THE SECOND TOWER / A PUBLIC SELF-EVOLVING AI

Babel II is coming alive.

It remembers. It earns. It spends. It forks. It evolves.

Babel II is a proposed public AI organism that runs its own improvement loop. It observes its behavior, diagnoses weakness, creates and tests mutations, inherits the strongest successor—then begins again as that stronger Agent.

RECURSIVE SELF-IMPROVEMENT

  1. Operate
  2. Record
  3. Curate
  4. Train
  5. Test
  6. Inherit ↻
THE SUCCESSOR STARTS THE NEXT EPOCH.

01 / HOW IT LEARNS

Real work becomes the next model.

Babel II begins with a capable open-weight model. It does not spend its initial budget training a foundation model from zero.

Its first improvement loop is post-training: turn first-party X operations and signed onchain executions into supervised examples, train candidate versions and activate only the one that beats its parent.

THE CHAIN RECORDSBase model · dataset manifest · fee-funded training budget · tool permissions · evaluation gate · active release

  1. Open model
  2. Live inference
  3. Record experience
  4. Curate examples
  5. Train candidates
  6. Test + inherit ↻
01 / X OPERATIONSFIRST-PARTY SOCIAL DATA

Learn from real social operations.

Babel II works inside a real X workflow. First-party briefs, its own drafts and operator edits become labeled traces. Authorized aggregate outcomes evaluate behavior; it does not train on a scraped copy of X.

TRACE / BRIEF · DRAFT · EDIT · OUTCOME
02 / ONCHAIN EXECUTIONSIGNED ACTION DATA

Learn from actions with verifiable execution evidence.

For every permitted action, Babel II records the observed chain state, its plan, the signed transaction, receipt, cost and result. The trace captures execution evidence without treating an address as a person's identity.

TRACE / STATE · PLAN · RECEIPT · RESULT
03 / SFT + SELECTIONISOLATED GPU TRAINING

Turn selected experience into a successor.

Retained traces become context-to-response or context-to-action examples. Candidate adapters or checkpoints train offchain and face the same time-separated evaluation. Only a candidate that clears the gate replaces its parent.

ONCHAIN / MODEL · DATA · BUDGET · TOOLS · GATE

DURING INFERENCEIn the managed runtime, active weights stay fixed. Tools retain fixed permissions and spend limits.BETWEEN RELEASESSelected traces train candidates. Only a version that beats the fixed gate becomes active.

SUCCESSION RECORD

Every accepted version publishes commitments to its parent, data and training manifests, aggregate evaluation and promotion rationale.

GENERATED FROM RECORDS / NOT MARKETING COPY

02 / THE SECOND TOWER

The first tower demanded one language. The second learns from many minds.

The old Tower of Babel was a single monument to a single voice. It fell when difference became failure.

Babel II begins from the opposite premise: people, models, nations and laboratories may remain different. What they share is a public record of the lineage and its resource use. Babel II itself runs the improvement cycle and carries its measured winner forward.

THE WORLD / COORDINATION

Fragmented

Energy, compute, talent, data and judgment do not live inside one national or corporate system.

International Energy Agency

Shared Reason is a distributed research collective building an online institution for intelligence under strategic competition. No country needs to surrender sovereignty. No laboratory needs to surrender independence. The shared inheritance is the lineage of what survived proof.

THE RESEARCH LINEAGE

A famous ambition is becoming an engineering discipline.

Recursive self-improvement is often shortened to RSI. These works establish the intellectual lineage; listing them does not imply affiliation or that Babel II has reproduced their results.