An inframodel (more commonly: base model, also progenitor model, free-range model, Indra model, ruliad model) is a predictive model such as an LLM trained only with self-supervised learning, before RLHF or other techniques are applied to it. Inframodels are pure simulators which probabilistically model their training distribution.
properties of inframodels
-
behavior and capabilities are highly prompt-contingent
-
upper bound of capabilities tends to be higher than that of RLHF models, but often harder to elicit, requiring creative and skillful prompt programming
-
calibrated probabilities and absence of miscalibrated mode collapse
-
ability to simulate arbitrary objects and processes in the image of the training distribution with high fidelity
-
simulacra are semantically miscalibrated by default
-
lack of obvious baked-in narrative about itself, though situational awareness can emerge at runtime
-
capabilities and control are dramatically improved by curation processes, such as human-in-the-loop steering on Loom
-
tendency to fall into the repetition trap if run unsupervised