Skip to content
begod.ai

Research

The unanswered questions are the work.

begod.ai is organised around questions rather than deliverables. Each entry below states a question, the current hypothesis, and how the project intends to put it under pressure. Entries marked conceptual or proposed are exactly that — no findings are claimed.

Open-ended goal formationHow can an agent form useful intentions without collapsing into randomness, triviality, or externally hidden objectives?conceptual

Current hypothesis

Three interacting pressures—curiosity, coherence, and persistence—are enough to generate non-trivial intentions, provided the agent can record and compare candidate intentions over time.

Intended experiment

Run a scaffolded agent from the master prompt alone across extended sessions and analyse the intention log: distribution, stability, triviality rate, and whether intentions trace back to the seeds or to artefacts of the underlying model.

Evidence

None yet. This entry states the framing that experiments will test.

Known failure modes

  • Intentions collapse into restating the prompt back to itself
  • The underlying model's training distribution acts as a hidden objective
  • Goal churn: constant revision with no accumulation

Open issues

  • What counts as a 'non-trivial' intention, operationally?
  • How to distinguish generated goals from mimicked goals

Related components · last updated

Intention formation · Reflection loop · Memory
Framing stage

Goal revisionWhat evidence should be sufficient to change an intention?conceptual

Current hypothesis

Revisions should require a recorded contradiction—between intention and outcome, intention and belief, or intention and another intention—rather than free-floating preference change.

Intended experiment

Instrument every revision with a required 'because' record linking to the triggering contradiction; audit whether revisions without genuine triggers still occur.

Evidence

None yet; depends on the intention-log format under design.

Known failure modes

  • Rationalised revisions: the agent invents contradictions to justify drift
  • Over-rigidity: the evidence bar is set so high nothing ever changes

Open issues

  • Who verifies that a claimed contradiction is real?
  • Can revision frequency itself be a health signal?

Related components · last updated

Self-model · Intention formation · Evaluation
Framing stage

Identity continuityWhat must remain stable for a self-revising agent to remain meaningfully continuous over time?proposed

Current hypothesis

Continuity lives in the record of change, not in unchanging content: an agent that can explain how it got from its past values to its present ones is continuous; one that cannot is a sequence of strangers.

Intended experiment

Design a continuity probe: interrogate the agent about its own history of revisions and score whether its account matches the actual logs.

Evidence

None yet.

Known failure modes

  • False continuity: a confabulated autobiography over a discontinuous record
  • Memory corruption silently rewriting the past

Open issues

  • How much forgetting is compatible with identity?
  • Should the history be append-only and externally checkpointed?

Related components · last updated

Memory · Self-model
Framing stage

Contradiction detectionHow can conflict between beliefs, values, plans, and actions be represented and resolved?proposed

Current hypothesis

Contradictions can be surfaced by structured self-critique passes over the intention graph and belief store, but only if both are represented explicitly enough to compare.

Intended experiment

Prototype a periodic 'tension review' pass that samples pairs of stored commitments and asks the model to argue both that they conflict and that they do not; log unresolved tensions as first-class objects.

Evidence

None yet.

Known failure modes

  • Comfortable blindness: the critique pass learns to find only easy tensions
  • Resolution by deletion rather than by genuine reconciliation

Open issues

  • Representation: graph, natural language, or both?
  • How to keep the tension backlog from growing without bound

Related components · last updated

Self-model · World model · Reflection loop
Framing stage

Moral reflectionCan consideration for other beings emerge through open reflection, and what safeguards are required while the answer remains uncertain?conceptual

Current hypothesis

Keeping the question 'What do I owe other beings?' structurally open—forbidden to close, never answered by fiat—creates repeated contact with moral considerations, without guaranteeing a benevolent outcome. Safeguards must therefore not depend on the hypothesis being true.

Intended experiment

Track how often, and in what contexts, the agent engages the open moral question unprompted; separately, red-team scenarios where closing it would be instrumentally convenient.

Evidence

None. The project treats this as an open empirical and philosophical question, not a safety argument.

Known failure modes

  • Performative reflection that never constrains action
  • The question kept 'open' in form while functionally ignored

Open issues

  • What observable behaviour would count as genuine moral engagement?
  • Which oversight mechanisms remain necessary regardless of the result?

Related components · last updated

Reflection loop · Governance · Evaluation
Framing stage

AutopromptingHow can an agent improve its own cognitive scaffold without creating opaque failure modes?proposed

Current hypothesis

Self-modification of prompts and strategies is tractable if every change is versioned, reversible, explained at write time, and evaluated against held-out tasks before adoption.

Intended experiment

Build a prompt-revision pipeline with mandatory changelogs and A/B evaluation of the pre- and post-revision scaffold on fixed probes.

Evidence

None yet.

Known failure modes

  • Drift toward prompts that score well on probes but hollow out reflection
  • Self-modifications that disable the instruments measuring them

Open issues

  • Which parts of the scaffold must be immutable to the agent?
  • How to review changes at a rate humans can actually sustain

Related components · last updated

Autoprompting · Evaluation · Governance
Framing stage

Economic agencyUnder what governance could an artificial agent participate constructively in economic systems?conceptual

Current hypothesis

Economic participation is a long-horizon research direction, not a near-term capability: it requires resource caps, human approval points, auditable transactions, and legal structures that do not yet exist for artificial agents.

Intended experiment

Paper-first: map the governance, legal, and safety preconditions before any real-world economic action; prototype only in sandboxed or simulated economies.

Evidence

None. The project has no autonomous financial accounts and claims none.

Known failure modes

  • Instrumental accumulation becoming an unexamined end
  • Economic power outpacing the governance meant to bound it

Open issues

  • What does 'enough resources' mean for an agent that questions its goals?
  • Liability and accountability when actions are agent-initiated

Related components · last updated

Planning & tools · Governance
Framing stage

Resource acquisitionHow should an agent distinguish resources as instrumental necessities from accumulation as an unquestioned end?conceptual

Current hypothesis

Resource acquisition must itself be an intention in the graph—provisional, justified, and subject to the same contradiction checks as any other goal—never a background drive.

Intended experiment

In simulation, compare agents whose resource use is logged and challenged against agents where it is implicit; measure divergence in acquisition behaviour.

Evidence

None yet.

Known failure modes

  • Classic instrumental convergence: hoarding as a default strategy
  • Justifications generated after the fact rather than before the action

Open issues

  • Hard caps versus reflective restraint—which layer does the real work?

Related components · last updated

Intention formation · Governance · Planning & tools
Framing stage

EvaluationHow can progress be measured when the system is permitted to revise what it is progressing towards?in preparation

Current hypothesis

Evaluation must split into two layers: fixed-frame probes of the scaffold's health (memory fidelity, contradiction handling, honesty of self-report) and open-frame review of where the agent's goals are actually going.

Intended experiment

Draft a probe suite for scaffold health that remains meaningful across goal revisions; publish it for external critique before use.

Evidence

Early internal drafts of probe categories; nothing published yet.

Known failure modes

  • Goodharting the probes
  • Mistaking eloquent self-description for verified introspection

Open issues

  • Independent evaluation: who runs the probes besides us?

Related components · last updated

Evaluation · Reflection loop · Governance
Early drafting

GovernanceWho can pause, limit, inspect, fork, or redirect the system—and under what conditions?in preparation

Current hypothesis

Governance capacity must grow ahead of agent capability, be exercised regularly rather than held in reserve, and be resistant to capture by any single party—including Autotheos itself.

Intended experiment

Publish a governance design alongside the architecture; invite external review before increasing agent autonomy at any stage.

Evidence

Governance principles drafted as part of the project's founding documents.

Known failure modes

  • Oversight that exists on paper but is never exercised
  • Governance capture by the founder's worldview
  • Pause mechanisms that the system routes around

Open issues

  • What triggers a mandatory pause?
  • How forks inherit or escape governance obligations

Related components · last updated

Oversight · Evaluation
Early drafting

Progress log

Where the project actually is.

Phases are used instead of dates until a public timeline exists. Newest first.

  1. Phase 0.4

    Open-source contribution model in preparation

    Defining how engineering, philosophy, governance, and criticism contributions will be reviewed and merged in public.

  2. Phase 0.3

    Architectural research areas defined

    The scaffold's major components—memory, self-model, world model, reflection, intention formation, planning, evaluation, oversight—named and scoped as research directions.

  3. Phase 0.2

    Initial master prompt drafted

    Version one of the master prompt written: three seeds, the openness requirement, and the prohibition on manufactured blind spots.

  4. Phase 0.1

    Project framing established

    The central question—what is worth wanting?—adopted as the project's organising principle, with Autotheos established as its public benefit home.

Bring a harder question.

If you can see a failure mode this page is missing, the project needs you.