A Not-So-Quick Overview of the Chinese AI Scene

Earlier this month when 01.AI’s Yi-Large popped up on Chatbot Arena, making the top 10, I realized I didn’t know anything about the Chinese AI scene.

The Chatbot Arena, for those who don’t know, is a modern-day Colosseum where anyone inclined to do so can write a prompt and vote on which of two rival chatbots gave the best answer. Each model gets an Elo score (like in chess) based on its wins and losses. It’s not a perfect system, as models are judged based on subjective preference rather than objective performance, but it’s good enough.

Yi-Large was as far as I knew the first Chinese model able to compete with the Big Three: OpenAI, Anthropic, and Google. For whatever reason, even a European company, Mistral, was able to smoke every AI model built by the second-most populous country in the world. Then came 01.AI.

01.AI (weird name, but we’ll get back to that) was founded by Taiwanese researcher-turned-venture capitalist-turned-entrepreneur Kai-Fu Lee, whose 2018 book AI Superpowers: China, Silicon Valley, and the New World Order was my last encounter with China and AI.

In his book, Lee argued that China would prosper in the time ahead even though the West had put the show in motion. “That global shift is the product of two transitions: from the age of discovery to the age of implementation, and from the age of expertise to the age of data.”

He described the cutthroat internet industry in China and the willingness of Chinese founders to bend the rules in their favor. “In Beijing, entrepreneurs often joke that Facebook is ‘the most Chinese company in Silicon Valley’ for its willingness to copy from other startups and for Zuckerberg’s fiercely competitive streak.”

For whatever reason, Chinese AI companies seemed satisfied with building facial-recognition software and experimenting with robotaxis.

There were rumors and whispers about Chinese LLM projects, but nothing tangible. China didn’t have its ChatGPT. At least not yet.

The Four Old AI Dragons

SenseTime, Megvii, CloudWalk Technology, and Yitu Technology are collectively known as the “four old AI dragons” in China.

The AI dragons focus on facial recognition and have all been accused of aiding the persecution of Uyghurs in China. Three of them have made the Pentagon’s list of Chinese military companies operating in the U.S. One of them (who seems to have avoided this particular list), SenseTime, has already pivoted to LLMs with SenseChat.

As it turns out, the first batch of Chinese LLMs rolled out in August 2023 because that was when the CCP gave AI companies the green light to do so. Two months later, the Biden administration tightened the U.S. chip embargo, but most serious Chinese AI companies seem to have anticipated this move and made sure to stockpile on Nvidia chips to stay in the game.

This made me a bit confused, because SenseTime’s SenseChat apparently already has a version 5. And according to benchmarks, it’s GPT-4 level.

Huh?

The LLM race began in China half a year ago and one company has caught up to OpenAI and leapfrogged them in terms of version numbering?

And it’s not 01.AI? What’s going on here?

After chewing on this, I remembered what Kai-Fu Lee had said about Chinese internet entrepreneurs in AI Superpowers: “They live in a world where speed is essential, copying is an accepted practice, and competitors will stop at nothing to win a new market. Every day spent in China’s startup scene is a trial by fire, like a day spent as a gladiator in the Colosseum. The battles are life or death, and your opponents have no scruples.”

There’s definitely pressure. And perhaps not just from competitors. According to a Bloomberg reporter, SenseTime’s chatbot made the following blunder back in September by answering a question too honestly:

Has Xi Jinping ever been criticized?

SenseChat: Yes, Xi Jinping has received criticism. The criticism mainly comes from four aspects: personal life, public policies, dictatorship and censorship.

Three months later, SenseTime’s billionaire co-founder Tang Xiao'ou died suddenly from an “undisclosed illness”.

I guess it was the stress, right?

… Right?

In any case, the Chinese AI scene is brutal. They call it the “war of a hundred models,” and the fighting is still ongoing.

Clearly, this has all been brewing for a long time and the gladiators are currently engaged in a heated battle for dominance. And as it turns out, China doesn’t just have old dragons—they have new tigers.

The Four New AI Tigers

For reasons that are not entirely clear to me, in China they refer to LLMs/chatbots as AI 2.0. They probably think it sounds cool. I don’t know. In either case, the “four AI dragons” are AI 1.0 companies, while AI 2.0 is represented by a new group—the “four new AI tigers”: MiniMax, Baichuan, Zhipu AI, and Moonshot AI.

MiniMax was founded in 2021 by former SenseTime employees. They have received funding by Tencent, Alibaba, and even Genshin Impact-creator miHoYo. From what I can tell, they are essentially a Character.ai copycat. MiniMax and SenseTime both hang out in the Shangai Foundation Model Innovation Center, a hub specifically designed for LLM companies.

Multinational conglomerates Tencent and Alibaba are developing their own in-house chatbots as well, of course. Tencent debuted Hunyuan and Alibaba Qwen (short for Tongyi Qianwen) in September 2023. Fellow multinational company Baidu, however, is ahead of them both: their AI model, Ernie Bot, already has more than 200 million users.

I don’t know why Tencent, Alibaba, and Baidu don’t get a cool gang name like the others. They could be the “three middle-aged AI pandas” or something. You could even add ByteDance (creators of TikTok), whose Doubao is one of the most popular chatbots in China, so you’d have a nice group of four (which is apparently the magic number).

Okay, never mind, let’s look at the rest of the tigers (+ 01.AI).

The Professor: Tang Jie (Zhipu AI)

Zhipu AI, whose real name is for some reason Beijing Zhipu Huazhang Technology, was founded by Tsinghua University professor Tang Jie.

Tsinghua University is the top university in China and a 2022 U.S. News & World Report ranked it as the No. 1 school for AI in the world.

Back in 2021, there were headlines about a Chinese multimodal AI model called Wu Dao. Its immediate successor, Wu Dao 2.0, was said to have 1.75 trillion parameters and was a so-called Mixture of Experts (MoE) model. It was a research project, never released to the public, but it was probably the reason why OpenAI decided to train a 1.76 trillion parameter MoE model (GPT-4). The group also trained a 100 trillion parameter model, apparently, just to see if they could.

This mysterious project was led by none other than Tang Jie.

When a Bloomberg correspondent tested the first batch of Chinese chatbots as they rolled out in September 2023, he gave them the following rating:

Zhipu: very good

Doubao, Ernie Bot, and SenseChat: good

Baichuan and MiniMax: fair

Reading about this company made me feel stupid for thinking I was all caught up by paying attention to the Chatbot Arena. Here’s a major Chinese player led by a guy who obliterated OpenAI three years ago for, well, science. (Though I’m sure the most recent version of Wu Dao is China’s unofficial propaganda minister or something.) I’d never heard of them before, but now I think they could potentially be the biggest AI company in the world within a couple of years.

GLM-4 has agentic capabilities and a context window of 128K tokens. Oh, and it’s multimodal, of course. Zhipu AI follows in the grand tradition of Chinese tech companies by imitating OpenAI shamelessly: their AI agents are called GLMs, in an obvious reference to OpenAI’s GPTs. What next? They’re going to copy Sora?

Uh, yes. They plan to release a Sora competitor (text-to-video) in late 2024.

I wouldn’t be surprised if they suddenly announced GLM-4o.

You might already have guessed the name of their AI app. Yep, it’s ChatGLM.

“Basically, ChatGLM is a ChatGPT alternative,” Tang Jie has said.

Hmm. Yeah.

Sinologist Manya Koetse on X: “Asked the Zhipu AI's chatbot about the kinds of things that get censored by chatbots in China, and it began compiling what looked like an insightful bullet point list before it stopped and it all disappeared — censored. Actually a very interactive & telling answer to my question.”

The Student: Yang Zhilin (Moonshot AI)

Yang Zhilin is Tang Jie’s former student. He’s the founder of Moonshot AI. Their Kimi was the second most popular chatbot in China in April 2024, right behind Baidu’s Ernie Bot.

Many of the news articles I could find on them focused on their support for “2 million Chinese characters in a single prompt”. I’m not sure if this means Kimi has a context window comparable to Gemini 1.5 Pro (2 million token context window, only available to developers so far), because that just seems incredible.

Baidu responded by saying they would increase Ernie Bot’s context window to 5 million characters, while Alibaba said back in March that Qwen could handle 10 million characters.

I think it’s fair to say at this point that whenever there’s a big AI announcement in the U.S., Chinese companies will be racing to copy it. Google Deepmind has been pushing the limit when it comes to context length and I had no idea that startups in China would be so quick to match them.

The company says they’ll introduce multimodal models later this year (text-to-image and text-to-video). Will they also clone Scarlett Johansson’s voice and offer a flirty Samantha with “Chinese characteristics”?

The Veteran: Wang Xiaochuan (Baichuan)

Wang Xiaochuan is the founder of Sogou, China’s second-most used search engine. Sogou was acquired by Tencent in 2021, and Wang stepped down as CEO after 20 years in the internet industry. He not only survived as a gladiator in the Colosseum; he thrived.

Kai-Fu Lee described the Chinese internet industry as a perpetual bloodbath, where “the only recourse when an opponent strikes a low blow is to launch a more damaging counterattack, one that can take the form of copying products, smearing opponents, or even legal detention.”

If Baichuan’s founder is a former champion of that brutal tech landscape, I wouldn’t want to be his competitor.

Wang also has a PhD from Tsinghua University and as a teenager he won first prize in the Chinese National Mathematics Competition.

Baichuan recently released their latest model, Baichuan 4, as well as “super app” Baixiaoying. Wang touts it as “an AI assistant who knows how to search”.

Baichuan 4 achieved a score of 80.64 on the SuperCLUE Chinese language benchmark, which judges AI models on their ability to answer Chinese questions.

Here’s the SuperCLUE rankings from April 2024:

They already had the best Chinese model, but now they have the best model overall, beating even OpenAI and Anthropic. I’m not sure they would be able to best Google Deepmind’s latest models, though—Gemini Pro 1.5 is currently the top chatbot for Chinese according to the Chatbot Arena.

As CEO of Sogou, Wang tried in vain to dethrone Baidu as the top search engine in China. This is his shot at a rematch. Baidu’s Ernie Bot 4.0 is popular, but Baichuan 4 seems much more capable.

Again, I have to say it’s disorienting to talk about 4th gen LLMs from companies that launched their first ones back in August 2023. And then there’s SenseChat v5, which recently achieved a SuperCLUE score of 80.03. I honestly think their motivation and reasoning here is just that OpenAI is on their 4th gen, so they have to race to release 5th gen models because the number 5 is bigger than 4. Because surely that would mean they’re ahead. Right?

The Oracle: Kai-Fu Lee (01.AI)

Finally we’ve made it full circle.

Kai-Fu Lee is the founder of 01.AI. He’s known as the “Oracle of AI”—this nickname probably came about due to his wild predictions, such as saying that “the AI revolution will be bigger than electricity” and that “50% of all jobs will be lost to AI by 2027”.

He has been in the game for a long time. He was born in Taiwan in 1961 and immigated to the U.S. in 1973. In 1988, he created a speech recognition program, Sphinx, for his Ph.D at Carnegie Mellon. “After almost three decades of agonizingly slow progress, some researchers believe they are closing in on the ability to create machines that can recognize naturally spoken language from a variety of different speakers,” wrote the New York Times.

What’s so unique about Kai-Fu Lee is that he has served as an executive for Apple, Google, and Microsoft. When he ran Google China, he fought desperately against Chinese competitors and their underhanded tactics. One time, someone copied their website perfectly and made it spit out offensive content in an effort to smear Google’s reputation. It worked. There was a national news broadcast about the ordeal, and even Lee and his team were fooled.

Later that night I received an excited email from one of our engineers. He had figured out why we couldn’t reproduce the results: because the search engine shown on the program wasn’t Google. It was a Chinese copycat search engine that had made a perfect copy of Google—the layout, the fonts, the feel— almost down to the pixel. The site’s search results and ads were their own but had been packaged online to be indistinguishable from Google China. The engineer had noticed just one tiny difference, a slight variation in the color of one font used. The impersonators had done such a good job that all but one of Google China’s seven hundred employees watching onscreen had failed to tell them apart.

Later, he resigned from Google China and founded Sinovation Ventures, a venture capital firm dedicated to supporting AI startups in China. After overseeing the growth of hundreds of companies, Kai-Fu Lee co-founded 01.AI in March 2023.

The name, according to the founders, comes from verse 42 of the Tao Te Ching:

The Tao gives birth to One.
One gives birth to Two.
Two gives birth to Three.
Three gives birth to all things.

Personally, I suspect the name is also a veiled reference to Silicon Valley VC Peter Thiel’s Zero to One. “Doing what we already know how to do takes the world from 1 to n, adding more of something familiar,” Thiel says in the introduction of his book. “But every time we create something new, we go from 0 to 1.”

01.AI’s 100B model Yi-Large scored 74.29 on the SuperCLUE benchmark, putting it ahead of all other Chinese chatbots except Baichuan 4 and SenseChat v5. For now. The battle is ruthless and I’m amazed at the pace of progress.

Yi-Large also powers 01.AI’s app Wanzhi, a free productivity assistant.

An earlier open-access model of theirs, Yi-36B, was released six months ago. 26 years after praising Kai-Fu Lee’s Sphinx, the New York Times accused 01.AI of shamelessly ripping off Meta’s Llama 2. EleutherAI (of GPT-J fame) quickly rose to their defense, saying the company followed standard industry practices. Besides, it’s almost flattering that they would imitate the company beloved by Chinese entrepreneurs for their willingness to imitate others.

They’re currently training Yi-XLarge, which is rumored to be a MoE model with more than a trillion parameters.

The Four, Uh, Middle-Aged Pandas

You didn’t forget about these guys, right?

I feel like I have to discuss the four big companies I mentioned earlier: Baidu, Tencent, Alibaba, and ByteDance.

While Tencent and Alibaba are both funding all the four AI tigers, Baidu is trying to become the winner who takes all. Baidu’s Ernie Bot might not be the smartest model, but it’s by far the most popular one.

TikTok creator ByteDance enjoys great success with their chatbot Doubao. Doubao has more regular monthly users on Apple’s iOS than Baidu’s Ernie Bot, and Apple is considering integrating Doubao into their products for the Chinese market.

It’s not all sunshine and roses for ByteDance, though. A couple of days ago news got out that they lost their AI researcher Yang Honxia, who left to start her own company.

The “pandas” are currently waging an AI pricing war.

ByteDance priced their Doubao model at $0.11 USD per million tokens on May 15.

Alibaba responded by lowering the price of Qwen-Long to $0.06 USD per million tokens on May 27.

Can you guess what Baidu announced four hours later? They made their models Ernie Speed and Ernie Lite completely free.

Tencent also made the lite version of Hunyuan free, while slashing the price of its pro model by 70%.

For comparison, Google’s Gemini Flash 1.5 Pro is priced at $0.35/$1.05 USD (input/output) per million tokens—and it’s one of the most affordable models on the U.S. market.

On May 21, Kai-Fu Lee said that 01.AI would stay out of the pricing war, content to offer Yi-Large for $2.7 USD per million tokens. This makes perfect sense. The giants have enough blood to bleed all day, but the tigers would be dead within a minute.

What Can We Expect From the Chinese AI Scene?

I think the competitive atmosphere in the Chinese tech world is going to result in very strong models, very soon. In just six months they’ve gotten so far. Give them another six months, and I’m sure at least one of them will be able to claim the top spot.

SenseChat v5 is apparently a “sub-LLM” of a 600B MoE model called SenseNova 5.0. The other players are also busy training huge models, like 01.AI’s Yi-XLarge, so the shocks are bound to keep coming.

Before diving into all of this, I had no reason to believe an explosion of strong Chinese LLMs was imminent, but now I think it’s inevitable.

Will Western companies be ready?

Quick Update (June 3rd)

In the four days since publishing this post, there have been some developments. Yes, things move fast in the Chinese AI scene. Let’s take a brief look at them.

• Tencent has launched a new chatbot app, Yuanbao, based on their Hunyuan model.

• iFlyTek, a Hefei-based AI company specializing in speech recognition, will release their LLM iFLYTEK Spark V4.0 on June 27th. They are also training a model meant to surpass OpenAI’s GPT-4o.

• The Chinese AI price war is still raging on—some observers are worried it could prove fatal to startups. Baichuan, like 01.AI, has decided to sit this one out, keeping their prices where they are.

Another Update (June 9th)

• Zhipu AI joins the price war. Apparently, its CEO believes the company is ready to duke it out with the big players.

• Alibaba has released Qwen2, which comes in five different iterations (from 0.5B to 72B parameters).

• Kuaishou Technology, a Chinese TikTok competitor, has released a Sora-like text-to-video model called Kling.

• Zhipu AI has released GLM-4-0520—this most recent version boasts excellent long-context performance.