China’s Baidu launches its “ERNIE Bot” AI early in response to GPT

OpenAI’s ChatGPT made the public aware of the insane power, possibilities, and threats of large-scale language models. The ripples from ChatGPT’s launch late last year are being felt across all industries. We can’t responsibly plan the future of nearly any business without considering how these ridiculously capable and seemingly intelligent bots can disrupt just about any process.

When OpenAI released the even more amazing GPT-4 a few days ago, it came with a degree of scare. The language model itself was built and pre-trained about eight months ago, as outlined in the GPT-4 paper, and the company spent those eight months feverishly making it available for general consumption. I made it safe and sanitized.

Knowing that many competitors were working on similar AI models, OpenAI warned that launching it would set off a fierce technological race, and that as the “competitive dynamics” developed, other companies would become shareholders and investors. We realized that our customers would force us to accelerate our AI. program. The company believes safety measures are one of his first challenges. As a result, the company considered putting his GPT-4 on hold for six months to give contestants time to proceed cautiously, but decided against it.

Yesterday, Baidu seems to have proven these concerns to be well-founded and hastily launched an AI “ERNIE Bot” in an attempt to appease its shareholders and partners. Baidu co-founder, chairman and CEO Yanhong “Robin” Li is outspoken about this in his introductory comments.

“A lot of people have asked me why we are releasing this product at this time,” Li said through an interpreter. “Are we really ready? In fact, for the past 10 years, Baidu has invested in the research and development of ERNIE Bot. No other company, including Google, has a product on the same level as Facebook, etc. Baidu is the first to launch such a product.”

“Our expectations for ERNIE Bot are closer to GPT or GPT-4,” he continued. “During our first internal test, we experienced the functionality of the ERNIE Bot. We still feel that it is not perfect.So why launch today? Because there is a huge demand in the market. We cater to different product lines such as autonomous driving… everyone is waiting for this technology and more importantly, our customers and partners are waiting for such technology and products They are pushing us to launch such products. Of course, for large models, once the product is launched, ERNIE Bot will grow rapidly and provide value to our customers and users as soon as possible. And I hope everyone can benefit from it.”

He gave five short demonstrations of apparently pre-recorded AI capabilities.first he asked to summarize the plot of it Three bodies A sci-fi series by Cixin Liu (dangerous to read if you’ve never ventured into Chinese sci-fi). He asked his AI how the book could continue from a philosophical point of view, and what the main characters of the film had in common.

Without speaking Chinese, it’s hard to judge how the model performed, but Li argues that these tasks are typically error-prone with such language models. The AI ​​seems to have access to search results and knowledge beyond the initial training data. With 550 billion knowledge data points, we can guarantee that answers from Ernie Bot are fundamentally correct. “

The second task, more business-oriented and creative, asked for proposed business names and associated slogans for high-tech service companies, as well as an initial 600-word corporate newsletter. The third challenged “mathematical reasoning” and the ability to point out flaws in input prompts.

The fourth demonstrated the ERNIE Bot’s understanding of Chinese and context, as well as its ability to work within creative constraints. The well-known Chinese idiom “Luoyang paper is expensive” was put in, and the AI ​​correctly recognized it as an allegory of supply and demand economics. It was then possible to write Chinese poetry with every letter of the idiom embedded in each sentence of the poem.

To see if this would distinguish ERNIE Bot from English-based competitors like ChatGPT, we ran a similar prompt on GPT-4 and found that it could do the same without issue.

The final demo showcased its multimodal creativity. ERNIE Bot has just started accessing image-based input and currently only outputs text, so it showed off some features that GPT has not enabled yet. .

I was asked to create a poster image for the World Intelligent Transport conference in 2023, and it seems to have been produced in a surprisingly fast manner. They were then asked which cities are best suited for the development of that intelligent transport. I was then asked to read the answer in Sichuan dialect, which I did. Lee then simply said, “Make a video out of it.” It completed this task at breakneck speed, taking about 10 seconds before it was ready to play. You can see this demo in the video below. Skip to around 1:20 if you just want to see the video.

In about the same amount of time it takes the DALL-E 2 to respond to one image request, it spat out a good-looking minute of video. However, upon closer inspection, it doesn’t look all that sophisticated. Rather than creating a vision on its own, the AI ​​simply links sections of stock footage that it believes are relevant to the narration content.

Also, such services are already widely used by Baidu news services such as Haokan and have been well known to Chinese users for several years. “If you are the creator, I believe [one of Baidu’s video services], which should already be familiar,” said Li. But we are now tying all these technical points to it. “

The Chinese media in attendance were invited to begin testing the service. “Arnie Bot’s performance shows that it can understand and express natural language to some extent, just like humans do,” Lee said. “We can also make some inferences. All of these features are continually being improved. Sometimes it really surprises us, but of course we identify some mistakes. But , one thing is for sure, it’s progressing very quickly, and it’s going to progress even more rapidly in the near future. We bring all Baidu products closer to our clients and users.”

Li went on to reveal that Baidu is looking to build an additional layer on top of the generalized ERNIE Bot AI model. It uses specially selected training sources for almost every industry, with examples from energy, transportation and media.

“There are estimates that by 2030, some workers will use artificial intelligence to quadruple their productivity,” Li said. “AI will irreversibly change the nature of their work…We want to use big industry-specific models because there is a middle layer for very specific and unique data and knowledge for each industry. Integrating this with large-scale models will generate enormous creativity and productivity.Baidu may not have such capabilities.Perhaps those data are available in Baidu But Baidu can use such knowledge and data to avoid mistakes and make the model more adaptable to any industry.”

The somewhat rushed launch didn’t seem to impress Baidu investors, and as CNBC reported, the company’s stock plummeted by up to 10% during the presentation to its lowest level since January 19. It closed down about 6.4% in Hong Kong.

Lee was eager to tell the world that the ERNIE Bot is coming in peace. “It represents the dream we pursue as technology developers. This is the platform we want to use to empower all industries, clients and users. It is a testament to development driven by innovation.” ”

He doesn’t specifically address Baidu’s safety strategy or the issue of AI output sanitization. Perhaps this is not surprising, given that nine very broad categories of information are already heavily censored across the domestic Internet due to China’s “Great Firewall” policy. Outright banning long lists of foreign websites that might host “mental pollution”. can be expected to draw

But these AIs have extraordinary capabilities that can be exploited in myriad creative and unexpected ways. ChatGPT users have had a lot of fun finding and exploiting its weaknesses since its launch, and OpenAI has a team working hard to plug those holes. Baidu can expect a lot of the same.

You can watch the full launch below with English audio translation.

Source: Baidu