GNN for Generative Model Supply Chains: A Golden Spike Moment for Multiagent Trajectory Planning with RxInfer.jl

all right hello welcome everyone it's May 15th 2025 this is active in for ant stream 14.1 and we're going to begin it as we do with a github push and we're going to go from there okay this stream is called generalized notation notation from plain text to triple play and before jumping into the background want to give a little bit of the payload up front here's the simple structure of the repo and where we're headed all the source code is in src all the documentation is in doc and all the output from running main goes to output so what is the output going to look like well first where does it start from and where is it going what it starts from is looking like a markdown file with a controlled vocabulary describing state spaces for a given generative model and here's where we take that first for each of those provided source files gnn files we get a bunch of visualizations so anyone who's worked with pymdp or active inference models knows that figuring out state spaces and doing all kinds of work and transformations with state spaces can be challenging so one thing this does is it takes that generative model distinction as specified and visualizes it it outports it exports the model into a variety of formats json text xml graphical all these different kinds of outputs there's a initial processing report essentially summarizing what kinds of files are being detected there are rendered simulations so here is a pymdp script that fills in not perfectly functionally at this point but i think more than enough to get the picture a pymdp script parameterized by the details of the state space of the source file and where this is all going is to be able to cross render into multiple implementation strategies so a stub method for going into rx infer type checking is performed so that is checking the validity of the code and making sure that the source files are valid and this allows us to bring in not just security features but also something that is often much desired for those running active inference models which is the computational resource estimation so here we can look at the state space sizes and the different kinds of operations that are entailed for a given generative model and then get different kinds of assessments like what is the relative computational cost going to be or what is going to be the size for a given number of variables of storing that specific model on on the ram or on the hard drive uh llms are used at the end to in this case summarize a set of questions so here are uh reports written with llms summarizing the states based model the logs for improved tracking and coding mcp is the model context protocol and we'll get into this a little bit more but this is a very large file which concatenates all of the mcp methods which we're exposing like visualize gnn file so this is something that you can run locally or we can stand up and we will stand up mcp servers so that these kinds of services can be provided ontology processing these are summaries of the ontology terms used and how they map onto different state spaces and this is really one of the key pieces of gnn which is putting a little bit of space between putting a flexible interface and gap between a given variable and its ontological annotation so commonly in looking at these generative models cognitive modeling type scripts you'll see the name of a variable like observation modality 1 or latent state or whatever it happens to be some english language name of a variable however that can get into some issues because what if it's modality 3 but then one of the other ones comes and goes or what how do you add modality 4 or it's the sense variable for one agent but it's the action output from another agent so in other words you don't want to have the variable always welded to its ontological assertion because there can be multiple english terms or roles for a given state space variable we're going to come back to that with cerebrum and then also there can be…