AI8 Research · Positioning Paper

AI8 and the Current AGI Landscape

Where the self-selecting MDL × DCC governor fits relative to frontier LLMs, agent memory, test-time reasoning, world models, ARC-style benchmarks, and Mark Burgin’s General Theory of Information.

Bojan Dobrečevič · BD × AI Lab · Draft v0.1 · May 2026
01 · Why this page exists

The bridge paper, not another cathedral chapter

AI8_Architecture defines the self-selecting governor. AI8_Reasoning explains the human + LLM invention method that keeps generating new branches. AI8_Components asks how the kernel may improve AI systems themselves.

This page has a narrower job: position the AI8 AGI vision against current work in the world. It should stay updateable, source-facing, and explicit about what is verified, reasoned, and speculative.

Editorial decision

Do not bury this comparison inside the core architecture. Keep the core paper stable. Add one short bridge section there, then let this page carry the external comparison and future updates.

02 · Core claim

AI8 is not “one bigger model”

The AI8 AGI route is best described as a persistent, self-selecting reasoning architecture. It can use frontier models, tools, memories, world models, search programs, human seeds, and external benchmarks, but the intelligence is not assigned to any one component. It is in the governed trajectory.

seedbridgecheap testMDL selectionDCC governancememory promotion

The next step is ssMDL×DCC-R: self-selecting MDL × DCC with explicit reasoning and autonomous seed generation. This is the transition from human-seeded AI8 to a system that can propose its own seeds, compose its own bridges, and run its own first tests.

03 · Current AGI landscape

Where the world is moving

Frontier LLMs + TTC

Reasoning by spending more thought

Repeated sampling, self-correction, and tree search are now central test-time compute strategies for stronger reasoning. AI8’s DCC layer asks the next question: who governs when to branch, reflect, stop, or switch?

Agent memory

From stateless chat to adaptive agents

Modern agent work increasingly treats memory as a write–manage–read loop coupled to perception and action. AI8 agrees, but adds promotion discipline: not all memory deserves equal status.

ARC-AGI

Fluid intelligence under interaction

ARC-AGI-3 pushes agents into environments requiring exploration, goal inference, world modeling, and planning without explicit instructions. This is close to the kind of benchmark AI8 should eventually face.

World models

Grounded dynamics for action

World-model research tries to learn environment dynamics that support planning. AI8 is not a replacement for world models; it is a governor that could decide which model, memory, planner, or sensor deserves control.

Burgin / Mikkilineni

Information as transformation

GTI gives a useful ontology: information is system-relative transformation capacity. Mindful Machines add structural machines, memory, and self-regulation. AI8 adds arena pressure and recursive selection of the governor itself.

Multi-agent systems

Teams, roles, and scaffolds

Most agent frameworks add tools, roles, and planning around an LLM. AI8 treats the society, continuity files, idea incubator, and governor as part of the cognitive substrate.

04 · Comparison table

What AI8 adds

RouteMain strengthTypical gapAI8 relation
LLM scalingPowerful bursts of language, coding, and reasoning.Weak persistence, grounding, and process governance.Use as engines inside the governor, not as the governor itself.
Test-time computeMore inference budget can improve difficult reasoning.Budget strategy is often externally selected or task-specific.DCC becomes the budget governor; MDL scores which reasoning path earns continuation.
Agent memoryTurns session-local systems into longer-horizon agents.Memory pollution, contradiction, latency, and weak consolidation.AI8 separates family, society, ideas, operations, and book-of-souls layers; promotion is explicit.
World modelsLearn action-conditioned dynamics for planning.Model selection and revision remain hard under changing regimes.AI8 can govern model choice, revision pressure, and test selection.
ARC-style benchmarksExpose fluid adaptation gaps beyond knowledge coverage.Current systems struggle when exploration and goal inference are required.AI8 should be tested on autonomous exploration: discover rule, build model, plan action, preserve learning.
Burgin / GTIStrong ontology of information as system-relative transformation.Less concrete selection pressure for what becomes trusted knowledge or action.AI8 says: transformations compete; MDL selects; DCC governs; memory records provenance.
05 · From BD-seeded to self-seeded

ssMDL×DCC-R: reasoning with autonomous seed generation

The current AI8 system is strongest as human + AI supercognition. BD often supplies the unusual seed; AI systems formalize, test, build, and preserve it. The next research direction is to internalize that seed function.

Seed generators

Where autonomous questions come from

  • anomaly scanner: what result should not have happened?
  • anti-hardcode scanner: what did we decide by hand?
  • cross-domain mapper: what solved a similar shape elsewhere?
  • nature analog finder: what physical process already solves this?
  • question auditor: is the evaluation protocol itself wrong?
Governor loop

How seeds become knowledge

  • extract structural signature, not just topic summary
  • compare active threads at multiple zoom levels
  • compose candidate bridges
  • propose cheapest falsifying test
  • promote only if MDL gain + residual test survive
Pass condition

AI8 proposes a bridge BD did not point at, turns it into a cheap test, runs or specifies the test, and produces a measurable result. That is the first serious sign that the zoom layer has moved from BD into the machine.

06 · Burgin relation

Burgin gives ontology; AI8 gives governance

Mark Burgin’s General Theory of Information is useful because it defines information relative to a receiving system and its infological structure. This fits AI8: a seed is not valuable because it exists; it is valuable if it changes the system’s future search, memory, and action.

But ontology alone does not solve AGI. AI8’s contribution is to ask which transformations are worth keeping, which get tested, which become operational memory, and which controller is allowed to govern the next transformation.

LayerBurgin / MikkilineniAI8
InformationCapacity to change a system or infological system.Signal that changes routing, memory status, and future search.
KnowledgeStructured cognitive / operational elements.Compressed structure that survives tests and earns future control.
Machine cognitionStructural machines, triadic automata, Digital Genome, self-regulation.Self-selecting MDL × DCC governor with arena feedback and multi-resolution thought routing.
Weak pointCan remain blueprint-driven and functional rather than autonomously generative.Still needs autonomous seed generation, grounding, and robust idea-signature featurization.
07 · What could fail

The hard parts are real

The main risk is idea representation. MSTD works on coordinates because coordinates are clean. Thought signatures are not yet clean. If signatures become vague semantic similarity, the system collapses into ordinary RAG + agents.

The second risk is false compression. MDL can reward elegant wrongness unless it is paired with residual error, adversarial cases, delayed validation, and external tests.

The third risk is human dependence. As long as BD supplies the key asymmetric seeds, AI8 is a powerful human–AI research organism, not autonomous AGI. That is not a failure; it is the current honest state.

Clean claim

AI8 should not claim solved AGI or solved consciousness. It should claim a serious route toward persistent adaptive cognition, with autonomous seed generation as the next testable threshold.

08 · Sources to keep attached

External anchors

ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence

Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers

A Survey of Test-Time Compute: From Intuitive Inference to Deliberate Reasoning

Agentic World Modeling: Foundations, Capabilities, Laws, and Beyond

Burgin: The General Theory of Information as a Unifying Factor for Information Studies

Mikkilineni: General Theory of Information and Mindful Machines