Google AI and DeepMind News and Discussions

[wjfox]

[spryfusion]

[wjfox]

[spryfusion]

[spryfusion]

[weatheriscool]

Google Announced 'Gemma' Open Source AI Models
Google says the first two Gemma AI models are more capable than they may first appear.
By Ryan Whitwam February 22, 2024
https://www.extremetech.com/computing/g ... -ai-models

Google has a history of supporting open source and free software projects—the entirety of the Android source code is open, for example. The tech giant is playing things a bit closer to the chest as it unifies its artificial intelligence projects under the Gemini brand. Google doesn't want to give away Gemini, but it has announced a lightweight AI model that it will open up. It's called Gemma, and you can start experimenting with it today.

According to Google, Gemma is a family of large language models (LLMs) based on the same technology that powers Gemini. There are two versions of Gemma, known as Gemma 2B and 7B, aimed at completing simpler tasks without the hefty processing requirements of Gemini. The scale of a machine learning model can be illustrated by the number of parameters it contains. More complex models need all the additional parameters to enable more complex behavior, but quality can matter as much as quantity. The 2B and 7B labels indicate the approximate number of parameters—2 billion and 8 billion—much less than competing models like Llama 2 (70 billion) and GPT-4 (1.76 trillion).

Despite the smaller parameter count, Gemma won't behave like a lightweight—Google has shared some benchmark numbers that show Gemma beating the much larger Meta Llama2 model. Don't expect performance like the latest Gemini models, though. Gemma is best suited for basic tasks like summarization and chatbots, and it's optimized for just English right now.

Gemma is a pre-trained model, but unlike Gemini, there's no website where you can just fire up Gemma and start talking to it. You'll need to deploy one of the models using a supported platform like Hugging Face, Kaggle, Nvidia NeMo, or Google Vertex AI. Google is offering a $300 credit to help new users get started with Gemma on Google Cloud. Researchers interested in using Gemma can apply for up to $500,000 in credits. Google also partnered with Nvidia to optimize the models for RTX GPUs, so you can run Gemma on a PC with a sufficiently powerful graphics card.

[spryfusion]

[weatheriscool]

[Powers]

I don't have very high hopes on Google about this.

[Vakanai]

Welp this should knock down Google's chances for an AGI model by about a decade or four.