Meta teamed up with Microsoft to challenge the large-scale model pattern. The latest Llama 2 is free and open source, and can be used directly for commercial use

Source: Titanium Media App, Author: Lin Zhijia

(Image source: Meta official website)

Overnight, the pattern of artificial intelligence (AI) big models has undergone tremendous changes again.

**Titanium Media App learned that in the early morning of July 19th, Beijing time, the technology giant Meta (formerly Facebook, NASDAQ: META), a technology giant with a market value of nearly 800 billion US dollars, released the latest AI large language model Llama 2 series, including 7 billion and 13 billion And 70 billion three parameter versions, the model information and starting code are all open source, and support free commercial use, and also include the Llama 2-CHAT model version fine-tuned for conversational chat scenarios.

Compared with Llama 1 released in February this year, the token used for the latest Llama 2 training has doubled to 2 trillion. According to Meta, Llama 2 outperformed all open-source Chat models in multiple external benchmark scoring tests including reasoning, coding, proficiency and knowledge tests.

At the same time, Meta also announced this morning that it will cooperate with Microsoft and Qualcomm. Llama 2 will be deployed on Microsoft's cloud service Azure, and Llama 2 will be able to run on Qualcomm chips, breaking the AI computing power market between Nvidia and AMD processors in the market. monopoly.

“Microsoft is our partner of choice for Llama 2 and expanding our efforts in generative artificial intelligence (AIGC).” Meta believes that an open approach is the right one for AI model development today, especially those in rapidly evolving technologies. large model. The company also argues in the paper that Llama 2 may be a suitable replacement for the closed model, based on Meta's human assessment of usefulness and safety.

Microsoft CEO Satya Nadella and Meta CEO Zuckerberg (Source: Zuckerberg ins)

Yann LeCun, winner of the 2018 Turing Award and Meta's chief AI scientist, tweeted bluntly, "This will change the landscape of the large language model (LLM) market."

Meta challenges OpenAI: Chat performance is stronger and safer, and the whole is close to GPT-3.5

For more than a decade, Meta has not only played in the "Metaverse", but has always had a layout for AI technology research, especially its Instagram, Facebook and other social products urgently need algorithms to bring advertising support.

In February of this year, Meta released the AI large model Llama 1. According to the company, researchers are shocked by the huge demand for Llama 1, with more than 100,000 access requests, and more developers have achieved amazing achievements by building model applications on Llama. However, Llama has not been free for commercial use because of the open source agreement.

Five months later, on July 18 local time, Meta released the open-source large-scale model Llama 2, which is not only free for commercial use, but also published a paper-style technical report, fully demonstrating the technical capabilities and evaluation information of Llama 2.

(Note: Paper address: ; Project address:

It is reported that Llama 2 is an updated version of Llama 1. Compared with Llama 1, Llama 2's pre-training model is trained on 2 trillion tokens (text data units), which is 40% larger than Llama 1's data size (1.4 times increase), and the context length is also doubled, and uses Grouped Query Attention (GQA) mechanism, while the fine-tuned Chat model is trained on 1 million human labeled data.

Technically, the new Llama 2 uses the optimized autoregressive Transformer framework, and of course supports PyTorch, performs more robust data cleaning, and updates mixed data.

Currently, Llama 2 series models are 7B, 13B and 70B three. And Meta AI said that they are still training the 34B version, but it will not be released for the time being. Meta's fine-tuned version of the large model LLAMA 2-chat is based on the fine-tuned version of LLAMA 2 for dialogue scenes. It also supports three versions of 7B, 13B and 70B parameters, using RLHF and supervised fine-tuning (SFT) and other techniques to iteratively optimize the model.

According to multiple evaluation information displayed by Meta, as of now, the performance of Llama 2 70B is closer to GPT-3.5, equal to or better than Google PaLM (540B), but Meta also admits that Llama 2 is far from GPT-4 and Google PaLM- There is still a large gap in the performance of leading non-open source models such as 2-L.

"But for many scenarios, you don't need GPT-4," said Percy Liang, director of the Stanford Center for Fundamental Modeling Research on the Alpaca research team.

In addition, the paper also mentions the security efforts of Llama 2. Among them, Meta uses techniques such as supervised security fine-tuning, secure RLHF, and security context distillation in security fine-tuning, and uses three commonly used benchmarks to evaluate the security of Llama 2. The overall performance is good, and it is more stable and safer than other large language models.

Microsoft "stepping on two boats": not only marrying Meta, but also raising the price of the GPT version of office

Microsoft, which is leading this wave of AI, finally revealed its new strategy at today's Global Partner Conference (Inspire 2023).

On July 18, local time, Meta and Microsoft jointly announced that starting today, Meta's Llama 2 will be available on Microsoft's Azure and Windows platforms.

Microsoft CEO Satya Nadella said that Llama 2 will allow Microsoft Azure developers to build on it and leverage their cloud-native tools for content filtering and security features. It is also optimized to run natively on Windows, providing a seamless workflow for developers.

Meta said that Llama 2 will be launched on Microsoft's Azure cloud, and will be distributed on other platforms such as Amazon AWS and Hugging Face in the future, and the cloud service based on the Llama 2 model will be released to developers around the world.

Microsoft CEO Satya Nadella

**So, there are currently two large-scale model options on Microsoft's development platform: not only OpenAI's GPT series large-scale model, but also Meta's Llama 2 large-scale model. **

At the same time, Microsoft also announced on Tuesday that Microsoft 365 Copilot will be priced at $30 per user per month for subscription merchants of Office 365 E3, E5, Business Standard Edition and Business Advanced Edition. According to the pricing on the official website, these subscription services for business users range from $12.5 to $38 per month per user, which means that Microsoft has increased the pricing of Office services for business users by 1-3 times.

Yusuf Mehdi, corporate vice president of Microsoft and chief marketing officer of the consumer sector, said that in the past few months, 600 companies around the world have participated in the early access program of the product, including well-known institutions such as KPMG and the National Development Bank of the United Arab Emirates. He emphasized that the more customers** using Copilot, the more enthusiastic they become about Copilot. Soon, no one will want to work without it. **

This is enough to show the pricing power and attractiveness of Microsoft and OpenAI to realize Office AI products in the market.

Microsoft has revealed that as of the first quarter of 2023, a total of 382 million people worldwide are using the commercial version of Microsoft (Office) 365 software.

However, Microsoft did not announce the official launch date of the product today. In addition to Microsoft, Google also announced earlier this year that it will integrate AI functions into Google Workspace, a cloud office suite. Internet software service providers in subdivided industries, including Zoom, are also rushing to launch AI functions.

It is worth noting that the chip giant Qualcomm jointly announced this morning that it plans to support AI deployment based on Meta Llama 2 on flagship smartphones and PCs from 2024, enabling developers to use the AI capabilities of the Snapdragon platform and launching a new generation AI applications.

Microsoft stock daily chart (Source: TradingView)

Affected by the new news, as of Tuesday, Microsoft's share price closed up 3.98%, hitting a record high. Since 2023, Microsoft's cumulative increase has reached 50.05%.