Meta’s Latest model is changing what it means to be open-sourced
Meta, the innovative technology company, has taken a bold step towards open-source AI by introducing its latest breakthrough: LLaMA 2, a large language model now available to the public for free usage.
In the race to challenge the dominance of OpenAI’s renowned AI chatbot ChatGPT, various tech companies have been racing to unveil their AI models. Meta has been moving at a steadier pace, previously releasing a smaller version of LLaMA, limited to researchers. However, they are aiming to catch up with the competition by launching LLaMA 2 and offering it as a foundation for commercial products, open to everyone.
This move marks the release of a suite of AI models, including different-sized versions of LLaMA 2. They are also introducing an AI model for chatbot integration, similar to ChatGPT. Unlike the latter, which users can access through OpenAI’s website, LLaMA 2 requires downloading from Meta’s partner platforms, such as Microsoft Azure, Amazon Web Services, and Hugging Face.
Nonetheless, some concerns persist. Meta has chosen not to disclose the specific data set used to train LLaMA 2, making it unclear whether copyrighted content or personal data was involved.
Furthermore, LLaMA 2 shares the same challenges as other large language models, such as generating false information and offensive language.
The introduction of LLaMA 2 poses a potential threat to OpenAI’s standing. Although Meta acknowledges a performance gap between LLaMA 2 and OpenAI’s GPT-4, this open-source model could empower companies to develop products and services more rapidly than their proprietary counterparts.
Preparing LLaMA 2 for launch involved extensive adjustments to enhance its safety and reduce the likelihood of propagating harmful falsehoods, in contrast to its predecessor.
To address these concerns, Meta utilized various machine learning techniques to improve the model’s helpfulness and safety. This included training the model on a more extensive data set, which was sourced from online scraping and fine-tuned based on feedback from human annotators to promote desirable behavior. Importantly, Meta asserts that LLaMA 2 does not employ any user data from its own platform and has excluded data from sites containing personal information.
Despite these efforts, LLaMA 2 may still produce offensive or problematic language, as is common among similar AI models. Meta has chosen not to eliminate toxic data from the training set, as this data could enhance LLaMA 2’s ability to detect hate speech, while removing it might inadvertently exclude certain demographic groups.
Meta’s Latest model is changing what it means to be open-sourced was originally published in Artificial Intelligence in Plain English on Medium, where people are continuing the conversation by highlighting and responding to this story.
https://ai.plainenglish.io/metas-latest-model-is-changing-what-it-means-to-be-open-sourced-728b76c573ed?source=rss—-78d064101951—4
By: Ronit Batra
Title: Meta’s Latest model is changing what it means to be open-sourced
Sourced From: ai.plainenglish.io/metas-latest-model-is-changing-what-it-means-to-be-open-sourced-728b76c573ed?source=rss—-78d064101951—4
Published Date: Fri, 21 Jul 2023 12:16:58 GMT
