Google has gone open source with two Gemma family lightweight generative artificial intelligence models built from the same research and technology used to create the Gemini models. Gemma 2B and Gemma 7B also come with pre-trained and instruction-tuned variants.
The pre-trained and instruction-tuned Gemma models can run on laptop, workstation, or Google Cloud.
That move follows Meta’s open sourcing of its LLaMA (Large Language Model Meta AI) in 2023 for approved researchers and organizations. In February 2024 Meta released a commercial version of LLaMA 2, making it freely available for both commercial and non-commercial use. This includes access to model weights and starting code for pre-trained and fine-tuning purposes, but not detail on the full training data used to create the model.
Google compares Gemma to LLaMA.
Meta also contributes to the OPT (Open Pre-training Transformer) open source language model.
As one might expect, controversy exists over the release of open source models, as some argue this will encourage use by bad actors generating harmful content, spreading misinformation, or launching cyberattacks.
Supporters argue that open source AI provides transparency and reproducibility, enabling researchers and developers to understand how the models work, identify potential biases, and contribute to their improvement.
And some arguments might support either point of view. On one hand, potential bad actors might not have access to the resources to create or train their own models. So open source might enable them.
On the other hand, the same resource limitations also inhibit use by individuals and firms who might be able to create useful capabilities.
Some might note that all technology can be used for good or evil.
No comments:
Post a Comment