FL33TW00D

9 ms·

61.
Domain specific architectures for AI inference(fleetwood.dev)
1 points by FL33TW00D 2y ago|0 comments
62.
by FL33TW00D 2y ago
Plot big tech stock valuations with markers for successful OS model releases.
63.
by FL33TW00D 2y ago
Yes it is according my familiarity, but set notation has been the same since the 1870s, why should this language fuck it up? Rust does it correctly.
64.
by FL33TW00D 2y ago
These cards are super exciting. If they can improve the software story then they have a real shot.
65.
by FL33TW00D 2y ago
The array syntax is very offensive: `const a = [3]i32{ 1, 2, 3 };` A set is denoted by braces, not an array.
66.
by FL33TW00D 2y ago
This is why Google has 500 people working on the TPU compiler team.
67.
by FL33TW00D 2y ago
The TPU obviously uses a systolic array: https://jax-ml.github.io/scaling-book/tpus/
68.
by FL33TW00D 2y ago
Everything is an abject failure... until it works. All the best ideas are tried repeatedly until the input technologies are ripe enough.
69.
Domain specific architectures for AI inference(fleetwood.dev)
4 points by FL33TW00D 2y ago|0 comments
70.
Domain specific architectures for AI inference(fleetwood.dev)
3 points by FL33TW00D 2y ago|0 comments
71.
by FL33TW00D 2y ago
Why not just construct more homes?
72.
by FL33TW00D 2y ago
It depends on the batch size and the accelerator you're running on! Decode is *typically* memory bound unless you can hit high batch sizes (in the hundreds), which is hard during serving due to the contention between batch size and low
73.
by FL33TW00D 2y ago
It seems to me that MLA will become the standard from here on out. If Deepseek R1 had used standard MHA, they would need 1749KB per token for KV cache storage. This means that once the conversation reaches ~46,000 tokens, the KV cache will
74.
by FL33TW00D 2y ago
You have it backwards. Training and prefill are compute bound. Decode is memory bound. FlashAttention massively increases the arithmetic intensity of naive MHA, such that you can remain compute bound at lower batch sizes during decode.
75.
by FL33TW00D 2y ago
They should just choose RocketLab: https://x.com/Peter_J_Beck/status/1876834824100401519
76.
by FL33TW00D 2y ago
Everyone calling Loom a glorified screen recorder... there was a $975M dollar bill on the floor and no one else picked it up. The technology was simple, but it solved a real problem for people. Solving well defined engineering tasks is vast
77.
Tenstorrent Wormhole Series(corsix.org)
9 points by FL33TW00D 2y ago|0 comments
78.
by FL33TW00D 2y ago
So many uninformed people in the comments here. "The Carbon Price Support (CPS), introduced by the UK Government in April 2013, has led to a substantial reduction in electricity generated from coal, which fell by 93% from a monthly ave
79.
by FL33TW00D 2y ago
Surprised by all the negativity here. The USA benefits enormously from skilled immigration: "doubling the size of the US H1B visa program increases US and EU growth by 4% in the long-run" From a recent paper here: https:/&#x
80.
by FL33TW00D 2y ago
The title is a reference to "The unbearable lightness of being" by Milan Kundera for those unaware.
81.
by FL33TW00D 2y ago
I raise you Warren McCulloch in 1962: https://www.youtube.com/watch?v=wawMjJUCMVw
82.
Layer Normalization as fast as possible (2023)(fleetwood.dev)
1 points by FL33TW00D 2y ago|0 comments
83.
by FL33TW00D 2y ago
Donate to GiveWell: https://secure.givewell.org/
84.
by FL33TW00D 2y ago
I don't think so. Listen to the most recent RKLB earnings call. Constellation on the way, Neutron launching in 2025. There will be competition.
85.
by FL33TW00D 2y ago
Rocket Lab and Peter Beck are your best hope!
86.
by FL33TW00D 2y ago
"This post intends to limit the mathematical knowledge required to follow along, but some basic linear algebra, trigonometry and understanding of self attention is expected." If you're not sure on self attention, the post wil
87.
by FL33TW00D 2y ago
Hm! Very interesting! Thank you for taking the time to debug that. I'm going to have to think hard about how to rewrite the motivating example to explain this best. Edit: updated the post, thanks for pointing out the pernicious init va
88.
by FL33TW00D 2y ago
Thank you! Seemed like voodoo to me too, hence this post!
89.
by FL33TW00D 2y ago
Run the code and look at the values!
90.
by FL33TW00D 2y ago
The first code example says False because of high precision, I've updated the example.

More ›