The energy consumption ratio reaches 10 times that of GPT-4! Huang Xuedong personally deciphers the success of Zoom AI, and the AI ​​companion can directly hold meetings for you

Recently, Huang Xuedong, the AI ​​tycoon and current CTO of Zoom, published a blog on his official website.

He grandly introduced his innovative achievements six months after joining Zoom - Zoom AI Companion, and the Zoom Federation that drives Zoom AI Companion ( federated approach) AI.

Before Zoom, Huang Xuedong served as the chief technology officer of Microsoft Azure AI. He is not only Microsoft’s first global Chinese technical academician, but also Microsoft’s highest-level Chinese scientist.

Huang Xuedong has worked at Microsoft for more than 30 years. With his help, Microsoft established its first Asia Research Institute in Beijing in 1998.

Huang Xuedong was elected as a fellow of ACM and IEEE in 2017 and 2020 respectively, and was elected as an academician of the National Academy of Engineering and an academician of the American Academy of Arts and Sciences in 2023.

Huang Xuedong said that the evolution and implementation process of artificial intelligence is like a winding road, full of constant destruction, adaptation and innovation

Over the past 30 years, I have been traveling on this path, dabbling in areas such as speech recognition, natural language understanding, and computer vision. In the six months since I joined Zoom, the speed of innovation has been particularly surprising

Zoom Federated AI, using the right AI to run the right task

Zoom hopes to use AI to improve the way humans collaborate and continuously increase productivity.

Zoom has been committed to providing artificial intelligence services such as speech recognition, computer vision, machine translation and large language models (LLM) for many years to enhance the competitiveness of its software

Zoom AI Companion is their latest product, which aims to help users be more productive, facilitate seamless collaboration, and gain deeper communication and insights

Zoom's federated AI enables users to enjoy the functions of AI Companion without spending extra money on the basis of previously paid services.

In this way, the user experience will be significantly improved, and at the same time, the participation of new technologies among users can be quickly increased

For Zoom software that has a large number of users and requires intensive use of AI functions, how is the additional cost of LLM inference shared?

According to rumors, GitHub’s Copilot is said to cost $10 per month, and each user needs to pay an additional $20 per month as a subsidy

The language model (LLM) currently used by Zoom includes its own model and third-party models, such as OpenAI GPT-3.5, GPT -4 and Anthropic Claude 2

Zoom Federated AI improves the end-to-end experience of Zoom customers by integrating the latest LLM technology from partners (such as OpenAI’s newly released GPT-4 Turbo)

Zoom adopts a more cost-effective strategy:

First, use a less expensive LLM to complete the task, then Evaluate the initial completion quality of the task through Z-Scorer

Zoom will adopt more advanced LLM if necessary to improve the results achieved by the initial LLM

- This is similar to a cohesive team working together to produce a higher quality product more efficiently than one person could.

According to Zoom’s internal testing, Zoom’s adoption of federated AI improves the output quality of AI Companion compared to OpenAI GPT-3.5 Turbo or several other state-of-the-art LLMs, and also effectively Controlled costs.

Huang Xuedong said that their criteria for measuring performance are to reduce costs, improve response speed and provide higher quality output

In the figure below, we can see the performance and cost comparison between Zoom AI Companion and GPT-4 on the English meeting query task

In this figure, we Two key metrics were compared, namely cost and output quality

Zoom AI Companion achieves almost equivalent output quality using less than 6% of the cost of GPT-4

Compared with the GPT-4-32k model used by Microsoft Copilot, Zoom AI Companion’s meeting question function can reduce costs and speed up response while maintaining comparable output quality.

The above picture is a comparison between Zoom AI Companion and GPT-4 on multi-language conference summary tasks.

In addition, the multi-lingual performance of AI Companion further proves the power of Zoom federated AI, supporting 32 languages ​​​​in addition to English.

Since most LLMs are pre-trained mainly using English data, Zoom added a translation model to expand multi-language capabilities.

In the multi-language meeting summary task, the above figure compares the performance of Zoom AI Companion and Microsoft Copilot. Both use GPT-4 as a proxy for Microsoft Copilot for comparison

In the Zoom AI Companion’s multilingual meeting summary, non-English text is translated into English using Zoom’s translation model, This allows both the translated data and the original data to be considered.

As can be seen from the figure, Zoom AI Companion completed output with a quality close to GPT-4-32k (97%) at less than 6% of the cost.

Zoom believes that the benefits of artificial intelligence should be made widely available to as many people as possible, and Zoom’s federated AI can seamlessly combine the advantages of different AI systems to provide high-performance results in plays an important role in realizing this vision. Rewritten content: Zoom believes that the benefits of artificial intelligence should be made widely available to as many people as possible. Zoom’s federated AI can seamlessly combine the advantages of different artificial intelligence systems to provide high-performance results and plays an important role in realizing this vision

Zoom AI Companion, the smartest meeting Assistant

Currently, almost all of Zoom’s artificial intelligence features have been integrated into Zoom AI Companion, which was launched in September last year

The "recharge service" provided by Zoom uses Zoom's artificial intelligence capabilities to automatically summarize the meeting content you missed after you join the meeting midway.

When it's your turn to speak, You can combine your own opinions and the content of previous meetings to generate a speech, realizing a true "meeting on behalf of users"

in the middle After joining the meeting, AI can tell you the key points of the meeting.

It can also help you transcribe and translate the speeches of other participants in real time, and easily obtain the text of the meeting discussion

The role of AI is not only to organize fragmented information, but also to generate meeting summaries and records for the team

Can summarize customer short messages obtained from the Zoom platform , and generate a specific execution plan

And for meetings that cannot be attended, AI intelligent recording can also be arranged directly on the schedule, and then segmented Look back at the key points.

Behind this entire AI companion, the technical ideas of Zoom and Huang Xuedong Federal AI are condensed, using appropriate AI resources to handle appropriate requests. .

And Zoom promises that it will not use any of the user’s audio, video, chat, screen sharing, attachments or customer-generated content (such as poll results, whiteboards) and other other data to train Zoom or third-party AI.

How much will it cost to use these convenient features and responsible AI?

With prices ranging from 150 yuan to 200 yuan, you can use it for a whole year, and enterprise users can also choose a customized version

The above is the detailed content of The energy consumption ratio reaches 10 times that of GPT-4! Huang Xuedong personally deciphers the success of Zoom AI, and the AI ​​companion can directly hold meetings for you. For more information, please follow other related articles on the PHP Chinese website!