@soumithchintala - AI Dynamics

Philippe’s Trolling Reshape Function Naming Convention

By

–

31 August 2023 19h32

looks like phillippe wanted to troll everyone by calling what should've been `tl.reshape` as `tl.view`.
It's like those trolly C macros ( #define float int )

→ View original post on X — @soumithchintala

31 August 2023

CPU Programming Complexity Versus GPU Performance Models

By

@soumithchintala

–

29 August 2023 23h40

high-performance CPU programming IMO is harder than GPU programming.
GPU performance model is a lot more uniform and easy to understand than CPU performance which has so many edges and discontinuities and so many layers of caches and a much more complex instruction pipeline.

→ View original post on X — @soumithchintala

29 August 2023

Hardware Expert Misunderstands ML and Open Source Movements

By

@soumithchintala

–

29 August 2023 4h23

one thing that was clear was that it was written by a Semi analyst — with all of the biases and shortcomings and misunderstandings that you'd expect from someone who deeply understands hardware, somewhat misunderstands ML & products and poorly understands open movements.

→ View original post on X — @soumithchintala

29 August 2023

FX Tool Enables Model Portability Across IR Frameworks

By

@soumithchintala

–

26 August 2023 18h28

i heard someone built this thing called FX that lets you move models from one IR to another. the rest is an easy exercise left to the reader

→ View original post on X — @soumithchintala

26 August 2023

Salary transparency beyond job title homogeneity using histograms

By

@soumithchintala

–

26 August 2023 18h26

http://
levels.fyi tries to democratize salary information, its only aided by title homogenity. you an still build salary transparency with unique titles — just report histograms across all submitted data points

→ View original post on X — @soumithchintala

26 August 2023

Papers with Code bug fix deployed successfully

By

@soumithchintala

–

24 August 2023 16h46

better late than ever, I onboarded onto the paperswithcode codebase and found the bug and made the fix and @rbstojnic graciously re-deployed the code. All instances like this should be correctly classified now (or over the next few days as the script runs through). Addendum:

→ View original post on X — @soumithchintala

24 August 2023

CodeLlama: Meta’s Llama2 Fine-Tuned for Code Tasks

By

@soumithchintala

–

24 August 2023 16h06

CodeLlama — a version of Llama2 that was fine-tuned for code tasks is live now. Available in 7B, 13B and 34B. https://
ai.meta.com/blog/code-llam
a-large-language-model-coding/
…

→ View original post on X — @soumithchintala

24 August 2023