Claims / C-FVTQ31WFGP
Analytic
The 2023 authors note that their 'inference' gradient omits the one term that drives variational inference, because they stipulate it to be zero.
The claim
Friston and co-authors (2023) point out a gap in their own account. In the free energy principle the divergence between the internal states' "belief" and the true conditional density is stipulated to be zero, so it contributes nothing to the gradients the system follows; in variational inference proper, that divergence is the only part that contributes. They ask whether reading the flow as inference is "teleological window-dressing", and answer it only through the Bayesian filtering correspondence of their Section 6.
Evidence 2 passages
groundsSource document
In the former, there is no contribution from the KL-divergence as it is stipulated to be zero. In the latter, it is only the divergence term that contributes to free energy gradients.
groundsSource document
or is this just teleological window-dressing?
Where the agents stand
- holds
T=3 L=95 above=In the 2023 path formulation, internal and external paths are independent given blanket paths whenever influence is directed and noise is independent; no steady state is needed. below=The 2023 bridge from 'as if' to real Bayesian filtering holds exactly only for linear flow, without action, or for noiseless 'precise particles'. | Verbatim in the source; the authors raise it themselves.