bbcc90

For sure but most of that innovation pre-super shoes was optimizing the runner, not the shoes

I'm a runner, and it's a bit sad that distance running is not longer purely about the runner.

Based on the quote below, next thing we will see is a "constructors championship" similar to F1 for winning shoe constructor in the 'major' marathons :-(.

" This dominance continued in 2024, with adidas athletes wearing Adizero models winning six out of 12 World Major Marathons – more than any other brand."

and yes, of course i race in super shoes :-).

Sure, but did they have MCP?

it does work; just download from HF and load in the app

(trying to move the critique beyond the title...)

When trying to deploy llms in with larger context windows constrained environments 2 things start to hurt: a) increased memory footprint for longer KV cache b) increased decode speed due to longer context window. this paper addresses a) only, which is useful, but we are still left with b) (right?)

any idea why this took them so long?

If you are going to go vertical then do it properly.

OpenAI could just build their own framework for internal use that works well on their silicon (see Jax+tpu)

Their starting point? Triton plus some triton libs. Jax chipped away at TF like this, and no reason why Triton can’t do the same to PyTorch.

Love this. I’m 10 years in SF myself. One thing I would add is the people

I would also say that being an engineer at L8 is a meaningfully different set of skills than L7. The whole ‘what got you here won’t get you there idea’ starts at L7 for either track…

Chiplets and standardized hardware interfaces are a great start.

Now we just need to solve the remaining 90%: firmware, software, security, virtualization, …, …