same share of the 676 Computor spots:

Computor spots held

What happens

68 (about 10% of compute)

Network runs normally

226

Enough to halt the network

451

Control of the network

And the attacker would have to win those spots not once, but every week, against all other pools.

He named five layers of protection for the Computor spots:

Layer

How it protects Qubic

Sybil resistance

Spots can't be created or bought. Each costs real work.

Byzantine fault tolerance

451 of 676 must agree, so up to 225 bad nodes are tolerated.

Weekly vote

No spot is permanent. Slack off and you're out.

Proof

Bit-exact replay and a 1 million QU deposit.

Bound to the identity

Solutions can't be copied or precomputed.

Our guide to quorum consensus explains why two-thirds plus one is the threshold. The talk focused on consensus attacks, but blockchain safety also depends on wallets. During the October Qubic wallet drain, a crypto hack exploited weakly generated seeds rather than the protocol, and the quorum coordinated a halt to freeze the stolen funds. Our guide on how to secure your Qubic wallet covers the user side.

What Useful Proof of Work Is Not (Yet)

frog-rabbit closed his technical part with an honest list:

  • Not a price oracle. The networks reproduce the past. They don't predict.

  • Overfitting is a risk, because many variants run on the same data.

  • "Useful" keeps being redefined, so scores are hard to compare across epochs.

  • Compute can be rented. Monero proved that, in both directions.

The value, he said, is in the experiment: an open lab with 676 participants every week.

What's Next for Qubic Mining

  • Oracle calls: the quorum signs requests to external machines, and a prototype is live.

  • Neuraxon: the best offline ARC-AGI-3 score, now 0.41 out of 100, came from a search on about 500 CPU cores with no GPU. The outlook is neural architecture search as Useful Proof of Work.

  • Getting involved: one CPU and one pool are enough to start mining.

[IMAGE PLACEHOLDER: slide comparing Bitcoin, Ethereum and Qubic consensus] Alt text: Proof of Work vs Proof of Stake vs Useful Proof of Work, comparing 51% attack costs on Bitcoin, Ethereum and Qubic

Also on the Program: Tooling, the Panel and Open Q&A

The afternoon continued with mio, a software developer working on APIs, on "Building on Qubic: Tooling & APIs for the Ecosystem." José and joetom then joined a panel on where decentralized AGI is heading, followed by an open discussion and Q&A with the speakers and an apéro. No slides were published for these sessions, so they aren't covered here.

Where to Learn More About Qubic AGI Lab

The Basel event page has the slides for the welcome and the three talks covered here, and the Lab plans to put the talks online. Membership, workshops and project ideas go through qubic-agi-lab.ch and the Lab's Discord. José's Aigarth and Neuraxon papers and code are on qubic.org/science, and the protocol, consensus and smart contract docs are at docs.qubic.org. For more on the research, the Qubic All-Hands recap covers September's Neuraxon and ARC-AGI-3 updates.

Figures in this recap come from the speakers' slides and the keynote recording. Qubic AGI Lab is an association independent from the Qubic network and core team. This recap explains the research and the technology and is not financial advice.

© 2026 Qubic.

Qubic is a decentralized, open-source network for experimental technology. Nothing on this site should be construed as investment, legal, or financial advice. Qubic does not offer securities, and participation in the network may involve risks. Users are responsible for complying with local regulations. Please consult legal and financial professionals before engaging with the platform.

© 2026 Qubic.

Qubic is a decentralized, open-source network for experimental technology. Nothing on this site should be construed as investment, legal, or financial advice. Qubic does not offer securities, and participation in the network may involve risks. Users are responsible for complying with local regulations. Please consult legal and financial professionals before engaging with the platform.