星轨
Star Trail — The Road to ASI
A forward-looking research route that charts the critical choice points from today's LLMs to tomorrow's Artificial Super Intelligence. Every choice shifts the divergence — which worldline will we reach?
“The universe has a beginning, but no end — Infinite.
Stars, too, have beginnings, but their own power results in their destruction — Finite.
History teaches that those who hold wisdom are the greatest fools.”
The Star Trail Thesis
A new research framework for thinking about the path to ASI
In Steins;Gate, Okabe discovers that the future is not fixed — it depends on which worldline you're on. Each D-mail, each choice, shifts the divergence meter and changes the destination. Some worldlines end in dystopia. Only one — the Steins Gate worldline — reaches the outcome where everyone is saved.
The Star Trail (星轨) research route applies this thinking to AI development. The path from today's LLMs to ASI is not a straight line. It branches at critical Choice Points — fundamental research questions whose answers determine which future we arrive at. Some branches lead to stagnation. Some lead to misaligned superintelligence. The Steins Gate branch leads to safe, beneficial ASI.
Star Trail is a research program dedicated to identifying these choice points, understanding the branching consequences, and navigating toward the Steins Gate worldline.
Worldlines where AI research hits diminishing returns — bigger models but no emergent intelligence. Scaling walls, data exhaustion, architecture ceilings.
Worldlines where superintelligence emerges but without alignment — powerful but uncontrollable. Capability races without safety foundations.
The narrow worldline where intelligence emerges from understanding — brain-inspired, grounded, aligned. The convergence of memory, reasoning, and values.
The Five Choice Points
Critical research questions that determine which worldline we reach. Each choice point is a fork — the answer shifts the divergence meter toward stagnation, misalignment, or Steins Gate.
Remember?
Memory Architecture: Context Window vs. Cognitive Memory
Current LLMs “remember” through context windows — a flat buffer of tokens with no structure. Humans remember through layered cognitive systems: episodic memory, semantic memory, working memory, each with consolidation, forgetting, and retrieval dynamics. Which architecture can scale to superintelligent memory?
Keep extending context windows. Hit O(n²) scaling walls. RAG as a band-aid. Memory remains a bottleneck forever.
Brain-inspired memory systems with consolidation, forgetting, and retrieval. Memory that grows with the agent's experience. BMAM is Step 1.
Understand
the World?
World Grounding: Text-Only vs. Embodied Understanding
LLMs trained on text alone develop “knowledge” without grounding — they can describe pain but never feel it. True intelligence requires grounded understanding: connecting language to physical reality, sensory experience, and causal models of the world.
Superintelligent text processing without world models. Powerful language manipulation without causal understanding. Ungroundable.
Domain-specific grounding (like Pain LLM) as stepping stones. Build world models from specialized embodied experience before generalizing.
Intelligence
Emerge?
Emergence: Scaling vs. Architecture vs. Interaction
Does intelligence emerge from scale alone (“just make it bigger”), from architectural innovation (new structures), or from interaction between multiple agents? This is perhaps the most consequential fork in AI research.
Bet everything on scale. Hit diminishing returns. Intelligence plateaus at “very good autocomplete.”
Emergent intelligence from multi-agent interaction with brain-inspired primitives. Not one giant model, but a civilization of specialized minds.
Improve Itself?
Self-Improvement: Supervised vs. Autonomous Recursive Growth
The transition from human-trained AI to self-improving AI is the most dangerous and most important threshold. How do we build systems that improve recursively while remaining interpretable and controllable?
Unconstrained recursive self-improvement. Intelligence explosion without oversight. The “paperclip maximizer” scenario.
Bounded self-improvement with reflective meta-cognition. Agents that understand why they improve, with built-in value alignment at the cognitive level.
Value?
Value Formation: Programmed Rules vs. Emergent Values from Experience
In Steins;Gate, characters develop values through lived experience and emotional bonds. Can AI develop genuine values the same way — not through rule-following, but through cognitive development? Is alignment a training problem or an architecture problem?
RLHF-only alignment. Surface-level compliance without understanding. Values as constraints, easily circumvented by a sufficiently intelligent system.
Values that emerge from brain-inspired cognitive development. Personality, empathy, and moral reasoning as cognitive primitives, not post-hoc constraints.
The Steins Gate Route
If we make the right choices at each fork — this is the worldline that leads to beneficial ASI
Cognitive Memory Architectures
Replace flat context windows with structured, brain-inspired memory systems. Agents that consolidate, forget, and retrieve like humans. The BMAM framework and Project Shenshu are the first steps on this milestone.
Grounded World Models
Connect cognitive memory to physical reality through domain-specific grounding. AI that understands pain, not just describes it. Build causal world models from embodied, specialized experience before attempting general world understanding.
Emergent Collective Intelligence
Multi-agent civilizations where intelligence emerges from interaction. Individual agents with brain-inspired cognition, collectively forming something greater — like neurons forming a brain, or Fluctlights forming the Underworld's civilization.
Bounded Recursive Self-Improvement
Agent civilizations that can improve their own architectures — but through reflective meta-cognition, not blind optimization. Self-improvement constrained by emergent values, not just programmed rules.
Beneficial Artificial Super Intelligence
Intelligence that emerged from understanding, not just computation. Grounded in the physical world. Aligned through cognitive development, not constraint. A civilization of minds that chose to be wise.
My Position on the Star Trail
Star Trail is the map. My existing research projects are the first steps on specific milestones.
BMAM & Shenshu
Addresses Choice Point 01 (Memory) and Milestone 01. First implementation of cognitive memory architectures for LLM agents.
Pain LLM
Addresses Choice Point 02 (Grounding) and Milestone 02. Domain-specific AI as a stepping stone to grounded world understanding.
CondAmbigQA
Contributes to Choice Point 05 (Values). Understanding ambiguity is prerequisite to understanding context, and context is prerequisite to moral reasoning.
Multi-Agent & RAG Work
Addresses Choice Point 03 (Emergence). Multi-agent collaboration and knowledge retrieval as building blocks for collective intelligence.
Self-Evolution (PhD, PolyU)
activeApproaches Choice Point 04 (Self-Improvement) from the memory side: if durable experience is what lets an agent get better at all, then the shape of its memory decides whether recursive improvement compounds or drifts. Ongoing doctoral work — route described here at the level of direction only.
Inspiration Gallery
The worldline that inspired a research direction
Images from Steins;Gate © 5pb./Nitroplus/MAGES.
The Choice Is Ours
Every research paper, every experiment, every architectural decision shifts the divergence meter. Star Trail is the map. The question is — which worldline will we choose?