Google의 Gemma: AI의 접근성과 협업성 강화

Google의 AI DevRel 팀의 Gus Martin은 개발자와 커뮤니티가 최대한 쉽고 편리하게 AI 도구를 사용할 수 있도록 하는 것을 목표로 합니다.

### Google의 Gemma: 쉽고 접근하기 쉬운 AI 모델 만들기
* 🤖  젬마 개발자의 날 행사에서 보여준 것처럼 Google의 AI DevRel 팀은 커뮤니티가 AI 도구를 쉽고 편리하게 사용할 수 있도록 최선을 다하고 있습니다.
* 📦  Gemma 모델은 상업적으로 친숙하게 제작되어 다양한 플랫폼에서 사용할 수 있으며, ai.google.dev/gemma 및 Kaggle 모델을 통해 쉽게 액세스하고 특정 라이선스를 사용할 수 있습니다.
* 💻  개발자는 선호하는 프레임워크에 관계없이 Kaggle 모델의 추론 위젯을 통해 Gemma 모델과 상호 작용하고 코드에서 단 몇 줄로 쉽게 구현할 수 있습니다.
* 📚  Gemma 문서는 추론, 미세 조정, 고급 솔루션 등 개발자가 AI 여정을 시작할 수 있는 튜토리얼과 리소스를 제공합니다.
* 🏆  Google은 개발자가 문제를 해결하고 커뮤니티에서 학습할 수 있도록 Gemma 모델을 기반으로 하는 Kaggle 대회를 제공하며, 클라우드 크레딧을 통해 연구자를 위한 추가 지원도 제공합니다.

[MUSIC PLAYING] GUS MARTINS: Thank you very
much for spending your Thursday afternoon with us. I really hope that,
by the end, you feel like, oh, this was amazing. It was really worth my time. Again, I'm Gus. I'm from the AI DevRel team. My main goal is to make Google's
AI tooling as easy as possible for you, for the community. And today, you've seen
a lot of-- there's talk about the vision
behind Gemma, how cool is it, the [INAUDIBLE] tool
kit, why it's an open model. But it's the Gemma
Developer Day, so I imagine you
are all thinking, when's the Developer Day? And it's now, exactly now. So we are going to
see a lot of demos. My colleagues there will
show you a lot of live demos, so let's all pray that
everything goes well.

And we'll show you lots of the cool capabilities we can do, but before we get there, I would like to teach you how to start this journey. I imagine some of you have started already. Thank you very much for that. And if you didn't, welcome. This is exactly where you should start. I'll give you all the tips for that. But before we start that, when I start working on this project for some time ago, one of the goals that we had-- we had three goals. We had many goals, but the three important ones for you are, we wanted to make models that were very, very, very good as we did that. You saw the numbers. They're very good. The other thing, we wanted to make those models safe and follow-- and following all the responsible AI perspective. We did that, too. But the one thing that I was a little bit annoying to my friends, every day, all the time.

is I want these models
to be as easy as possible to use and to be everywhere. I want them to be in every
platform for everyone to use, for everyone
to benefit from it. And that's what we did,
so we launched them in all our platforms, our
tooling, our frameworks, AI related, of course,
and our cloud. We collaborate with our friends
Hugging Face, from NVIDIA, to make them even more
available, even faster, and this was very good. We were very happy with
the beginning of that, and we gave to the community
everything that we thought, and that was very good. And two days after that,
we saw that the community got the models and added
another 20 other frameworks and platforms. Many of those that I wasn't-- I didn't know they
were so popular, but the community
just ran with it. They just got it, start
putting other platforms, sending PR requests to
other GitHub repositories.

More than 300,000 downloads,
more than 300 spaces on Hugging Face, so it became very popular. And I think maybe a part
of it is because we tried to make this easy for you. The idea is it has to be
easy for the developer. I know I received questions
from people, hey, Gus, I really want to start this
machine learning journey, and then it's like, yes,
I understand your pain, and I imagine you all
had the same question. And where should I start? That's the follow-up
question, which is very hard, and we wanted to make
this journey easy. That's why we did this. And to start with
Gemma, you will go to the ai.google.dev/gemma,
where is the first page where you can see all the benchmarks,
all the information. The main documentation is
there, and that's where you should start your journey. But Gemma has a
specific license.

As mentioned before, it's not an open-source license. It's an open license. It's a commercially friendly license, and you have to go accept this. And the main place to do that is on Gemma's page on Kaggle Models. You can go there, and you can accept the license. You can use your Kaggle user. You can use your Hugging Face user, your Gmail, another email account, any of those, and then you get access to that. So when you see that screen, you are able to start playing with the model, and as a developer, as I mentioned, that's what you want to try. And before we go into trying the model and doing any code whatsoever, there's another cool thing you can do is, too, on Kaggle Models, which is we have an inference widget where you can play with the model right there on the web page.

Of course, this is not running on the web page. We will hear about this a little bit later today. This is a widget that was created by our colleagues on Google Cloud, and you can play with the model, try some prompts, see the results there. This is a good start for you to, oh, let me see what this model can do, before going to any code related environment. But then you say, OK, Gus, this is the developer day. You told us a couple of minutes ago. Yeah, yeah, yeah, got it. Got it. So let's start with the basic one. We have the Keras implementation. If you read here the box to the left, there's basically two lines of code there. There's some import, some styles thing that you set up, but it's two lines of code. You load the model, and you call the generate method.

and that's it. With that, you can start
playing with the model with code right there. Two lines of code, so
it's very straightforward. Of course, if you are
not a Keras developer, you are, I don't know, PyTorch
developer, or Jax developer, we also have the same way
of doing that, of course, for specific PyTorch,
specific for Jax. But the idea is you
get this quite fast. And when you are ready
to continue your journey, you got the license,
so now, what's next? Great question. You go to our documentation. We have a lot of tutorials. We added two new
tutorials be-- if you saw the documentation yesterday,
now, it's even better. We have more tutorials that
were added this morning, and you can continue
your journey. So you can play with inference. You can do fine-tuning. You can do fine-tuning
multiple hardware.

You can play with other configurations. You can understand how the details about the model, how to do other advanced solutions, all of that in our documentation. Now, I give you the first steps, and you can start playing with that. You can start building your own applications. You can start having fun. But one question that I got from the developers, from you people, from the community is, Gus, can you point me to a project that I can start playing with? How do I start? I played with the tutorials. I understand how it works. Now what? Don't worry. We got you covered. We have two Kaggle competitions running right now based on the Gemma models. We have the AI Assistant competition, and we have the LLM Prompt Recovery. Both of them are competitions based on the Gemma models, and the idea is you try to solve a problem using the Gemma models, and you compete against other teams.

You can create a team. You can compete by yourself. But why does this so cool? The idea is, when-- as you've seen, when we start to learn machine learning, there's a lot to learn, so it's very hard to do it by yourself. When you start to use Kaggle, you have a community of people, a community of people trying to do very similar things as you. They are all competing there. They're trying their best to solve a problem. They will share in the forum. They will share code. They will share their ideas, and you can learn from them. So maybe you're not an expert now, but you can learn from the experts. They are not going to give you the secret sauce at first, but they will give you a good start. So there's a lot of great code there that you can go and start learning with it, so this community is amazing.

I highly suggest you start playing with it. And there's also a prize, but the idea is learning. I'm pretty sure that learning is more important than the prize. And of course, if you are a researcher, we also got you covered. We have a program to give you credits, cloud credits, so you can do your researches using Google Cloud and Gemma because we don't want-- the model is [INAUDIBLE] we say developers. Maybe I should have defined this earlier. When I mean developers, I mean developers, researchers, builders. I want everyone that wants to do something, create something, and researchers are a very important part for us. We would like to help you build new models, new capabilities. The last week, I saw people building stuff, and I was like, is that possible? And I've been playing with this model for a while. But people are-- I still get surprised by some things.

that people do, so it's amazing. We have this on our main page. You can see the link to
get more information. If you are a researcher, we
can give you more information also during the break. [MUSIC PLAYING]

Getting started with Gemma models

Subscribe to join a community of creative developers and learn the latest in Google technology — from AI and cloud, to mobile and web.

Explore more at developers.google.com



Google for Developers

Google AI DevRel 팀의 Gus Martins가 모든 플랫폼에서 AI 모델을 쉽게 사용하고 액세스할 수 있도록 한다는 목표를 강조하면서 Gemma와 개방형 모델의 비전에 대해 설명합니다.

Gemma의 출시는 단 이틀 만에 20개의 다른 프레임워크와 플랫폼에 통합되는 등 커뮤니티에서 광범위하게 채택되었습니다.

젬마는 30만 건 이상의 다운로드와 300개의 허깅 페이스 공간을 확보하며 큰 인기를 얻고 있습니다.

개발자들은 머신 러닝 여정을 더 쉽게 만들기 위해 ai.

dev/gemma에서 벤치마크와 설명서를 제공하고 있습니다.

Gemma는 상업적으로 친숙한 라이선스로 제공되며 Kaggle 모델을 통해 액세스할 수 있으므로 사용자는 웹 페이지에서 추론 위젯을 사용하여 모델을 개발용으로 사용하고 사용해 볼 수 있습니다.

Google Cloud의 모델 위젯을 사용하면 몇 줄의 코드만으로 머신 러닝 모델을 쉽게 탐색할 수 있습니다.

이 위젯은 간단한 시작을 위해 Keras 구현을 제공하며 PyTorch 및 Jax 개발자를 위한 지원도 제공합니다.

또한 추론, 미세 조정 및 여러 하드웨어 사용에 대한 추가 학습과 탐색을 위한 포괄적인 문서와 튜토리얼을 제공합니다.

현재 진행 중인 두 개의 Kaggle 경연 대회인 AI 어시스턴트 경연대회와 LLM 프롬프트 복구 경연대회는 Gemma 모델을 기반으로 하며, 참가자들이 이 모델을 사용해 문제를 해결할 수 있는 기회를 제공합니다.

이 대회는 개인 또는 팀 단위로 참가할 수 있으며, 머신 러닝에 비슷한 관심사를 가진 사람들로 구성된 커뮤니티에서 배우고 경쟁할 수 있는 특별한 기회를 제공합니다.

참가자는 포럼에서 코드, 아이디어, 토론을 공유할 수 있어 귀중한 학습 경험이 될 수 있습니다.

연구자 및 개발자를 위한 인센티브

구글 클라우드와 젬마를 사용한 연구를 위한 클라우드 크레딧을 제공하는 프로그램입니다.

학습과 창작에 중점을 두고 있으며, 부차적인 인센티브로 상금이 제공됩니다.

이 조직은 연구자, 빌더, 개발자가 새로운 모델과 기능을 구축할 수 있도록 지원하는 것을 목표로 합니다.