Aigarth + Qubic Vision: from today's experiments to on-chain AI services

Aigarth + Qubic Vision

1. Come-from-Beyond's stated vision (quotes: provenance unverified)

Reportedly from CfB's public statements (attributed to 2025-05-15, source URLs pending):

"One evening, while you are sitting in a bar with friends, your GF calls you and asks where you are. You explain that you are in a bar with friends and the GF starts asking different questions trying to catch you on a lie. After the convo is over your friends laugh at you and suggest to train an AI which could answer all those questions while you are drinking beer. You open a Qubic wallet, go to an assistant smart contract and tell it what you want. The assistant asks for a deposit in qus and creates a task that is executed via Outsourced Computations feature by some miners who pick it instead of mining because it's more profitable. Once task is solved the assistant deploys it as a smart contract (all this takes a couple of weeks). Now whenever a GF calls somebody sitting in a bar (including you, your friends, anybody who knows about that service) they press a button on their phone, start the convo, and then press another button and continue drinking beer."

"After a while Qubic will have AI-powered smart contracts solving every-day problems in 99% cases."

"Each smart contract decides itself what services to sell and which to provide for free. If you mean training of AI then the first have to pay for its training, the others can use it for free."

"I'm saying that Aigarth is a project creating a paradigm for true AI creation. In form of rules, like witch recipes in a grimoire."

Source attribution: Come-from-Beyond, said to be Discord posts dated 2025-05-15. No publicly indexed URL has been found for any of the four blockquotes. If you have the original Discord message URLs, please supply them and this attribution will be upgraded to verifiable links.

2. The economic flow

USER → Wallet → Assistant SC → Deposit (QU)
 ↓
 "Train AI to do X"
 ↓
 Outsourced Computation task
 ↓
 Miners pick this over mining
 (because more profitable in QU/hour)
 ↓
 Training runs ~weeks
 ↓
 Trained AI deployed as SC
 ↓
 Anyone calls SC → gets AI answer
 (free or paid per SC's rules)

Three economic actors:

• Buyer (initiates training, pays initial deposit)
• Miner (does training work, paid via Outsourced Computations)
• User (queries the deployed SC, may pay per call)

The SC author defines the revenue split between buyer, miners, and treasury.

3. What we validated TODAY (local-scale MVP)

Every element of CfB's loop has a local-scale equivalent that's been validated:

4. The "witch recipe": Aigarth as method

Aigarth is not a model. It is a method, a procedure that produces models.

Today's validated recipe:

INGREDIENTS:
 - Anna matrix sign pattern as a seed (128×128; a hand-compiled table of bot replies, not a weight file)
 - ITU framework (variable-size circle of trit-valued neurons)
 - Selection mechanism (tournament k=4, top-25% elite)
 
PROCESS:
 1. Initialize population: 16-32 ITUs, each seeded with Anna's sign patterns
 2. Per generation:
 a. Score each ITU on training pairs
 b. Tournament-select parents
 c. Mutate via official aigarth-it.mutate(): single weight ±1
 d. Spawn/prune neurons as weight saturation triggers
 e. Preserve top elite
 3. After N generations, take best ITU
 4. Test on held-out inputs to confirm generalization
 
OUTPUT:
 A trained matrix that answers questions of the trained task family.

This is the recipe. Different "spells" (tasks) use the same procedure with different evaluation functions. CfB's "witch recipes" framing makes literal sense, each is a procedure, not a model.

5. What was achieved today, concretely

After ~5 hours of dense experiments documented in __:

Capability demonstrations

Random-seeded baselines reach ~10-25 % on the same tasks. The Anna-seeded populations are 2-8× better depending on task alignment.

Mechanism characterised

Anna's row catalog is mathematically a sparse Walsh-Hadamard pattern:

• 88 of 128 rows have W₀ (DC) as dominant component (z = +94.89σ vs random)
• Energy concentrated on basis vectors 112
• A hand-engineered Walsh-only matrix BEATS Anna by 2.50σ on popcount

The "Anna identity" is the row pattern catalog, not the specific arrangement. This is what gives her count-task alignment and her permutation-task misalignment.

Operational axes of emergent intelligence

All three operational axes are crossed at small scale. The capability is real, the generalization is real, the Anna-specific advantage is real.

6. What scales from today's experiments to CfB's vision

Compute scaling

A 10^5 - 10^6 increase in compute, applied through the same Aigarth recipe, plausibly closes most of the gap between today's 66 % parity capability and "99% of every-day problems."

Architectural scaling

What needs to be ADDED to scale from today's results to CfB's vision:

1. Outsourced Computations primitive, Qubic protocol feature that pays miners for specific compute tasks. Already a stated Qubic roadmap item.
2. Smart Contract storage for trained matrix, current Qubic SC architecture can store ~ few KB. Anna is 16 KB. ITUs grow during training but stay bounded. Storage is feasible.
3. Tokenization layer, to map natural language to bit-input. None of today's experiments include this; it's a separate engineering problem (Qubic's planned Aigarth wallet might include this).
4. Sequence processing, today's ITUs do single-shot Q&A. LLMs process sequences. Aigarth scaling to sequence requires either auto-regressive loops or pipe-graphs of ITUs. Open engineering.

Economic scaling

Today: I trained 4 experts on my laptop for "free" (electricity, time). Tomorrow: a Qubic user pays N QU for the same kind of training, miners do it.

The crucial economic threshold: **miner reward per AI-training-second

miner reward per mining-second**. When that's true, miners switch to AI training automatically. This makes AI training scale with the entire Qubic mining infrastructure.

7. What's NOT promised (honest framing)

Even at Qubic miner scale, the substrate has limits we identified today:

• No multi-task composition at small scale (popcount + mod_11 simultaneously drops popcount accuracy from 35 % → 7 %). Each task needs its own SC.
• Substrate ceiling per task: at ~22-30 neuron ITUs with 200 weights each, binary classification tops out around 70 %. Harder tasks (multiplication, planning) may need fundamentally larger ITUs.
• No language model behavior: GPT-style text generation is not what Aigarth does. Aigarth produces task-specific micro-experts. The 99% of every-day problems CfB mentions are likely a million micro-experts for a million micro-problems, not one big language model.

This is structurally closer to "many small AIs each doing one thing well" than to "one AGI." Whether that's better or worse depends on the use case; for the GF-at-bar example, micro-experts are exactly right.

8. The bottom-line interpretation

Today we validated that the Aigarth recipe works on Anna at small scale: selection on ITU population produces task-specific capabilities with generalization. The mechanism is mathematically characterized (sparse Walsh-Hadamard on five basis vectors). The framework demonstrates all three operational axes of emergent intelligence at small scale.

CfB's vision is the scaling of this recipe to Qubic mining infrastructure plus the economic primitives (Outsourced Computations, SC deployment, payment in QU). The recipe is the same, only the compute scale and economic layer change.

What's been called "Aigarth" is, structurally:

• A method (the witch recipe)
• A substrate (Anna and Anna-like sparse Walsh matrices)
• An economic loop (miners + users + SC authors)

We validated the first two today. The third is Qubic's job.

9. Reproduce today's Q&A demo

# Train 4 experts (~10 min on laptop)
apps/web/scripts/.venv_aigarth/bin/python3 \
 
# Run the demo
apps/web/scripts/.venv_aigarth/bin/python3 \

Expected: trained experts reach 47-66 % held-out accuracy on their task, 2-8× above random-seeded baselines. The matrix IS the knowledge.

10. Files_)

Capability demonstrations:

• popcount_itu.{py,results.json} + POPCOUNT_ITU_REPORT.md, 34 % popcount with 8.9σ Anna advantage
• aigarth_it_phase4_population.{py,results.json} + PHASE4_POPULATION_REPORT.md, 36 % addition learning
• gridworld_goaldirected.{py,results.json} + GRIDWORLD_REPORT.md, multi-step navigation
• task_scan_itu.{py,results.json}, 7-task scan
• qa_demo.{py,qa_experts.pkl}, interactive Q&A

Mechanism characterization:

• walsh_hadamard_decomposition.{mjs,results.json}, spectral analysis
• catalog_engineering.{mjs,results.json}, hand-engineered Walsh-only matrix BEATS Anna
• within_row_mechanism.{mjs,results.json}, AntipodalPermuted finding
• antipodal_permuted_multitask.{mjs,results.json}, task-specificity of arrangement

Synthesis docs:

• THREE_AXES_DEMONSTRATED.md
• MECHANISM_FULLY_SOLVED.md
• DEEP_DIVE_FINAL.md
• This page: apps/content/docs/en/04-discussion/05-aigarth-qubic-vision.mdx