Slides for Iris

Methods and contributions

Read the source for the full argument, qualifications, and evidence.

The slides identify that in the GPT framing the latent semantic state is not separated by speaker, so speakers are admixed and cannot be weighted or attended to differentially.
Reusing a renormalization-group figure from Friston, Friedman et al. 2023, the slides sketch speaker attention as a portfolio (regime) of nested attentions at all levels.
Listed next steps are adding visualizations/dashboards, specifying the model in GNN toward implementation, and adding an action component.

Active Inference generative model drawn as a Bayesian graph — The slides open with a labelled graph of the A, B, C, D, E, G and policy terms of an Active Inference generative model as the reference frame.
Mapping GPT onto a minimal prior/state/observation model — A reduced D-s-o model is used to describe what GPT does: prior as training set and parameters, latent semantic state, and word strings as observations.
Perceptual-inference extension with a transition matrix B — Temporal dynamics are added via the perceptual inference part of Active Inference, with hidden states changing over time and no action selection.
Renormalization-group model of nested attention (Friston et al. 2023) — A figure from Friston, Friedman et al. 2023 is reused to sketch how speakers, groups and sets of groups could be modelled as nested attention.