Baidu should apply for intangible cultural heritage status for ERNIE Bot.

"Don't forget where you came from"

Comrades and friends, the meta has reverted!

AI application builders are out of moves. The scales of victory have once again tipped toward the foundation model companies. In light of this, we're paying homage to Funeral AI's series from a year ago — running through the model companies one by one.

We already covered Moonshot AI, Zhipu AI, Doubao, and Qwen, and realized we'd shortchanged the second-tier models. So we're emergency-adding Baidu, Meituan, Xiaomi, Tencent, and other companies in transition. More model companies coming soon 😘

Funeral AI has never written about Baidu, not out of laziness, but mainly because we keep forgetting ERNIE Bot is still alive.

Nowadays nobody even tells ERNIE Bot jokes anymore.

The earliest one I heard was: if you type "continue" in the chat box, ERNIE Bot would explain what the Chinese word "继续" means.

Now even the "AI is stupid" joke ecosystem has been captured by others. They've left Baidu absolutely no room to breathe.

Over the past three years, Baidu has left me with only two memorable pronouncements from boss Robin: one, that open-source models are an IQ tax; two, that closed-source models will only grow stronger.

Calling open-source models an IQ tax would be outright heresy today. I wonder what Robin thinks of his own words from two years ago now.

Of course, at least the second half wasn't wrong — it's just that everyone else's open-source models are stronger than your closed-source one.

Testing Baidu's ERNIE Bot these days, the first rule is: don't laugh.

This is also the first time I've seen an AI product where the only model options are ERNIE and DeepSeek. Did traffic surge after integrating DeepSeek in 2025, so they can't bear to take it down yet?

Let's start with a "stupid questions" classic:

How am I maliciously guiding the question?

Fine, let's go harder. You keep bragging about text comprehension, so let's test it with the most elegant Inside the Nail and see your level.

But first, this query suggestion — are you just slapping Baidu Search on here?

Now the result:

Hahahahaha, absolutely incredible ERNIE Bot. I asked about Zhang Xuefeng earlier and it just gave me Zhang Xuefeng's PowerPoint. No need to check the PDF either, I guess.

After I pointed out the error, it gave me an analysis of Inside the System of Government:

Are there any real humans left in this app?

Oh right — ERNIE Bot was also the first domestic model to charge for access. It's several tiers above Doubao on the monetization front.

November 2023 pricing: 59.9 yuan per month. Though after DeepSeek R1, it swiftly surrendered and announced free access — very much a wise man who knows the times.

Of course ERNIE Bot isn't entirely useless. It can also generate videos from your answers and one-click publish to Baijiahao — truly eating your own dog food, soup and all.

Unfortunately the video generation failed outright. Second try, third try — still failed. Failed attempts still count against your quota though. But no matter, already priced in.

Not sure what else to use it for. Let's test fortune-telling...

This Spring Festival, every consumer-facing app was running red envelope promotions. Quest Mobile compiled the DAU numbers. If you don't look closely, you'd never notice the red line at the very bottom.

Attentive viewers know that's ERNIE Bot's DAU. Those who don't would assume it's been dead for a while.

So the key to testing ERNIE Bot is lowering your expectations. That way you get this "holy shit, you actually know that?" surprise effect. IQ roughly on par with Xiao Ai.

Xianyu also added the next-gen foundation model ERNIE 5.1 to the Funeral AI benchmark. He originally included it out of morbid curiosity, but it actually managed to secure last place with zero controversy. Four out of ten tasks failed outright, scoring zero — a loss with no room for dispute.

I initially thought there was something wrong with the testing process, so I reran ERNIE 5.1 several times. No variance — half the rounds failed outright, a despairingly consistent performance.

At this point watching Baidu release models is pure entertainment. Just want to see what new clown moves they can pull.

In case you didn't know, when ERNIE 5.0 launched, it claimed 2.4 trillion parameters and called itself "the strongest humanities student."

Hard to imagine a company's model capabilities are still at this level in 2026. Calling it roadkill would be insulting to roadkill.

But on the flip side, in an increasingly homogeneous competition where model companies struggle to differentiate themselves, only ERNIE Bot lets us truly feel the gap between models. Isn't that Baidu's greatness? Sacrificing itself to elevate others.

When Baidu first launched ERNIE Bot, someone said:

"No matter your views, you must admit: on AI, Baidu has walked to the forefront of China. OpenAI allied with Microsoft attempting to be the gatekeeper of the large model industry — from today's perspective, most companies including Google and Facebook have basically been shut out.

Even if by hugging the wall, Baidu is the only company that squeezed through before the gate closed. Hope Baidu continues developing, to prevent an OpenAI monopoly."

At the time this wasn't that explosive. Would Baidu dare boast like this now? If we're relying on Baidu to fight OpenAI, then Chinese AI is truly finished.

But no worries — Baidu already has its exit strategy planned.

Because at Baidu's annual meeting this year, the messaging shifted: Baidu is no longer fixated on large models, "we're competing on Agents, competing on real-world deployment."

"There's a very practical motivation behind this: if Baidu continues defining itself as a large model company, it gets dragged into an increasingly homogeneous competition that's harder and harder to differentiate in."

I'm serious, this headline and content could make haters, fans, and trolls all simultaneously find peace.

What else can I say, brother? I can only praise Baidu's large model strategy: no fighting, no competing, detached from fame and fortune 👍

Translation: our models are dead, so we're playing against the meta and pivoting to applications.

Lost on large models, lost on cloud, now losing on AI applications — what's left to lose after that? Nothing. Face completely gone.jpg.

They should hire Fan Zhiyi as ERNIE Bot's spokesperson.

So what's Baidu's application game like?

First, the menu: Wenku GenFlow 2.0, Cangzhou OS, general-purpose agent DuMate, coding agent Miaoda 3.0, digital human agent Baidu Yijing, decision-making agent Famou 2.0, enterprise one-stop AI marketing app Hogee, video product Huixiang...

I've already demonstrated ERNIE Bot's level above. Did you even know Baidu has a separate Wenxin app, Wenxin Yige? If you did, you're a certified Baidu stan — go ask Robin for some stock.

As the first domestic model vendor to propose rebuilding all products with large models, Baidu has too many AI products to list. Too many.

Results aside, can anyone tell which is which just from the names? Also truly impossible to test them all, will review individually next time.

Maybe the only one that gained any traction is Miaoda? But this so-called first no-code tool looks worse than local agents like WorkBuddy or Qoder Work by a mile.

Now they're pushing something called DAA. Same old story: never won on models, never lost on coining new terms.

At this rate Baidu is positioning itself to live purely off government contracts.

Why Baidu fell behind on large models is a question worth perpetual study — it's the perfect case study of how to squander a winning hand.

Nobody believed in you, and you really didn't believe in yourself either.

Every company should study Baidu's organization and strategy — way more useful than studying ByteDance. It's a complete guide on what not to do. CEOs should all go consult Robin.

The feeling is, with Baidu's organizational capability, you could send them to do knowledge distillation and they'd still mess it up.

So buying Baidu stock now is purely an act of charity. Search moat eroding — Xiaohongshu ahead, Doubao behind. Classic fortress that failed to keep up with the times.

You can justify buying Tencent, Alibaba — Alibaba's models are top-tier, Tencent my bro Pony hasn't even made his move yet, Delta Force proves evergreen games have bright futures?

The reason someone might still be buying Baidu... kindness?

I think Baidu's last shot is whether those LibTV shares can buy a ticket on the AI ship. Though that Yueguang Zhang bet is probably a wash.

After discussing with Xianyu, we're now planning to include ERNIE models in every Funeral AI ranking going forward. Besides the minor benefit of dramatically showcasing model differentiation, it's mainly to remind everyone not to forget where Chinese AI came from 😭

Finally, nothing much to give Baidu, but here's a small suggestion.

Establish an ERNIE Bot History Museum.

No need for Baidu to fund it — let ByteDance and Alibaba pay. Main purpose: remind ByteDance, Alibaba, and other model factories to learn the lessons and not end up like Baidu.

Going forward, annual performance reviews at these model companies — anyone scoring M- or below gets mandatory training at the ERNIE Bot History Museum. Each person handwrites a 3,000-word repentance letter, analyzing how Baidu ended up here and how we can avoid becoming Baidu. Ensure the reflection penetrates the soul and reaches the bone.

This is also the relatively significant contribution Baidu can still make to China's large model industry.

(Cover image generated by ChatGPT, text 100% human-written)

⬇️

Subscribe to our Substack: funeralai.substack.com