GraphRouter is an interesting concept. A Wiki implementation that would allow many people to input the performance of different LLMs on different tasks might be helpful. Users should also be able to specify their hardware, such as Oneplus 11 smartphone with 16 GB RAM, so the GraphRouter would select an LLM that is optimized for it.
From a real world product perspective a graph ML approach towards effective prompt routing is completely over engineered. The underlying problem is simple and given the speed and intensity of ongoing progress a solution must consider performance and adaptability at the same time.
I agree that most are just adding a higher level complexity where it is really unneeded. This does have its place however. I use this same process in general for all of my LLM selections, Prompt reiteration, and even command functions... but my graphs are DAGs built off of a custom local BTRfs subvolume set (and a btrfs/Freenet repo). When your FS is the Graph your system traverses it makes this higher level complexity go away and allows the benefits of the Graph Services to be had at a "root level". Nothing like having an LLM Agent using your FS and eBPF/Kernel hook on a GNN's Index that is the system itself... Makes the Desktop a convenience for user display instead of a need for computing or tasking ;)
Its not this hard... Distributed neural network. You add on tensors onto each edge device depending how much they can handle then you distribute the computing across many many many devices. You can scale this infinitely. Its like a bonnet neural net.
GraphRouter is an interesting concept. A Wiki implementation that would allow many people to input the performance of different LLMs on different tasks might be helpful. Users should also be able to specify their hardware, such as Oneplus 11 smartphone with 16 GB RAM, so the GraphRouter would select an LLM that is optimized for it.
From a real world product perspective a graph ML approach towards effective prompt routing is completely over engineered. The underlying problem is simple and given the speed and intensity of ongoing progress a solution must consider performance and adaptability at the same time.
I agree that most are just adding a higher level complexity where it is really unneeded. This does have its place however. I use this same process in general for all of my LLM selections, Prompt reiteration, and even command functions... but my graphs are DAGs built off of a custom local BTRfs subvolume set (and a btrfs/Freenet repo). When your FS is the Graph your system traverses it makes this higher level complexity go away and allows the benefits of the Graph Services to be had at a "root level". Nothing like having an LLM Agent using your FS and eBPF/Kernel hook on a GNN's Index that is the system itself... Makes the Desktop a convenience for user display instead of a need for computing or tasking ;)
A neural network trained to find novel neural networks .
Its not this hard...
Distributed neural network. You add on tensors onto each edge device depending how much they can handle then you distribute the computing across many many many devices. You can scale this infinitely. Its like a bonnet neural net.