wow didn't know this was happening. this is huge!
scikit added support for pytorch and GPUs via array dispatch
@soumithchintala
-
Scikit-Learn Adds PyTorch GPU Support via Array Dispatch
By
–
-
Ex-PyTorch Leaders Launch ML Infrastructure Startup
By
–
a few ex-PyTorch folks started an ML Infra company last year, and now they have a twitter account.
also, @jiayq the infamous Caffe author and PyTorch 1.0 co-lead just left Alibaba to start an ML Infra company.
worth following to see what magic they will create. -
AGI Achievement Definition and Financial Implications Risk
By
–
you'll make a lot of money, unless what qualifies as "AGI achieved" is scammy
-
GPT-4 Remains Most Capable AI Model Despite Competition
By
–
wdym by "they dont present a business threat"? GPT4 is still the most capable model around by a mile.
-
Punk Credit Attribution for ChatGPT Generated Content
By
–
Punkier way of crediting something written by ChatGPT: – anonymous individuals on the internet mishmashed by a Transformer, 2023
-
Missing AI meme accounts: discovering ML humor on Threads
By
–
Wondered why my Threads feed was incomplete.
And then I realized i was missing all the vibe accounts…
I've started enjoying the ML memes too much!
@untitled01ipynb @typedfemale @kipperrii @tszzl @repligate -
Llama.cpp and Activation Checkpointing Optimization Techniques
By
–
yea, llama.cpp is it for the llama-style models.
About activation checkpointing — the way the LLM people do it is to intentionally sidestep autodiff and hand-write a clever backward. Sometimes people even do mathematically approximate stuff. -
Beyond Homogeneous Transformers: Loss Functions and Inference Optimization
By
–
i think your position comes from a fundamentally incomplete estimation of the parts of the graph that are not a homogeneous transformer — the loss function, the beam search, things like activation checkpointing that isnt simply autodiff, token fusion style tricks. caffe style
-
Open-Source AI Control Versus Commercial Partnership Constraints
By
–
I find this as a tangible example of why we'd want to retain control of our AI Assistants, most obviously via open-source.
All commercial offerings will eventually cripple the assistant in ways that would make their partners and an approximation of stakeholders happy (partners -
CUDA PyTorch AMD hardware compatibility for AI computing
By
–
I think the details matter a lot.
CUDA enables pretty general programming and experimentation — AI, fluid dynamics, biochem, various scientific simulations.
PyTorch itself is used in the widest variety of settings, including all of the above.
My main position around AMD working