IMO apple will not call gpt3 because NIH and expense — and the local models (1B to 3B range) won't be as good. I think they'll build their own large model and local models and use them smartly (auto-routing or something). Maybe they'll fork Mistral's latest.
@soumithchintala
-
Database Benchmark Restrictions: Legal Implications Oracle
By
–
i remember some databases had a no-benchmark license / terms of service. Benchmarking those databases was illegal / grounds to sue. Oracle was one. https://
danluu.com/anon-benchmark
/#:~:text=Overtheyearsmanymajor,forbidsbenchmarkingoftheirdatabase
…. -
Apple’s AI Features: Two Years Behind, Mediocre, Yet Heavily Marketed
By
–
if the recent iPhones are any trend, they're going to be 2 years behind, mediocre but well-integrated, and timmy is going to announce on stage that its a breakthrough
-
Rising Stakes: AI Investment and Financial Implications
By
–
yep, except now stakes are higher as there's crazy money involved
-
Anyscale Benchmark Flaws Raise Concerns About Research Quality
By
–
it pains me to see a poorly constructed benchmark coming from a credible source. I hope @anyscalecompute fixes things, and also consults other stakeholders before publishing such benchmarks. If I didn't know Anyscale closely, i would have attributed bad faith.
Lets dive into -
New Tools to Detect Subtle Code Issues Easier
By
–
@gdb i think you'll like this post. more tools to detect subtle but annoying issues easier.
-
Neural Network Resilience Through Summation and Dense Connectivity
By
–
I follow the summations and network connectivity to understand resilience to information loss.
When each output is the sum over a large set of input computations, which themselves densely cross-connect in previous layers — I just expect information to be really resilient to -
Fastest BS=1 Inference with PyTorch Speculative Decoding
By
–
I think this breezes past as the fastest bs=1 inference I know.
With plain and simple PyTorch code.
Speculative Decoding isn't even applied, and will be purely additive… -
History of Deep Learning Software: A Mini-Book Project
By
–
Tianqi, you and I should probably compile a "History of Deep Learning Software" for fun 🙂
it'll be a mini-book, if only for us to keep in our library. -
PyTorch Design Origins: Historical Context and Architecture
By
–
Mirrored here for resilience and googleability: https://
soumith.ch/posts/2023/12/
pytorch-design-origins/
…