Meta重大发布!挑战OpenAI

On April 18, local time, Meta Platforms released an early version of the latest large language model, Llama 3, on its official website, which contains an image generator that can update pictures in real time when users enter prompts. The Llama 3 model will be integrated into the virtual assistant Meta AI, which users can use in Meta's Facebook, Instagram, WhatsApp, Messenger programs, as well as a new standalone website, allowing it to compete more directly with Microsoft-backed ChatGPT.

"The largest version of Llama 3 is currently being trained with 400 billion parameters and has already received a score of 85 MMLU, a scoring metric that measures the capabilities of large language models," said Zuckerberg, founder and CEO of Meta, citing metrics used to convey the strength and quality of performance of AI models. ”

MetaAI, a portable assistant, enters the user's life

According to Meta's official website, the latest Llama 3 model will be integrated into the virtual assistant Meta AI, which allows users to conduct direct searches on Facebook, Instagram, WhatsApp and Messenger applications without having to switch between apps. Let's say a user is planning a ski trip in a Messenger group chat, using the search function in Messenger, they can ask Meta AI to find flights from New York to Colorado and find out which weekends are the least crowded, all without leaving the Messenger app.

(Laiyuan/Meta)

According to the mobile phone operation screen shown by Meta, users can directly @Meta AI in the chat app when planning a trip with friends, helping customers find a picnic spot where they can see the sunset in San Francisco, or find the best place to see cherry blossoms when traveling in Japan.

Chris Cox, chief product officer at Meta, said: "The ultimate goal is to help customers get rid of their worries and make their lives easier, whether it's helping with work, writing, or planning travel. ”

It is understood that Meta is speeding up image generation, and users can use Meta AI's Imagine feature to create images from text in real time, and a beta version of the feature has been launched within WhatsApp and Meta AI programs.

"When a customer starts typing, they see an image appear, and with each time they type some text, the image changes," the Meta team said. The resulting images are also now sharper, higher quality, and better able to include text in the images. At the same time, the Imagine feature can provide useful tips as well as ideas for changing the image, so customers can constantly optimize their ideas. ”

The Meta team has also launched a separate website with the latest Llama3 model. Similar to OpenAI's product ChatGPT, users can log in to Meta AI's web version while using a computer. "We want our customers to be able to use Meta AI when they're doing their work on their computers, so we're launching Meta AI on the web today, so users don't have to worry about solving math problems anymore," the Meta team said. ”

Chris Cox said in an interview: "The Meta team has equipped Llama 3 with new computer coding capabilities and provided it with images and text, but currently the model can only output text, and future versions will provide more advanced inference capabilities, such as the ability to make longer multi-step plans." ”

Seek cooperation to accelerate AI development and challenge OpenAI's leading position

It is understood that this time Meta also announced a new partnership with Google to include Google's real-time search results in Meta AI's response. Google also said on Thursday local time that it would move its Responsible AI team, which focuses on the development of secure artificial intelligence, from the research department to the Deep Mind department so that they can build and scale AI models.

Chris Cox said image training will be added to Llama 3 to enhance the update to the Ray-BanMeta smart glasses launched this year, which has partnered with European eyewear maker Essilor to enable MetaAI to identify the wearer and answer questions about them.

Not only is there a collaboration between Google and Essilor, but at the same time, Nvidia also announced on its official website that all of its platforms will be optimized to accelerate the Meta Llama 3 model, which will be combined with NVIDIA accelerated computing to enable developers, researchers, and enterprises to innovate responsibly in a variety of applications.

According to NVIDIA's official website, the version of Llama 3 accelerated on NVIDIA GPUs is now available and can be used in the cloud, data center, edge computing, and personal computers. Developers can try out Llama 3 on the web via a browser, packaged as an NVIDIA NIM microservice with a standard application programming interface that can be deployed anywhere. In addition, NVIDIA RTX and GeForce RTX GPUs for workstations and computers accelerate inference on Llama 3, which provides developers with the goal of more than 100 million NVIDIA accelerated systems worldwide.

According to Meta's official website, with the update, Meta AI assistant is expanding to more than a dozen markets outside the United States, including Australia, Canada, Singapore, Nigeria and Pakistan. As for the European market, Chris Cox said: "Meta is still looking for the right approach in Europe, where privacy rules are stricter, and the upcoming AI bill will introduce requirements such as the disclosure of model training data. ”

400 billion parameter masterpieces, free and open without fear of challenges

The huge demand for data for generative AI models has become a major source of tension for the development of this technology. Meta has been releasing models like Llama3 for developers to use commercially for free, a strategy that has also drawn voices from some critics who fear that unscrupulous developers might use the model to build content.

Zuckerberg acknowledged the competition in a video accompanying the website, in which he called Meta AI "the smartest AI assistant you can freely use." Zuckerberg said that the largest version of Llama 3 is currently being trained with 400 billion parameters and has already received a score of 85 MMLU, citing metrics used to convey the strength and quality of performance of the AI model. "The two smaller versions currently being launched have 8 billion parameters and 70 billion parameters, with the latter scoring around 82MMLU," he said. ”

It is understood that earlier developers complained that the previous version of the Llama 2 version of the model could not understand the basic context, while Google experienced similar problems and recently suspended the use of its Gemini AI image generation tool after being criticized for its inaccurate depictions of historical figures.

Meta's team says the latest Llama 3 will reduce some of the previous problems by using "high-quality data" to allow the model to identify nuances. But the Meta team didn't elaborate on the dataset used, other than to describe how much data is fed in Llama 3 than in Llama 2, and uses "synthetic" or AI-created data to enhance areas such as coding and inference. Chris Cox said there has been "no significant change in attitude" in the way companies get their data. ”

Editor-in-charge: Ye Shuyun

Proofreading: Peng Qihua

The 2024 Securities Times "Looking for the 'Golden Eagle' of Venture Capital and Discovering the 'New Seedlings' of Enterprises" plan was grandly reopened, lighting up the dream of the times and blooming the light of the times.

Click on the poster to register, and quickly poke ↓↓↓

Copyright Notice

All original content on the platforms of the Securities Times shall not be reproduced by any unit or individual without written authorization. Our company reserves the right to pursue the legal responsibility of relevant actors. For reprinting and cooperation, please contact the Securities Times assistant, WeChat ID: SecuritiesTimes

END

Click on a keyword to view it

Periscope Series In-depth Report丨Stock Council Column丨Investing in Xiaohongshu丨E Company Investigation丨Times Meeting Room丨Top Ten Celebrity Private Equity Interview丨Another A-share Company, Locked in Delisting!丨Supervision 11 Consecutive Questions!丨Bad Guys" Recommend "Gatekeepers", This A-share, Three Months Two Changing Accounting Firms, What Is the Secret?丨Lei Jun: There is no cold 4 billion in the card! Xiaomi SU7 sales are 3 to 5 times higher than expected, "It's too early to talk about success"!丨A shares, a big explosion! The Shanghai Index broke through 3100 points in intraday trading丨 The second is empty! Huawei Pura 70 suddenly went on sale!丨Lithography machine giant exploded!Nasdaq crashed丨Will Musk's 400 billion salary be paid?Tesla's hard court!丨Just now!Xiaomi's vice chairman donated 100 million yuan!丨The National Development and Reform Commission spoke out!In response to the question of overcapacity, the latest arrangements for ultra-long-term special treasury bonds...... Huge amount of information