Netizens voted for the king of AI, and LMArena became a $1.7 billion unicorn overnight.

A "Produce 101" in the AI world has become a hit! LMArena allows you to blindly vote for the strongest AI. It started as a campus project three years ago and has now completed a remarkable comeback. It just raised $150 million in funding, and its valuation soared to $1.7 billion. The crowdsourcing voting challenges the authority of experts, sparking numerous controversies, yet it has become an industry benchmark. Your vote can determine the next AI superstar!

A "Produce 101" in the AI world has become a hit!

A group of AI "trainees" such as ChatGPT, Claude, Gemini, and Grok are standing in line, nervously waiting for their performance.

This is not a talent show but a real AI battle taking place on lmarena.ai.

This once small campus open - source project recently raised $150 million in funding, with a valuation of $1.7 billion.

Top AI labs such as OpenAI, Google, xAI, and Microsoft are vying to send their models for "auditions."

Now, the strength of AI is no longer solely determined by big companies. The decision - making power is in the hands of global netizens.

How does this "AI training camp" work? Who will become the next top - star? Let's uncover the secrets together.

The "Talent Show Origin" of LMArena: From a Campus Project to the Silicon Valley Stage

It all started in 2023. At the Sky Computing Lab of the University of California, Berkeley, a group of graduate students and professors launched a small open - source project called Chatbot Arena.

The founders include Ion Stoica, a computer science professor at Berkeley (co - founder of Databricks), graduate student Anastasios Angelopoulos (current CEO), and Wei - Lin Chiang (current CTO).

Initially, they just wanted to conduct a simple experiment: let netizens anonymously compare different AI chatbots to see which one gave better answers.

Unexpectedly, this project became a hit as soon as it was launched and quickly turned into the most popular crowdsourcing benchmark platform in the AI circle.

In just three years, Chatbot Arena accumulated a huge number of users. In May 2025, it officially transformed into a for - profit company, renamed LMArena, and completed a $100 million seed - round financing, with a valuation of $600 million.

The turning point occurred on January 6, 2026 - just yesterday!

LMArena announced the completion of a new round of $150 million in financing, co - led by Felicis and the investment arm of the University of California, with participation from star institutions such as Andreessen Horowitz, The House Fund, LDVP, Kleiner Perkins, and Lightspeed Venture Partners.

The company's valuation soared directly to $1.7 billion, and the total financing exceeded $250 million!

Now, LMArena has more than 5 million monthly active users, covering 150 countries, and generates more than 60 million conversations per month.

These users are like the "national producers" who vote. Even the top AI labs quietly send their latest models for a showdown.

From an academic experiment to a new Silicon Valley star, LMArena completed a comeback that many talent - show champions would envy in just three years.

But the secret weapon that really made it popular is that simple yet addictive "blind - box PK" voting mechanism.

Blind - Box PK and Netizen Voting: The Power Game of "National Producers"

The climax of a talent show is the stage performance and on - site voting. LMArena's "performance stage" is equally exciting: it's called Arena mode, and the core is one word - blind!

Open lmarena.ai, enter the battle mode, and randomly input a question. The system will start randomly matching two anonymous AI models and give their answers simultaneously.

You don't know which model generated the answer. You can only vote based on your feeling. After voting, the website will reveal: Oh, it turns out that the one on the left is Gemini - 3 - Pro, and the one on the right is Grok - 4.1!

This form is like opening a blind box - fair and addictive.

The total number of votes is also included in the scoring system. LMArena uses the Elo scoring system to calculate in real - time. You get points for each win and lose points for each loss.

In the total score list seven days ago, Gemini - 3 - pro ranked first.

After summarizing the total scores, different category lists will be made public: text conversation, web development, visual understanding, text - to - image generation, image editing, search, and even text/image - to - video generation.

In popular categories, Gemini - 3 - Pro leads far ahead in the text and visual fields, Grok - 4.1 - thinking closely follows, and in image editing, GPT - Image - 1.5 and a variant of Gemini take turns topping the list.

Why do these top models participate in this "talent show"? CEO Anastasios Angelopoulos revealed the truth:

Leading AI companies use our platform because they themselves have difficulty judging whether their models are good or not.

New models that haven't been publicly released will be secretly hosted on LMArena for testing first, and they will use netizen feedback to quickly update and iterate.

Netizens are not just guinea pigs in the experiment. They even have a great time - without understanding technology, they can become "national producers" in just a few minutes and vote their favorite AI to the C - position.

Millions of votes form the hot - search rankings. Who goes up and who goes down depends entirely on the mood of netizens.

The Confrontation between "Scandal" Suspicions and "Paid Tutors"

No matter how popular a talent show is, it can't escape "scandal" suspicions and fan feuds. LMArena is no exception - it was involved in various controversies as soon as it emerged. Some people said "it's too democratic," while others scolded "it's too chaotic."

The most common complaint is that crowdsourcing voting is easy to manipulate.

In 2025, a paper directly exposed a scandal: Before the release of Llama 4, Meta secretly submitted 36 private variant models and repeatedly tested to "boost scores," successfully gaming the rankings.

Researchers from institutions such as Cohere, Stanford, and MIT pointed out that top labs can optimize through multiple private tests, and small and medium - sized players simply can't afford it.

Similar accusations include: some large companies are suspected of vote - rigging or prioritizing hosting new models, making the rankings seem "biased."

Some people also think that netizen voting is not professional enough. How can a random netizen's vote be compared with that of an expert?

This leads to its biggest competitor - Scale AI. Scale's evaluation method is completely different: they spend a lot of money to hire paid experts, such as lawyers, professors, and doctors, to score AI answers.

In September 2025, Scale directly launched the "Seal Showdown" platform, openly challenging LMArena and claiming that their method is more representative, more rigorous, and avoids the noise and bias of crowdsourcing.

Co - founder Ion Stoica said in an interview last year:

The highest - quality evaluation - the gold standard - is to let people vote on topics they are familiar with.

They believe that users know their own questions best and can give the most honest feedback; paid experts may have biases or be out of touch with reality.

Moreover, the diversity of users from 150 countries around the world makes the rankings more comprehensive and avoids the dominance of a single culture.

Despite the controversies, LMArena's rankings have become the de - facto industry standard - big companies still rush to participate.

But the talent show won't stop at voting. LMArena is already planning something big.

From Rankings to an "AI Talent Agency"

After a talent - show champion debuts, the most exciting thing is the "follow - up plan": holding concerts, shooting variety shows, accepting endorsements, or transitioning to acting?

The same goes for LMArena. It is not satisfied with just holding competitions and is already preparing to evolve into an "all - around talent agency" in the AI world.

The new round of $150 million in financing is mainly invested in this area.

The company's announcement clearly states that the funds will be used to significantly expand computing resources, recruit top engineers, and launch enterprise - level AI evaluation services.

In the future, LMArena will not only allow netizens to conduct blind tests but also provide paid professional evaluations for large companies like OpenAI, Google, and xAI. It will help them run models, collect feedback, generate reports, and even customize in - depth benchmark tests.

LMArena is also very ambitious in the field of reinforcement learning. Co - founder Ion Stoica previously revealed that the company is considering using a large amount of user - voting data to train AI models - this is the legendary RLHF (Reinforcement Learning from Human Feedback).

Using "good answers" as rewards and "bad answers" as punishments, AI can continuously optimize itself like a trainee practicing dance hard.

Peter Deng, a partner at Felicis, an investor, said bluntly in an interview:

Once it becomes the de - facto benchmark layer, the product will naturally expand. The real value lies in the in - depth cooperation with AI labs - combining their internal data with our comparative external data."

This "AI training camp" has just started, and the climax is yet to come.

LMArena proved a crazy fact in three years - in the AI era, the power of crowdsourcing can crush traditional experts, and democratic voting can become the sharpest yardstick.

More importantly, it has turned us from spectators into protagonists. Every vote of yours not only determines today's C - position on the ranking list but may also shape tomorrow's super AI silently.

ChatGPT, Grok, Gemini... Who can continuously top the list, and who will be suddenly overtaken by a dark horse? It all depends on the mood of us "national producers."

The future of AI is no longer far away. It lies in your next vote.

Reference:

https://www.theinformation.com/articles/ai-evaluation-startup-lmarena-valued-1-7-billion-new-funding-round?rc=epv9gi

This article is from the WeChat public account "New Intelligence Yuan", author: New Intelligence Yuan. Republished by 36Kr with permission.