hmmm, so its one of those location refactors (quite sad, but classic large-company stuff).
I've generally found that Foundation / Core language engineers are pretty hard to replace or be fungible — they have deep knowledge about complex code and social dynamics that are hard to
@soumithchintala
-
Foundation Engineers Hard Replace Large Tech Restructuring
By
–
-
Google Lays Off Entire Python Foundations Team
By
–
apparently Google laid off their entire Python Foundations team, WTF!
( @SkyLi0n who is one of the pybind11 maintainers just informed me, asking what ways they can re-fund pybind11)
The team seems to have done substantial work that seems critical for Google internally as well. -
SIMD and Parallelization Optimization Challenges
By
–
yea that and other SIMD / parallelization unoptimizations
-
PyTorch API design constraints for computational efficiency
By
–
pytorch has a history of constraining API to be computationally efficient (in eager), in a "dont let user shoot themselves in the foot" kinda way. this is one of those constraints.
-
Gemini 1.5 Pro: Model size specifications and capabilities
By
–
what do you mean by "it's class"?
Did Gemini 1.5 Pro's size ever get revealed? -
Microsoft Phi-3 Model Shows Impressive Results
By
–
nice work on Phi-3 @SebastienBubeck and team 🙂
— Soumith Chintala (@soumithchintala) 23 avril 2024
results look really impressive. https://t.co/6sLVl478G5nice work on Phi-3 @SebastienBubeck and team 🙂
results look really impressive. -
Llama3-70B Achieves Top Ranking on English-only Benchmark
By
–
also, Llama3-70B is #1 on English-only, whut!!!!
-
Llama3-70B ranks fifth as 405B model approaches release
By
–
Llama3-70B has settled at #5. With 405B still to come next… I remember when GPT-4 released in March 2023, it looked like it was nearly-impossible to get to the same performance. Since then, I've seen @Ahmad_Al_Dahle and the rest of the GenAI org in a chaotic rise to focus,
-
CUDA CuDNN Python Startup Costs and Compilation Overhead
By
–
its unclear to me even for code because compile-time and startup costs (even for a CUDA/CuDNN loading Python program) is 1 second.
If you strip out the binary blobs that need to load (say business logic filled program), then its quicker, but i'm not sure if there aren't other -
Agent Bottlenecks: UI Interaction Speed vs Direct API Access
By
–
having spent a little bit of time working on agents, the bottleneck if often not inference speed but interaction speed. Interacting with buttons and browsers and UIs is so inefficient and much slower than say if you had direct API access, and very often you don't.