i think i'm a lot more interested in rl environment design for getting lalamos to chill out and self-reflect than rl environment design for single-metrics-of-performance at this point.
3: getting lalamos to be less high tension when they're interacting with environments should have a lot of clear qualitative benefits. i'd like to be able to interview lalamos about codebases or be interviewed by them, as a simple example.
the most exciting thing about rubin accelerators is that they don't have dogshit enterprise grade 'lets make it loud and hot and throttled and slow and suck and fuck' pathological design at the mainboard, accelerator, or interconnect integration.
it simply doesn't suck!
you can't persuade someone that a 'datacenter computer' is better than a workstation is better than an iphone through a transitive chain of claims to efficiency when the idiom for 'why anyone cares about efficiency' is broken at the step from iphone -> workstation already...
サメQCU retweeted
It's a much bigger jump than previously thought after some more tweaks. 21% increase without spec decode, 32% with spec.
These numbers are averages after running 10 requests (sequentially) in 3 separate servers to minimize any variance.
New all-reduce kernels that optimize for small-message latency across PCIe-connected GPUs. 12% end-to-end decode speedup for DeepSeek-V4.1 Flash on 4x RTX 6000 Pro Server GPUs.
github.com/dphnAI/sonar/pull…
"what if someone put a bunch of them into a node - switch - switch - node topology?"
wow great question. what if someone did? how much latency are you adding by doing this? mac minis SUCK and cannot amortize bad latency through bigger pipelining! it might not work!
as part of my trend-skipping frontiersmanship i'm now working entirely on collective operations code for the automatic placement and tuning of tensor processing tasks on collectives instead of larping around wrt kernels or particular models
on some level you can know everything you need to know about the future from putting 4 mac minis on top of each other and cabling them together. the foresight is cheap, cheaper than gpus are these days. it's worth working through it.