marmaduke

8 ms·

61.
by marmaduke 2y ago
> how many other things there are that people subconsciously learns, without it ever becoming obviously noticeable for them? 95%+? Stuff like how to jiggle keys on a keyring to get the right one for the door you're about to unlock?
62.
by marmaduke 2y ago
Sounds like science to me
63.
by marmaduke 2y ago
Yeah, nice, you could also consider a low rank factorization as well.
64.
by marmaduke 2y ago
Well, certainly puts things in perspective, doesn't it?
65.
by marmaduke 2y ago
Having used CSR it's not surprising, and some newer formats might have more mechanical sympathy like block ELL, since they avoid uncoalesced reads / gathers, tho the code is trickier.
66.
by marmaduke 2y ago
> there's nothing more dangerous than a teenager with very little money and a lot of time on their hands, so a simple DNS-based block definitely won't suffice, you probably need Chinese-style deep packet inspection and such Or
67.
by marmaduke 3y ago
Numba is great but missing explicit simd, stack allocated arrays, and user defined types, to name a few things. I tried to keep my colleagues in Jax these days.
68.
by marmaduke 3y ago
> curated output of the brainstormy process that immediately precedes it Daniel Dennett gives a nice albeit more detailed version of your idea in his book Consciousness Explained, could be worth a read
69.
by marmaduke 3y ago
Hm, let's wait and see what the gemm/W perf is, and how many programmer hours it takes to implement say an mlp. Wafer scale data flow may not be a solved problem?
70.
by marmaduke 3y ago
Just for comparison, Swiss CSCS new Alps system will get 5k GH200 nodes (each with a H100).
71.
by marmaduke 3y ago
Was there not similar rhetoric about asymmetric cryptography?
72.
by marmaduke 3y ago
Since performance was a consideration for you, how does it compare to TBB or even say OpenMP? I've used those as well as the task system that comes with ISPC for HPC stuff, but I like the lightweight C11 approach here.
73.
by marmaduke 3y ago
for instance, running a nextcloud instance through docker compose, I'd like to have a lightweight web ui for browsing the server logs. not just for 15 minutes but for longer periods of time. on an hpc cluster, I want to want logs from
74.
by marmaduke 3y ago
I was looking for something like this for a while, and it could be just the landing page, but I much prefer a tool that's been built slowly with thought and consideration (which is evident from your landing page). How does your tool ha
75.
by marmaduke 3y ago
This seems like a great argument that bureaucracy is a just another kind of biological organization. It follows from Maturana's theory of autopoeisis, any version of homeostasis, or more recent ideas from Friston on self-evidencing.
76.
by marmaduke 3y ago
Is it really the fault of corps? I mean, is that a useful explanation? Corps exist to maximize profit don't they? Perhaps we have to consider why single use plastic is appealing in the first place. The culture of use once and throw aw
77.
by marmaduke 3y ago
For envs which are pets, who cares, but this will be great for those cattle situations.
78.
by marmaduke 3y ago
Any chance you'll support "traditional" batch HPC environments like Slurm?
79.
by marmaduke 3y ago
The math is not designed to intimidate but rather approach the "how to build sequence model" in a principled way from state space models, which draws from an arguably longer literature than neural networks. Some of concepts are be
80.
by marmaduke 3y ago
Just read some summary of his points elsewhere and had a hard time taking it seriously. For instance, > rule seven—which is basically that once you have power, money, and success, people will forget and forgive how you got there—is such
81.
by marmaduke 3y ago
What do you think about the gravity well "drive" in three body problem trilogy? I thought it was fairly clever, especially given the stories describing it beforehand (maelstrom iirc?)
82.
by marmaduke 3y ago
I can't buy this kind of argument anymore. How about the external effect of AI steering the entire semiconductor industry to increase GPU/NPU capacity?
83.
by marmaduke 3y ago
CPU mark scores are 32k and 39k respectively. It's at least twice as efficient
84.
by marmaduke 3y ago
I think the group working on rwkv has yes, even if they don't show the details in this article. (I followed them on their discord channels for quite some time). The interesting take away for me was that training rwkv from zero to inte
85.
by marmaduke 3y ago
These models don't have a fixed context size and are progressively fine-tuned for longer and longer contexts. The context length also doesn't impact inference cost. Another aspect of performance is not just how well does the trai
86.
by marmaduke 3y ago
Rosen's Life Itself is another, along with Maturana's work on autopoeisis.
87.
by marmaduke 3y ago
You load the qemu portably onto the usb as well
88.
by marmaduke 3y ago
Hm I'd take a stab at a Jax version based on this. Thanks
89.
by marmaduke 3y ago
> insights are already as deep (actually more deep) as I get from most humans I ask for help This was my thought as well. But then I figured if I can't get someone to give me thoughtful feedback, I might have bigger problems to so
90.
by marmaduke 3y ago
> That is equivalent On what number or op for the h100?

More ›