AGI, Astra, the table, and where Qubic is in it

What is AGI. Four definitions

One term, four definitions, and they don't match:

OpenAI, the charter: a highly autonomous system that outperforms humans at most economically valuable work. Definition through economy.

2. Chollet, author of ARC: intelligence is the efficiency of skill acquisition. Not "what it can do", but "how fast it learns something new with little experience". Definition through learning.

3. DeepMind, "Levels of AGI": a ladder of levels, from "emerging" to "superhuman", separately by depth and by breadth. Definition through a scale.

4. Helen Toner, ex-board of OpenAI: the term is "almost useless", because everyone puts his own meaning inside. Definition through refusing to define.

When OpenAI showed the Astra numbers on ARC, CfB wrote: the test was about necessity, but they sold it as sufficiency. Meaning: fail ARC and you are surely not AGI. Pass ARC and it doesn't mean AGI. These are two different statements, and the market mixes them on purpose.

More precisely, the marketing people mix them for you.

What the metrics in the table mean

ARC-AGI-3 is games. Interactive, no instructions, you have to discover the rules yourself. The key parameter here: adaptation.

Three sets: 25 public demo games (not counted in the score), 55 semi-private (for models via API), 55 fully private (Kaggle, offline).

The score is counted per level: how many actions a human spent, how many the agent spent, the ratio squared, weighted by level, averaged over games. The result is in percent.

Human is 100%. This is the reference point.

Now the rows.

Gemini 3.1 Pro at launch: 0.37%.

GPT 5.4 at launch: 0.26%.

Opus 4.6 Max at launch: 0.25%.

Grok-4.20: 0.00%.

This is the standard harness, no add-ons. This is how the frontier looked in summer: zero with a tail.

GPT-6 Astra, September 3, verified by ARC Prize: 62.7% in the standard harness, cost of the run 26,098 dollars (you feel it already?). With the provider adapter, meaning continuous memory between moves: 99.9% for 18,817 dollars.

Chollet called it a "step-function change" and wrote about 360 dollars per game.

ARC Prize, at the same time, wrote separately: we are not claiming this is AGI.

Kaggle track: 50 dollars compute budget, offline, open source. Same test, different weight category.

Where Qubic is in this system

Two rows, two different tasks.

First. BPP9000, epoch 229. Aigarth built a model of 64 neurons and 138 connections, trained it on 8,088 hours, and on one hour the model never saw, the ensemble of 676 best gave 55.3%. 359 out of 649.

The first forecast of the first Aigarth product turned out right! I ask to fix this in the history!

Second. Neuraxon, Vivancos and Sánchez, on ARC-AGI-3. On the demo games they reached level 7 in lp85, in r11l they spend 0.45 of human actions. On Kaggle the team "Qubic Nxon Vivancos & Sanchez": 0.25%, 67 entries, rank 1484.

Now put them side by side.

0.25% for Qubic. 0.25% for Opus 4.6 Max at launch. 0.26% for GPT 5.4.

The same level the frontier was showing in summer.

Only for the frontier it is a result on tens of thousands of GPUs, and for Neuraxon it is on 50 dollars, no GPU and no LLM.

Astra: 62.7% for 26,098 dollars.

The budget differs five hundred times. The place in the table is lower, for now.

And one more detail. Under Chollet's tweet about Astra, Vivancos put NoGPU/NoLLM/Neuraxon. This tags say a lot.

What is the difference of approaches

An LLM learns on everything people ever wrote and predicts the next token. More data and more GPUs, the better.

Aigarth doesn't read text. It grows a network. From noise, by selection, on the compute of miners. Not billions of neurons, but 64, and CfB is preparing the move to 128.

Last week he confirmed the mechanism: the network as a graph, degenerate and near-degenerate topologies get thrown out, the task is solved with fewer neurons, the scaling goes sublinear "because of graph isomorphism".

Astra talks. Aigarth grows.

With Astra you can't verify that it understood. Aigarth is verified by the score: 0 errors means the best model for the task.

And the last thing. About mining. Today.

On the AMA the team said it plainly: miners are priority number one.

CfB said it even plainer: "Now we depend on miners. All I can do is optimizations."

All the rows above, 55.3% on BPP9000, 128 neurons, the next task, they don't grow in a datacenter. They grow on the cards of people who switched on the miner.

This is not about income. This is about who feeds the network while the market is looking at Astra.

Fire up the miner. Point your card to the pool.

Qubic is not in the table. It is in one of its rows, on the other side of the electricity bill.

JUST RUN MINER!