A short position essay. Sun’s Compression Principle says every computational model is a simplified picture of reality, and explicit symbolic reasoning makes up for what the simplification threw away. From it he draws a modelling-reasoning trade-off: the more of the world a model keeps, the less explicit reasoning it needs. Symbolic methods dominated because computation was scarce; foundation models spend their compute on building the model instead. His conclusion is that symbols are moving from the engine of AI to its interface with people: specifications, contracts, runtime monitors, regulations. The running example is a junction full of autonomous cars that may not need a traffic light, while pedestrians and regulators still do.
Findings
- No experiments or data. It is an argument by example.
- Once two real situations map to the same symbol, no amount of correct reasoning can recover the difference, so inference can be valid inside the model and still wrong in the world.
- Proposed consequences: formal methods should specify and verify behaviour that is visible from outside the system; explanations are for communication and need not trace every internal computation; AI governance is a problem of designing symbolic interfaces.
Limits
- The trade-off is never stated formally or tested.
- The claim that rich models need less explicit reasoning is not weighed against how rich models fail (hallucination, distribution shift), which is where symbolic checks earn their keep.
- Adversaries are absent. A symbolic interface that humans trust is also a target.
Relevance
The “interface” view suggests writing a robot team’s specifications over what a monitor can observe, the stance behind runtime-monitoring and safe-control. An open question for robust robotics: what happens when the symbol (a “task done” flag, a green light) is spoofed while the rich model underneath says otherwise?
Links
- arXiv
- Subfield: Robot learning