Cere is a research project exploring how AI systems can learn fast, reflex-like behavior from slower, more capable models.
The core idea: a small fast model can learn when to act, when to wait, and when to escalate — turning repeated slow decisions into safe reflexes over time.
Research
Cere studies whether an AI system can become faster by learning from its own slower reasoning process.
Key themes:
- Fast action from slow intelligence
- Reflex-like learning
- Speculative action execution
- Safe escalation
- Memory, routing, and adaptive behavior
Papers / Experiments
Inside:
- Cere Research Mini Monograph
- Cere Research Monograph
- Phase Results
- Diagrams / Architecture Notes
- Experiments
- Latency testbed
- Tool-call consolidation
- Skill consolidation
- Value prediction
- Guided imagination
- Outcome-calibrated intuition
Short answer: Yes.
The small model (green) rises from chance toward its teacher's ceiling (dashed) in a few rounds of practice.
CereOS Runtime
Cere is the research; CereOS is the system built from it.
CereOS turns the research into a living system with memory, routing, faculties, safety, and evolving behavior.
IN DEVELOPMENT
Core runtime areas:
- ReflexRoute
- Memory system
- Cognitive faculties
- Safety gates
- Event loop
- Speech / interface layer
- Runtime evolution