news

Recently, the domestic 3A game "Black Myth: Wukong" has become popular on the Internet and has attracted global attention shortly after its launch. According to statistics from the domestic game best-selling list, in just one day, the total sales of the game on multiple platforms exceeded 4.5 million copies, and the total sales exceeded 1.5 billion yuan. At the same time, some "victims" implicated by the popularity of the game have also appeared.

On the day Black Myth: Wukong was released on Steam, a well-known game streamer suffered from 3D sickness while playing the game live. This made him the most searched topic on Weibo, and many netizens jokingly called him the "world's first victim" of Black Myth: Wukong. On August 21, another "victim" of the game appeared, and his experience was similar to that ofMicrosoftrelated.

Entering "Black Myth Wukong Customer Service" in Microsoft Bing search incorrectly displays the personal mobile phone number of a Jifeng.com employee, not the official customer service phone number. In addition, there are two wrong phone numbers marked as customer service, including the contact number and email address of the copyright department of China Business Network.

The person whose phone number was leaked said that he received about 20 calls in 5 hours. It is reported that the main reason for this incident was that Microsoft Bing AI The assistant incorrectly captured information, which led to the leakage of personal information. Although the captured articles have been deleted and the victim has submitted complaints and other feedback, the incorrect "Black Myth Wukong Customer Service" information still appeared on the Bing search homepage. At present, judging from the search results, the Bing team has corrected the incorrect information.

As the world's second largest search engine, Microsoft Bing covers 36 countries and regions, with more than 600 million users. On February 7, 2023, Microsoft announced that it would ChatGPT Integrated into the new version of Bing (New Bing), the integrated new version of Bing adoptsOpenAI AI ModelGPT Upgraded version of 3.5GPT-4This incident may reveal that AI search engines have certain deficiencies in information capture and processing.

Last month, Microsoft announced a major update to Bing, which will see the search engine undergo a comprehensive transformation and begin to prioritize AI-generated answers. That is, when a user enters a search query, a master answer generated by AI will pop up on the results page, detailing all the selected information sources used to obtain the results. Of course, you will still see traditional search results on the Bing search page, but they will be displayed next to the AI-generated materials (in a smaller pane on the right).

Microsoft explained this change in an official blog post: "This new experience will be based on the inherent foundation of Bing search results, the same as the large/small language model (LLM It combines the power of AI and SLM. It understands search queries, indexes millions of information sources, dynamically matches content, and displays search results in new AI-generated layouts to more effectively meet the user’s query intent.”

Microsoft also listed some examples in its blog post about Bing generative search. In addition to the summary function, Microsoft will also provide the main source links of the large language model and the small language model, and the answers that users see are created by them. After the AI ​​generates the results, it is a regular list of result entries.

For example, when querying "How long can an elephant live?", the summary of the answer is followed by a video listing factors that affect the lifespan of an elephant. If a user searches for "What is a spaghetti western?", Bing generative search will display information about the history, origins, and classic works of this subgenre of film, as well as links and sources to this information.

At the time, Microsoft said that this adjustment was only available to a small number of Bing users, but it would be gradually expanded in the near future. Microsoft also stated in its blog post that they will continue to evaluate the impact of AI search on websites and readers. Some industry insiders are worried that ifAIIf robots scrape content and present it directly in chat windows or search pages, websites that create content for free will eventually go out of business.

In response, Microsoft said that this new AI search experience was built from scratch with this issue in mind, thus maintaining the same number of website clicks as traditional search. Time will tell if this is true. In addition, it is understood that Bing can choose to turn off the AI ​​generation function in the results page and only display traditional search summaries.

Microsoft isn’t the only browser company adding AI-generated results to its search pages. The race to bring more AI capabilities to search is heating up as Microsoft rolls out more tools for Bing.

However, countless real cases are proving that AI search is not as reliable and accurate as we imagined - it may make mistakes, and in some cases the generated results may even display wrong information and suggestions.

Earlier this year,GoogleA similar tool called AI Overview was launched to retain users who wanted to seek answers to their questions directly from the AI ​​chatbot. However, the tool also made some jokes after its launch, such as suggesting adding glue to make cheese stick to pizza and answering "Geologists recommend eating at least one small rock every day".

The Arc Search browser, in AI mode, confidently gave inappropriate medical advice that "severed toes will eventually grow back."

The artificial intelligence search engine Genspark recommends some weapons that may be used to kill people.Perplexity It plagiarized news articles written by some media without indicating the source or copyright.

In addition, AI-generated summaries may also cannibalize traffic from the source websites. One study found that AI summarization could reduce traffic to content publishers by about 25% by de-emphasizing article links.

These emerging AI search engines are popular for their ability to quickly generate large amounts of text and convincingly mimic human writing, but behind the scenes, AI "illusions" have become a key obstacle to these chatbots going to the next level. Unfortunately, some experts warn that this situation may never be solved.

A new report published by the Associated Press highlights that the problem of "nonsense" in large language models (LLMs) may not be as easy to solve as many technology founders and AI supporters claim. Emily Bender, a professor of linguistics at the University of Washington's Computational Linguistics Laboratory, expressed pessimism, saying, "The hallucination problem is simply unsolvable. It is the inevitable result of the mismatch between the technology and the proposed use case."

According to Shane Orlick, president of Jasper AI, in some cases, some nonsense is not a bad thing. Orlick explained, "Illusions can actually bring additional benefits. Customers have been thanking us for the inspiration we bring, and the root cause is that AI may output stories or angles that customers have never thought of by chance."

Similarly, AI hallucinations are also of great benefit to AI image generation, and models such as Dall-E and Midjourney generate eye-catching and wonderful images with this imagination. That said, only in the field of text generation is hallucination a real problem that bothers users, especially in scenarios such as news reporting where accuracy is highly emphasized.

“The fundamental principle of large language models is to ‘make up’ content, which is the basis of all their functions,” Bender said. “But because their ability comes from making up, when they output text that happens to correctly match our prompt word, it is actually a coincidence. Even if the models are fine-tuned to be correct most of the time, they will still not be completely free of glitches. And future hallucinations are likely to exist in a blurred state that is more difficult for readers to notice.”

Large language models are powerful tools that can achieve extraordinary functions, but companies and the entire technology industry must realize that just because something is powerful, it cannot be considered a good tool. Just like a hammer drill is also very useful and can easily break through sidewalks and asphalt roads, but no one dares to take it to an archaeological dig site.

As Bender points out, when a large language model is first trained, it is trying to predict the next word in a sequence based on the prompt word we give it. Each word in the training data is assigned a weight or percentage to track the previous given words in a given context. But these initial words themselves do not have enough concrete meaning or important context to ensure the output is accurate.

In other words, these large language models are just excellent imitators, they don't actually know what they are saying, so over-trusting them will only put users in trouble. This weakness is inherent in large language models, and although the "illusion" may gradually decrease in future iterations, the problem itself may never be truly fixed.

Reference Links:

https://www.engadget.com/microsoft-is-adding-ai-powered-summaries-to- 必应 -search-results-203053790.html?src=rss

https://www.techradar.com/computing/artificial-intelligence/ 必应 -has-been-revamped-to-prioritize-ai-search-results-whether-you-like-it-or-not

https://www.techradar.com/computing/artificial-intelligence/chatgpt-and-other-ai-chatbots-will-never-stop-making-stuff-up-experts-warn

https://techcrunch.com/2024/07/24/bing-previews-its-answer-to-googles-ai-overviews/