I mean, it's not like interaction nets don't come up every now and again. I remember the Bend language making the rounds here several times, v1 of that language was built on their own HVM, which uses interaction nets, it's a niche idea but not a completely forgotten one (Bend 2 appears to have dropped them due to perf reasons). We also get the occasional articles on revolutionary new (old) CPU architectures that will displace everything, last one I saw was pushing VLIW as the silver bullet that would displace the current CPU giants and make x86 obsolete. I don't see much reason to believe that this approach would overcome the main barrier to adoption that those approaches failed on, which is the cost of porting existing software across, unless the creators are willing to front the token costs for that.
Also, the article title is one that heavily emphasizes the AI autoresearch aspects, and doesn't mention interaction nets at all, so it makes sense that people would go into the article primed to think about the autoresearch aspects.
Did not know Bend dropped them. https://raw.githubusercontent.com/HigherOrderCO/HVM/main/pap... was interesting.
People keep rediscovering the same ideas and renaming it. Interaction nets are another kind of graphical dataflow model. There are lots, but they are all basically the same idea.
//I don't see much reason to believe that this approach would overcome the main barrier to adoption that those approaches failed on, which is the cost of porting existing software across,
Llms are pretty good at porting software.
You don't have to make x86 obsolete to be successful. You can just add extra accelerator chips. GPUs did this, NPUs are trying to do this, and there are plenty of exotic accelerators inside modern computers doing specialized DSP tasks.
If this stuff works, it'll surface first as some PCIe extension device that work can be dispatched to when it makes sense. Then maybe it'll get integrated into SoCs and be exposed via new CPU instructions or devices. Perhaps compilers can then start to identify blocks of code that can be reduced to interaction nets, compile the net and then run them on Frostbite style chips opportunistically.