[ad_1]
This articleauthorChe Pinjueis a director of Hong Kong Science and Technology Parks Corporation, a visiting associate professor at the School of Chinese Business at the University of Hong Kong, a senior consultant at Alibaba Cloud, and writes a column for Hong Kong Economic Journal“Big Data for All”.
Meta released the latest open source large language modelLlama 3.1 SeriesThis series of models includes three parameter sizes: 8B, 70B, and 405B. The 405B parameter model performed well in many benchmarks, surpassing OpenAI’sGPT-4oand is comparable to other leading closed-source models, such asClaude 3.5 SonnetMeta founder Zuckerberg said that Llama 3.1 version is a turning point in the industry, indicating that open source artificial intelligence (AI) will become mainstream in the future.
In this release, the Llama 3.1 model has not only expanded in scale, but also increased the size of the context window from the original 8K to 128K, a 15-fold increase, while supporting 8 languages. The 405B model in particular, which was trained with more than 15 trillion tokens and used 16,000 H100 GPUs (graphics processing units), is the first model to reach such a scale. After evaluating more than 150 benchmark datasets, the Llama 3.1 405B model’s performance in common sense reasoning, operability, mathematics and other tasks is comparable to GPT-4, GPT-4o and Claude 3.5 Sonnet. At the same time, the smaller 8B and 70B models perform on par with other open source and closed source models of the same size.
In actual scene applications, the overall performance of the Llama 3.1 405B model is better than GPT-4o and Claude 3.5 Sonnet. Meta has also updated the open source license to allow developers to use the output of the Llama model (including 405B) to improve other models. Although the image, video and voice integration functions are still under development and have not been officially released, Meta said that these features will be integrated in future versions.
I strongly agree that open source AI can promote innovation, reduce costs and improve security. Developers can also use open source models to train and fine-tune their own models to meet different needs.
In addition, open source models are cheaper to use and more efficient, especially when running inference tasks, the cost is about half of that of closed models, which enables developers to operate on their own infrastructure and enhance data security.
Open source AI represents the world’s largest economic opportunity and security guarantee, and can also create greater economic value and a higher level of global security. To date, all versions of the Llama model have been downloaded more than 300 million times, and its popularity and influence are self-evident. As for the battle between open source and closed source big models, it actually depends on the degree of data openness and computing efficiency.
Support EJ Tech
If you want to submit articles, report information, issue press releases or interview notices,Click here to contact us.
[ad_2]
Source link
