logoalt Hacker News

catapart • today at 2:13 PM • 1 reply • view on HN

For me, the edges are still too blurry for me to make out what this actually is. I need someone with more in-depth knowledge to let me know what is actually being proposed here.

On the optimistic side, this looks like a kind of "bigger step" in the direction that GPU/shader programming took, where you change the way you think about and break down a problem to get insanely fast results. Feels like if you target this interaction net structure and give it to this CPU that uses them natively, we could amplify the speed of computations to possibly allow for things that are currently way to slow to do. Realtime physics sims with magnitudes faster steps. That's pretty cool. And all of the downstream effects that would allow for, in product design, scientific modelling, etc. And that's just one train of thought.

On the pessimistic side, though, I have no clue how to "program for an interaction net". Seems like the authors took a "you don't need to, an LLM will do it for you" position, and I can get behind that, but that's a pretty huge step in the direction of "Natural Language Programming" in that I literally can't program it without using an LLM and natural language, because I don't understand the underlying details and probably won't have any structural support to understand it for a while, as the tooling slowly rolls out and gets iterated on, until it can do something we can recognize as "Debugging". I imagine there's some kind of "program it in C, and translate it" thing, too, but it feels like the point here is not to have me be able to debug it, and a detour through C just wastes LLM tokens.

So those are the initial impressions, but then I'm making a TON of assumptions in both sides of that. So I don't even know if my initial impressions are conceptualizing the right paradigms, much less if they are worth discussing with people who are seriously considering this topic. Only speaking for myself here, but at least my own silence is due to not having enough information myself, and not seeing enough in this paper to interest me in research in the topic just yet. Maybe if they've got something I can play with, I could get my head around it more. But, for now, it's just neat.


Replies

jcranmer • today at 4:26 PM

> Feels like if you target this interaction net structure and give it to this CPU that uses them natively, we could amplify the speed of computations to possibly allow for things that are currently way to slow to do.

I mean, it's basically what out-of-order CPUs already do: read lots of instructions in at the same time, dispatch instructions as their inputs become ready to execute.

Sort of the reality of modern processors is that we have the units available to do the work, we just don't have dispatch logic to keep them fed with work, given how slow memory is. When you have very regular parallelism--dense linear algebra is the canonical example here--SIMD works very efficiently to minimize the dispatch logic needed to keep units fed. When you've got irregular parallelism, dataflow architectures seem to work well at first, until you start trying to make actual chips and you find you've been consistently underestimating how much overhead your dispatch logic is actually chewing up.

➕ show 1 reply