AI - AI Dynamics

→ View original post on X — @dair_ai

PaLM 2 Excels in Multilingual QA and Mathematical Reasoning

By

@dair_ai

–

04 October 2023 18h25

Curious about the capabilities of PaLM 2? In our new paper explainer, we summarize the main contributions of PaLM 2, a model that excels in various tasks like multilingual question answering and arithmetic reasoning. PaLM 2 mathematical reasoning performance surpasses SoTA (at

BNB Quantized Layers Initialization Performance Issues

By

–

04 October 2023 18h25

The issue is the creation of the bnb quantised layers, not the adapters. (They are fast to init anyway since they're small.)

WhatsApp’s Engineering Challenges: Lessons Beyond ML Data Engineering

By

–

04 October 2023 18h24

Consider what Whatsapp achieved prior to acquisition with their team. And their eng challenges were many orders of magnitude harder than ML data eng!

RLHF Training with Small Teams and Data Engineering

By

–

04 October 2023 18h23

I've seen plenty of folks doing RLHF with <5 dedicated people! (And I do all my own data engineering — it doesn't take too long after a few decades of practice…)

→ View original post on X — @juanmerodio

AI-Powered Olympics: Technology Replaces Human Athletic Competition

By

@juanmerodio

–

04 October 2023 18h19

Esto que hoy es un CGI, en pocos años podría volverse una realidad.

Te interesaría ver unos JJOO basados en la innovación tecnológica en lugar de en las capacidades humanas?

Yo personalmente prefiero el deporte tradicional, pero una sana competencia entre países aceleraría… pic.twitter.com/nkShPeNoRB
— Juan Merodio (@juanmerodio) 4 octobre 2023

Esto que hoy es un CGI, en pocos años podría volverse una realidad. Te interesaría ver unos JJOO basados en la innovación tecnológica en lugar de en las capacidades humanas? Yo personalmente prefiero el deporte tradicional, pero una sana competencia entre países aceleraría

Model Sharding Challenges for Large GPU Deployment

By

–

04 October 2023 18h16

Not sure if there's a way to do that with model sharding — which is necessary if your model is too big to fit on the GPU.

Hugging Face Performance Fix Coming Soon Official Update

By

–

04 October 2023 18h08

I chatted to the Hugging Face folks and they tell me an official fix for this performance issue is on the way! 😀 So hopefully my hacky workaround will be redundant soon…

PEFT Kaiming Init Optimization Patch for Training

By

–

04 October 2023 18h08

The issue is that PEFT spends all that time doing unnecessarily doing kaiming init. So our code first patches the `init` module so that `kaiming_uniform_` does nothing at all. Here's a gist with the full code:

Python contextlib.contextmanager decorator explained

By

–

04 October 2023 18h08

BTW this is a great example of the use of the handy `contextlib.contextmanager` decorator. Check out the docs if you haven't used it before: https://
docs.python.org/3/library/cont
extlib.html#contextlib.contextmanager
…