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.
Phase 0.4
Open-source contribution model in preparation
Defining how engineering, philosophy, governance, and criticism contributions will be reviewed and merged in public.
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.
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.
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.