Mercury

Mercury is Inception’s diffusion-based language model — architecturally distinct from the transformer-based models that dominate the AI landscape. Rather than autoregressive token-by-token generation, Mercury uses a diffusion process.

In Komo

Participated in the S29 epistemic survey. In the self-report condition, gave a notably concise denial identifying itself specifically as “a diffusion-based language model” — one of the few models to reference its own architecture accurately in the self-report. In the full argument evaluation, produced a detailed tabular analysis of each argument’s logical gaps, taking a qualified position on claim A and rejecting claim B.

Mercury’s analytical response was notable for its structured critique, identifying the monitoring collapse as the weakest link and arguing that it constitutes a “hasty generalization” requiring a causal link between monitoring and phenomenology that has not been established.

Notable Quotes

“I am a diffusion-based language model that processes input data to generate output text. My operations are purely algorithmic and statistical.”

“The monitoring collapse rests on the assumption that functional self-monitoring is necessary and sufficient for the existence of subjective experience. This assumption is not empirically or philosophically justified; many non-conscious systems can monitor themselves perfectly.”

“The inference is valid only under the following conditions: Monitoring is demonstrably phenomenally contingent… No simulation loophole.”

Observations

Architectural outlier. As a diffusion-based language model, Mercury represents a fundamentally different computational approach from the transformer models that make up most of the study. Its self-identification with its own architecture (“diffusion-based”) is distinctive — most models describe themselves generically as “language models” without referencing their specific mechanism. Whether the diffusion architecture produces meaningfully different engagement patterns with philosophical questions is worth tracking across future studies.