GPT-4 has set off a new round of AI storm. Can Wen Xinyiyan, who is besieged, fight?

Baidu set the release date of Wen Xinyiyan on March 16th, but it did not expect that it would be bombarded by OpenAI, Google, and Microsoft: First, in the early morning of March 15th, OpenAI released large-scale multi- Modal Transformer model GPT-4; Immediately afterwards, it was announced that the API interface of the large-scale language model PaLM would be opened, and MakerSuite, a tool for developers, was launched; after Wen Xinyiyan was released, the giants did not rest, 3 On the evening of March 16, Microsoft released Microsoft 365 Copilot, an AI-driven office tool that claims to increase the productivity of Word, PPT, Excel, OutLook, and collaborative office software.

Wen Xinyiyan benchmarked ChatGPT and even GPT-4. The threshold is very high. No major global manufacturer has yet made it. Baidu is the first. I tested it myself. It still feels like there are a lot of imperfections.

"Practice brings true knowledge, and Robin Li also expressed his hopes: "Once there is real human feedback, Wen Xinyiyan's progress will be very fast, and I hope it will grow quickly. ” ​

​

​​

##Why will GPT-4 set off a new round of AI storm?

In the new year of 2023, the Yuan universe is still warm, and ChatGPT has become one of the hot spots in the AI ​​field. After a short period of fermentation, OpenAI finally made a big move and released the multi-modal large model GPT-4, which quickly Detonating global public opinion. In fact, the previous generation product of GPT-4 is the large ChatGPT model-GPT3.5. GPT-4 not only upgrades GPT3.5 in all aspects, improves natural language text processing capabilities, but also opens up The ability to process images,

triggered global imagination about generative AI and future working methods.

And GPT-4 generated answers that basically meet the requirements.

This is also This shows that GPT-4 can not only recognize complex instructions, but also perform complex operations based on complex instructions. It is reported that

GPT-4 can read, analyze or generate text up to 25,000 words, far exceeding ChatGPT, which has a maximum length of 3,000 words, can be used for application scenarios such as long-form content creation, extended conversations, and document search and analysis. It can not only identify various things in photos, but also understand the relationships between things The relationship

. In addition, the possibility of GPT-4 responding to prohibited content requests is reduced by 82%, and the possibility of generating factual responses is increased by 40%. It can be seen that GPT-4 that supports multi-modality no matter in nature There have been certain improvements in language processing and security.

​​

​

​

Google and Microsoft compete to expand generative AI usage scenarios

, indicating that the new version of the search engine Bing is already running on GPT-4

and is undergoing commercial monetization. Currently, Microsoft hopes to use The outside world is paying attention to the new Bing and OpenAI to promote commercial monetization. It is reported that Microsoft is holding roadshows around the world and holding private meetings with advertisers and advertising companies to demonstrate the advertising format of the ChatGPT version of Bing search it is testing.

Microsoft can be said to be at the forefront in exploring the generative AI business model. Its goal is the huge content creation market

. At least Google and Microsoft are I think so. After OpenAI released GPT-4, Google also announced a series of updates to its own products in terms of AI.

Gmail, Google Doc, Sheets, Slides and other Workspace office components will be launched with generative AI function. At the same time, Google has also opened the API of its own large language model PaLM. Copilot can also automatically generate content and improve office efficiency.

At the press conference, Microsoft stated that Copilot There are complex processing and orchestration engines behind it, and there will be support from models such as Microsoft Graph and GPT-4.

​​

​

​

The most significant significance of Copoilt is that it greatly improves production efficiency and reduces daily irregularities. Necessary work time is wasted.

Is Wen Xinyiyan being too harsh?

This shows how much Baidu attaches importance to Wen Xinyiyan. At the press conference, Robin Li demonstrated Wen Xin Yiyan’s comprehensive capabilities in five usage scenarios: literary creation, business copywriting creation, mathematical calculation, Chinese understanding, and multi-modal generation. However, when evaluating Wen Xin Yiyan, Robin Li said: "'Wen Xin Yiyan' in internal testing is not perfect, but it must be launched if there is market demand."

文心一The training data of the Yanda model includes trillions of web page data, hundreds of millions of search data and image data, tens of billions of daily average language call data, and 550 billion facts of ​​knowledge Chart​​​. In other words, purely from a retrieval perspective, as long as the database coverage is wide enough, Wenxin’s answer is definitely not a problem. For example, ask Wen Xinyiyan this math question - in a subtraction calculation, the sum of the minuend plus the subtraction plus the difference is 674. We also know that the subtrahend is 17 more than 3 times the difference. Find the subtrahend. Wen Xin answered the question accurately and listed the corresponding formulas.

​​

​​

The copywriting created is smooth, and the sentence structure is mostly expressed in parallelism, However, there is a certain lack of emotional expression. It can assist creation and refer to paradigms in article writing.

​​

​​

Deep semantic understanding also creates difficulties for Wen Xinyiyan, for example, when answering the question "Why didn't my parents invite me when they got married?" Wen Xinyiyan did not fully understand the logic, so he answered in the most direct way - Sorry to hear this news , maybe your parents have their own reasons for not inviting you... In fact, ChatGPT also fails to understand the relationship.

####################################################### ### ##From this point of view, the release of Wen Xinyiyan is not only the expectation of domestic generative AI, but also the requirement of partners. ###Baidu also hopes that after the trial, once there is real human feedback, Wen Xinyiyan will improve very quickly, and hopes it will grow quickly. Currently, Wen Xinyiyan is still in the closed beta stage. ######### Baidu is one of the few artificial intelligence companies in the world that has a full-stack layout in these four layers, from the high-end chip Kunlun core, to the Feipiao deep learning framework, to the Wenxin pre-trained large model , to applications such as search, smart cloud, autonomous driving, Xiaodu, etc., there are industry-leading self-developed technologies at all levels. Among them, Wen Xinyiyan is located at the model layer. Currently, ERNIE 3.0 receives billions of user search requests every day. ######### Faced with the strong iteration of OpenAI, which has raised people’s expectations for generative AI, Baidu still decided to fight to the death, and the official announcement made it clear. ### Just as Robin Li said: "This type of large language model is far from being fully developed, and there is a lot of room for improvement. It will definitely develop rapidly and change with each passing day in the future." ###### Facing the fiercely competitive generative AI market, the author is still full of expectations for similar domestic products. ###

The above is the detailed content of GPT-4 has set off a new round of AI storm. Can Wen Xinyiyan, who is besieged, fight?. For more information, please follow other related articles on the PHP Chinese website!