Weekly #001 | The LLM "Kill Line" Wars: Tokens Are the New Money

Hello everyone, welcome to Next Token | Beyond the Token. This is Next Token Weekly, our weekly recap of big AI happenings. I'm Pan YANG, and I'm the host on duty for this episode.¶

The show has three other hosts: Qiaomu, Orange, and Guizang. Hosts, please say hello to everyone.¶

Hello everyone, I'm Xiangyang Qiaomu.¶

Hello everyone, I'm Orange.¶

Hello everyone, I'm Guizang.¶

Alright, let's get right into it. The few of us have long wanted to record a series like this: at a fixed time every week, give everyone a review of the various events and news that happened in the AI world over the past week, plus some interpretation based on our own understanding. That way, if you follow the show, you can catch the big fresh happenings of the past week and stay aligned on the latest AI knowledge.¶

Today we'll go through several parts. This time we're covering the week of August 27 to September 3. First, the models released over the past week — there were a lot of them; then AI Agent-related products; in the third part, some hardware we find interesting; and finally some big business events and news.¶

Let me first read out the names of the models released in the past week, and then the hosts can chat about them together. Not in chronological order — I'll just pull up a list.¶

The past week was a momentous one. Right before we recorded — yesterday evening, I believe — Gemini 3.8 Flash was released, and right after that benchmark just hit the top, Meta's Muse Spark 1.3 came out.¶

The past week also had Fable 5.1 and Mythos 5.1, and a bit earlier Qwen 3.8 Max and Qwen 3.8 Flash Next. There was also Zhipu's GLM 5.3 Flash that was all the rage this past week — the model everyone knows as "Niulai" ("the bull is here," a meme nickname) — plus Hunyuan's Hunyuan 4 Preview.¶

Zhipu 5.3 was released the week before, but last week it was open-sourced. DeepSeek V4 Flash Vision was also open-sourced, as an early-access edition.¶

There were a few more niche ones: one is Google's Gemini Omni 1.1 Flash, an omni-modal model. There's also Gemini 3.5 Transcribe, a speech-to-text model, and Meta's Muse Voice Transcribe.¶

Also, last week had one big event — not a newly released model — which is MiniMax H3's infinite livestream. It took the whole internet by storm and may have pioneered a whole new way to play.¶

That's the model list I put together as this week's host. Now let's talk: what did everyone think of last week's models? Guizang, kick us off.¶

I think the one worth watching is GLM 5.3 Flash. Mainly because, the way I see it, if people can't afford something, they stop paying attention. Especially for users in China — take Fable 5.1 or Mythos 5.1, way too expensive. After launch they said themselves that inference cache costs dropped 75%, but a 200-dollar 5-hour quota gets burned through in 3 minutes, so a lot of people still can't afford it.¶

The ones with good capability and low cost are what everyone pays attention to. Including Hunyuan 4 Preview and some of the trial experiences on WorkBuddy — people probably paid quite a bit of attention to those too.¶

A 2-week trial.¶

MiniMax H3 is actually a model too. It's not the original H3 — it's a speed-up fine-tune done by fal.ai, similar to the old Lightning models. Inference used to take maybe 40 to 50 steps; now it's 2 to 3 steps, so it's fast, real-time.¶

And there's a bit of gossip: apparently another open-source project made a similar fine-tuned model, and the fal.ai guy went on Twitter saying their stuff was useless, and they got into a fight. Pretty funny. But the livestream thing really is…¶

About this infinite livestream thing — how come my impression is that something similar already existed last year, but it didn't catch fire?¶

It's because of quality problems. These Lightning models are basically extreme quantization — the quality is almost not the same model as the original anymore; quality is terrible and the overfitting is severe. MiniMax H3 is still fairly big, so even after using this trick, the generated video quality is still decent.¶

fal.ai also has a demo: it hooks up an Agent, with a video on the left and an Agent on the right. When you talk to the AI by voice, the visuals change in real time.¶

For example: "walk me through how humans came out of Africa and evolved." A TTS narrates on the right while a MiniMax model on the left generates animation in real time. And the moment you interrupt and send a new question, the animation on the left immediately starts changing. For education, I think the demo is quite…¶

There's a bit of product thinking built into it.¶

It makes you feel like it's different from what came before.¶

Last year several companies did demos like this — it felt like just showing off frontier tech, and none of it actually got used.¶

Real-time.¶

Last year a lot of them had no real theme, it was like watching a dream — that kind doesn't work. Look at the scene just now; it actually had the feel of a proper show video.¶

It's actually usable. Those earlier things were all very demo-ish, and the compute they burned was terrifying, so many were closed betas — they'd let us try it for a bit, then never open it up.¶

It's kind of genuinely used. And it has time limits — they'll tell you, for this scenario, at most how long the generation can run.¶

It just crashes.¶

From what I read, spending over 2,100 dollars a day gets you a 24-hour livestream — inference costs have come down to that level. So compared with last year's infinite livestreams, the video quality H3 generates today has actually reached a very high level.¶

The quality is much higher than before — at least it's watchable now. Before, honestly, the real-time stuff was unwatchable.¶

But I feel like with H3's release, it's — it's once again, another breakout moment, once again it, it — expanded the whole video generation field globally, I think the whole, the market grew, grew by another ring.¶

On model promotion — I think vision-related models have an advantage in promotion and marketing. Didn't we just mention that everyone's benchmarks now are 3D, Three.js — people have run Three.js into the ground. Why? Because…¶

But visuals are still pretty…¶

Consumer users can only perceive this kind of thing. You say your Agent got so much stronger, your actual execution time grew so much longer, that you can do, do healthcare stuff or whatever — no perception. People can still only go by these visual things.¶

So when model vendors build training data, they must have been piling more of this in all along.¶

So for a while everyone focused on multimodality, Google included back then. Then they all stopped focusing on multimodality, because it really doesn't help model quality, intelligence, or IQ progress. Now they're carrying multimodality along again, because without it, what you build just doesn't look good.¶

Good for marketing.¶

And H3 open-sourcing is still a big deal. This time — the model quality is actually decent, and now that it's open source, it makes the whole ecosystem flourish.¶

Won't this affect hardware too? I feel like a lot of individuals are starting to want to buy better GPUs and deploy it themselves.¶

Plus it can, you know, do that thing.¶

Then it's not just H3. I have a 128GB Mac Studio at home, and I've actually tried a lot of this wave of models. Early on it was Alibaba's Qwen series, 27B, 35B; then this latest DeepSeek V4 Flash — I installed one on my machine, inference hits about 30 TPS, already usable.¶

Then this "Niulai," the 5.3 Flash, is also decent — I tried it on my machine. And Qwen 3.8 Flash Next. These are all things a local machine with 256GB of memory can handle.¶

I think 100B is a very sweet-spot parameter count. As of now, 100B is very…¶

But I still think local inference remains a niche use case. For ordinary people to break even, the math honestly doesn't work. It definitely doesn't.¶

I did an extreme calculation: buy that Mac Studio, and if it's running every second of every day, you break even in about two years. But for normal everyday use, you're basically looking at 10 or 20 years, because there's no way you're using it nonstop yourself.¶

Mainly, isn't everyone saying OpenAI bought 30,000 Mac Minis? When the news broke, lots of people were ranting that they drove Mac Mini prices up.¶

Can't get one. But one thing I know is that OpenAI buying Mac Minis isn't actually for running models — it's to do large-scale Agent tracking in the Mac environment, building training data around Agent behavior.¶

The environment. They're all using it.¶

Computer Use and all that need it too. Speaking of Computer Use, I think today…¶

Over the past six months, models' Computer Use and Browser Use capabilities have improved enormously, and I think that has massively boosted efficiency. We used to have to stop a huge amount of work and hand it to a human; now a lot of that basically isn't needed at all.¶

Now the acceptance testing after Vibe Coding basically all works like this — for example, Codex automatically invokes Computer Use.¶

By the AI itself.¶

So the browser built into Codex is amazing, and now it can log in directly — so good. It doesn't need to read your username and password.¶

No need for cookies or anything either.¶

Let me pull up the history. I remember around late '23, early '24, OpenAI acquired a small company. That company originally built multi-user real-time interaction technology frameworks for VS Code or similar products, including screen reading and simulated mouse clicks. At the time of the acquisition, nobody really knew what it was for. But over these two years, Codex's lead over every other vendor in this area has had a lot to do with that acquired team.¶

They're all called Computer Use, and there are open-source ones too — why does it seem like none of them do it as well as Codex? Why is that?¶

Especially Claude — it seems to be really bad at it.¶

Really bad. Claude, as of now — I saw they posted yesterday — only just started supporting Computer Use. Released last night, I think.¶

But I've been using Zhipu's ZCode recently, and its support is quite good too. Zhipu ZCode is a complete Codex Alternative — I think that positioning is great. I'm used to Codex, and switching to ZCode is almost seam…¶

Super smooth. Zhipu has actually been leading here in China all along. The Lobster had this feature — one-click install of Feishu — done entirely with Computer Use; you don't have to touch anything. That was really ahead.¶

I think Zhipu released that AutoGLM last year, right? So there's been someone working on this line all along.¶

Using that — iterating on that base.¶

Working on it all along.¶

So when people recommend Chinese Harnesses, lots of them recommend ZCode — quite a lot, actually.¶

So sometimes when building products, don't be ridiculous — if something's good, just borrow it openly and use it.¶

Zhipu's problem before was that their GPUs for the models were never enough — always short — so the Flash means a lot to them too.¶

Niulai wave just straight up, like…¶

Straight up.¶

How many tens of thousands of GPUs they've deployed, no idea, but…¶

I think…¶

They put together a cluster. Didn't the rumors say it was at… you know. Mainly it's efficient — their inference efficiency is 10x the previous version.¶

But look, you can clearly feel that although "Niulai" has a smaller parameter count, it's not on the same route as Google Gemini. Gemini is pushing speed, 100+ TPS; Zhipu is going for affordability — the speed isn't pushed up, but wasn't supply the bottleneck before? Now they've opened up supply.¶

No, that's tied to the hardware — with domestic GPUs the speed probably can't…¶

Fair — they'd want to push it too, they really can't, really not as good as TPUs. TPUs really are, in this wave, no question…¶

But the volume is huge. Look how long it's been open on OpenRouter, supporting global usage at this scale — really impressive. 66T, 66T consumed in one week. At one point lots of people on X were speculating: is this a Google model? Is this compute backed by NVIDIA?¶

You're saying it's not China, because China doesn't have that many GPUs.¶

Right, no — not that many GPUs.¶

That's because this time they upgraded to a whole new architecture, plus multimodality — small yet strong. And the official GLM coming up, didn't they say it will achieve complete self-training? Very impressive. Qwen's Flash Next is also a new architecture — the Qwen 4 architecture.¶

But Zhipu — aren't 5.3 and 5.3 Flash still, like, the architecture hasn't changed for the last few versions?¶

No, Flash is a new architecture — the ones before had no multimodality.¶

Flash did change. Niulai is a new architecture.¶

Niulai is a new architecture, so it's a very impressive model.¶

I think Astra said — yesterday, about that Astra, the one that didn't release GPT 6 — it's that loop depth, they used a new architecture again, so there's no way to monitor the amount of thinking anymore.¶

Astra's inference cost is 4x Sol's, the price is 3x, and they still give you a discount.¶

At this rate no one will afford the new models.¶

Very expensive — and supposedly that's the price after the cut.¶

It's fine, they've got the cost covered.¶

Look — if cheap models are this good going forward, or put the other way, good models are this cheap, and the best and the usable have a 10x gap between them, then from the Agent's perspective, mixing models at different price points is something that really should be pushed hard.¶

NetEase Youdao Makes sense.¶

Right now it's still humans doing the switching.¶

But I think the core problem is, if you let the AI control model switching, the AI doesn't control it well.¶

Look, there are two routes right now. One is doing it from the Router angle — I think that's even more absurd; at least an Agent has some judgment.¶

Routing.¶

So how is it done now? Judge the task's difficulty, then assign it to other models at different cost levels.¶

And is anyone working on this? No idea.¶

Is anyone on either side working on it?¶

Here's the thing: a good model is good at everything; it's just that I think the ROI differs. Take replying to a message — using a very expensive model is fine, because it's a one-shot reply, a bit pricey doesn't matter. But if you're going to run it long — an Agent task — running it on Fable is a losing proposition.¶

Yeah, don't — don't bother.¶

There's also the cache problem right now. So it's still basically Agent Teams,¶

…done in the Sub Agents pattern, because you have to account for caching.¶

Speaking of Harness just now — that was also last night's news. Guanlan and their folks, that is Runta, put out a new benchmark called Harness Eval, benchmarking which Harness is the most efficient. It looks from all angles, including task completion rate and token cost.¶

In there you find that with the same model, Claude Code is extremely expensive.¶

With the model held constant.¶

Claude has a bug — it's not that they deliberately broke caching for non-Claude models; that bug still hasn't been fixed.¶

That's a caching thing — no, no.¶

That doesn't count as a bug.¶

In that benchmark, the most efficient is Codex — we can set aside time to talk about it separately later. It now also has a "kill line" — a curve; being on the edge is basically the kill line.¶

DeepSeek Harness is extremely efficient, and it even benchmarked three different modes of DeepSeek Harness separately. The three modes differ in token consumption and efficiency — that's how fine-grained it is.¶

Codex is high because its tools are more complete, and its own…¶

I think it might be that it's simple.¶

Actually it might be simplicity.¶

Its architecture really is efficient.¶

Its architecture, yes.¶

Claude Code's architecture is actually heavy. Look — they've already deleted half the prompts, because it came from that era. There were tons of prompts, including early TODO-style stuff, way too much edge-case baggage.¶

Piled on a bunch.¶

Codex's time — things were already lighter. And DeepSeek is newer — DeepSeek doesn't deal with any of this.¶

DeepSeek only ships the essential tools, nothing else.¶

Aren't there lots of developers who like that Pi Harness?¶

Pi is there too, but for some reason Pi…¶

Pi — I read the report this morning. Pi is inside the kill line, which is like being killed by it — it's not on the edge. The kill line means: either cheap or good. It's in the middle, probably overshadowed by DeepSeek Harness.¶

It's lightweight in itself, but much weaker on extensibility — MCP support and all that is poor.¶

And then you depend on the community, and the community just…¶

That's more suited to developers DIY-ing it, but now DeepSeek Harness feels more DIY-friendly. DeepSeek Harness, I think…¶

DIY. Right, didn't I write an article about this a while back? I think DeepSeek did the Harness well and people are happily DIY-ing it, but actually, it's not meant for people to…¶

Use.¶

DIY-ing it is pointless — what are you going to do after you DIY it?¶

Can't do it well.¶

It's for users to play with on their own — fine for playing. But actually DeepSeek Harness is for the upcoming joint training of models and Agents — it's really…¶

Self-evolution.¶

Before, say with Claude Code, at most your data was the Agent's behavior trajectories, but the Agent itself couldn't self-evolve.¶

One more recent observation of mine. Same requirement — say background research. This time, preparing materials for our show, I asked it to sweep the whole web for info on the four of us. Give the same requirement to Codex, and it might trail off halfway — I'd rate it maybe a 20 out of 100.¶

But with DeepSeek Harness, everything a human could find by hand — and everything they couldn't — it finds. And the most amazing part is, I don't know why, I haven't looked into it, it can read WeChat official account articles. First it can search them out, then it pulls down the complete content at that URL, and it's super lightweight — definitely not using a browser, it just pulls it straight down.¶

Throw the same link to Codex and ask it to read it, and it says it can't.¶

Can't read it.¶

Codex deliberately doesn't read — it can but won't. All of Codex: if you've declared you don't allow reading, it won't read, it just flat-out won't — like that Atlas browser before.¶

Pretty good about respecting the LLM.txt protocol.¶

That thing of theirs before.¶

But when I used Codex, what came back was: it says it can't read it — it tried and couldn't.¶

No — it just doesn't read.¶

Same thing.¶

Pretty much the same thing.¶

Atlas was like this before — it could see it, but it wouldn't read.¶

Because WeChat official accounts — normal people can view them, so in theory an AI viewing them shouldn't be a big problem either.¶

Like, for example if I use…¶

No, WeChat official accounts do block it a bit — write any random one, any random one, and it can't be read.¶

The UA, the UA.¶

Like, if I use Dia it reads fine — the Dia browser reads it just fine — so I feel like actually it…¶

It has to disguise itself as a human.¶

So now, say I want to write some articles and need background research, I just use DeepSeek — it's great.¶

Use DeepSeek.¶

Then what does it use for search? This kind of thing definitely has to call a search tool.¶

I haven't taken it apart — haven't looked closely.¶

It's built in — the DeepSeek model has search built in.¶

It has that search engine built in, right? But that search engine, it — right, it's at the model layer, the API just comes with it, the model comes with it. It's at the model layer. You can't see it in DeepSeek's open-source code,¶

Then are they still using info sources from that Hangzhou company, Bocha?¶

Probably not.¶

Nowadays.¶

Probably not.¶

Probably more than one provider, too.¶

I…¶

For overseas info it definitely isn't Bocha — Bocha has no English-language sources.¶

Feels promising. Using that Herdr for this kind of thing is really interesting, really fun.¶

Which one?¶

That Agent that wraps around the CLI — written as a CLI.¶

Is it called Herdr, or Herdr-something? Yeah, that one.¶

Herdr, Herdr. You start it first — launch it in your terminal, then start the CLI inside it. They can talk to each other, and you can pipe DeepSeek Harness into it.¶

Right, super geeky — lots of geeks will love it: start it first, then wire up all sorts of things.¶

Whereas I haven't used a terminal in ages.¶

And he's…¶

Me neither—¶

Yeah, honestly.¶

But you shouldn't — you came from a technical background, you shouldn't.¶

People still like interfaces, still like things with visual variety.¶

So people still like—¶

Mainly, that Codex 200-dollar plan is just great to use.¶

So great. I honestly don't know — I use it to death, and the only time I came close to using it up was 85% of the weekly quota, one single time.¶

Did you use 85% or have 85% left?¶

Codex.¶

Used 85%, and then it just snapped and reset on me. So I can't use it up unless I go really crazy with it.¶

Anyway, you—¶

You can't burn through it coding — normal coding won't use it up, unless you have it write articles.¶

I figured out how it actually gets used up.¶

Editing videos, writing articles — that'll burn through it for sure.¶

Editing, tons of screenshots, crazy multimodal usage.¶

Really? Well then, right — later we'll let Codex edit this very video.¶

Editing videos.¶

One more thing I find interesting: why has everyone's model progress been so rapid lately, with release frequencies so high? I asked this question on X and got lots of replies. The most-liked take was:¶

For one, in July the overseas models all shipped new versions, forcing Chinese labs to speed up or risk losing competitiveness.¶

Second, the claim is that the open-weights-plus-commercial route has been proven to work.¶

Third, roughly: since China still lacks GPUs, people have innovated a lot in post-pretraining and post-training.¶

A lot of work.¶

So with model releases getting faster and faster, plus the domestic inference cards now being usable too.¶

The B300s are already more than enough.¶

So I don't know — Pan YANG, you seem to understand this whole thing so deeply.¶

I think on the first point, I remember someone saying that model training has no real secrets in this industry — at least not among the Chinese model labs. Once one figures something out, soon everyone…¶

Actually almost everything can be bought and sold — your GPUs, your data, all kinds of things are readily available. Only people can't be bought and sold. Everything else…¶

People can move between companies too, actually. Didn't we see yesterday that Hunyuan poached someone from Zhipu, one of the 5.3 people? I think — and they don't even need to formally quit; everyone's really just one circle exchanging with each other. That works too, and everyone publishes papers. That's one aspect.¶

Then about the point Qiaomu just made — I don't exactly… it's not that I disagree, it's just a different angle. Everyone thinks Chinese models are chasing so hard now, putting huge pressure on the US. Actually I think it's that side — yes, I think so, that's why I say it's the reverse. That side stopped. Regulation. Safety. Understood. It's the American side that got pressed down by policy and various reasons, safety. In fact, we might still be two generations behind,¶

I think the anti-AI sentiment in the US or overseas is terrifying. It often goes like this: if you're a YouTuber who wasn't in the AI scene and you take an AI ad — whatever AI ad it is — there isn't a single kind word in the comments.¶

We know the public discourse overseas is quite loose — the stuff that would get you, you know, over here basically won't be too, you know. But over there, once one blogger calls you out, that's it — you get absolutely savaged.¶

So Niulai turns out like that and the foreigners are cheering, saying it's the best thing artificial intelligence has ever made.¶

Actually Niulai is already hugely popular abroad.¶

So our side is still pretty… especially for image and video model products, even more so.¶

But I think on our side it's just a matter of who holds the microphone — people can't speak up. Actually I think that same sentiment is rising in China too; there's just no voice, people aren't allowed to speak up.¶

Could be.¶

Deep down it's still about the actual work.¶

I think with Chinese open-source models rising this fast, in reverse… how to put it — it will also amplify the voices supporting Chinese open-source models. Globally, the voices in support of Chinese open-source models will rise. They will, for sure.¶

And this will boomerang back — more and more people will step up to attack that…¶

"Why are you so expensive and…"¶

But with models competing this fiercely now, we've reached the point where there isn't enough attention to go around. After K3, Qwen 3.8 Max came out — the capability is actually about the same, and nobody cared.¶

I couldn't agree more on this one. These past two weeks, people keep asking me which model is good or which Agent is good. I tell them I can't even keep up with trying them — can't try all the models, can't try all the Agents.¶

I tried several these past few days — Hunyuan 4 Preview, and Qwen 3.8 Max 0902 too — and honestly I can't even come up with a really good test case anymore.¶

You can't tell them apart anymore.¶

No difference.¶

They're all about the same.¶

The only thing you can still feel, I think, is speed — and a lot of the ones called Flash aren't even fast.¶

Like you said.¶

Let's dig deeper then. Once we've entered this stage, how will the way developers or consumers choose change going forward?¶

Before, people read reviews and community feedback, then tried things themselves. But if you can't even try them all anymore, how will people go about choosing models or Agent products?¶

That's why I think letting everyone try GLM 5.3 for free is a great move. Trying a model has a cost — even with KOL recommendations or sponsored endorsements, I still have to try it myself before I believe it.¶

Trying it yourself is the best way.¶

To try it myself I have to pay for a Token Plan. If I can't afford it, I probably just won't try. And if I'm not allowed to buy a Token Plan, how am I supposed to try it?¶

So letting people try it free for a while lets them build their own felt experience, and lets good cases emerge from the community. It's a fairly positive cycle, and consumer users also get a free model they can keep using.¶

What you just said is worth a million to the model vendors.¶

This makes total sense. I said it at the time too: letting everyone use it free on OpenRouter is actually the fairest, most sensible testing method — better than any benchmark. Some people might think it's grabbing everyone's attention, but no — it's simply the fairest.¶

And our product Cola has had a free model available every day since last week. Users give you direct feedback — for example, Qwen's Flash is clearly not as good as Zhipu's Flash. Google's Flash has its strengths and weaknesses; users know it all after one try — more accurate than any benchmark.¶

So see — we started this conversation with Qwen. When Qwen releases its next new model, it should let everyone experience it free across the whole web.¶

Actually that Qwen model is still free on our platform — we've joined in on this free-model testing game too.¶

So you're saying free alone isn't even enough anymore.¶

No, its daily consumption is still huge — you can clearly see it's our number one by daily usage.¶

But following Guizang's logic through: if you go free but don't create that wave of impact and re-sharing, then your free was free for nothing.¶

I think so — they just didn't put out enough volume.¶

And the impact wasn't big enough either.¶

Exactly. If the model is decent and free for a while, users will stay — there's retention. But if, say, Qwen's 3.8 Max is too expensive now, users won't stay.¶

So in the end it still comes down to price-performance.¶

Nowadays you either be number one and command a premium — even a 10x premium — or, like Niulai, hit the kill line.¶

Or you define a new kill line.¶

Only these 2 options; the middle doesn't exist. Because Codex has already maxed out the 200-dollar tier — the middle has a lot of volume, and they're making 20x on it.¶

Honestly, to put it bluntly: OpenAI's 200 dollars plus the constant resets — its all-in cost beats every model in China outright.¶

That's the real kill line.¶

It's enormous.¶

Yes, yes.¶

You have to compare with Luna — Luna has already cut prices 80%. You can't be more expensive than Luna; Luna is very good too.¶

Luna's 200 dollars — you can hammer it as hard as you want and never use it up. That feeling—¶

Yeah.¶

It's like it doesn't consume at all. Luna's price cut was brutal.¶

Yes.¶

The usage bar doesn't even move.¶

Absolutely brutal.¶

So see, after DeepSeek raised prices, a lot of people stopped using it. Pro has no price-performance left; Flash still has some users — pretty scary.¶

Is there now a strong rivalry between, say, Zhipu's 5.3 Flash and DeepSeek V4 Flash?¶

No strong rivalry.¶

I think—¶

Because that Flash just gets completely crushed.¶

DeepSeek is a bit weak — the vision is even worse, kind of rough.¶

Completely unusable. So that means the sweet spot on the kill line right now is the 5.3 Flash.¶

5.3 Flash, Niulai—¶

Exactly, the five-point… And there's one hidden little wallflower that's always inside the kill line. MiMo 2.5.¶

That one's basically provided for free.¶

It's always inside the kill line — truly absurd, that, uh—¶

Why?¶

The one we run is unlimited.¶

Xiaomi one.¶

He, uh, OpenCode — he's been using it all along, right?¶

Nobody even talks about it.¶

Its cost is low — even counting the API money, it's still inside the line. It just sits there all alone; the kill line keeps shifting, the circle keeps shrinking, but it's forever inside it.¶

It just never changes.¶

So impressive.¶

Doesn't seem like anyone's talking about it lately.¶

It's several times cheaper than DeepSeek Flash.¶

Cheap — dirt cheap.¶

Also, why did both of them release one? Gemini and Muse both put out those speech transcription models.¶

This one's pretty minor, really — just done on the side.¶

I think the core is that Google, lately, needs buzz, but its LLM lineup really isn't doing great, so it has to have something to release, something to make news with.¶

And another thing — what does this tell us? The DeepMind situation is still, it's still DeepMind's — there's not much pressure on projects; they let you have a team working on this.¶

Not integrated yet.¶

If it were anyone else — OpenAI or Anthropic.¶

They'd never let you release this.¶

As if they'd let you take money and GPUs to work on this — they're grinding on the main thing every day and shipping daily.¶

Though for them, this is a tiny amount of resources. Actually, look — in China, Doubao's transcription is already the best; there's basically no competition left.¶

Qwen's release is pretty good too.¶

So Google's positioning is basically "the American Doubao" — that explains a lot.¶

Ouch. That stings.¶

It's a total mirror image, right?¶

It really is mirrored — look, Shiyi and their team are doing it too.¶

But Qwen's release, CozyVoice Studio, right — seems to be a complete audio solution, and it's actually pretty good.¶

What's the core issue here? I've been building that video-editing Skill lately, and what did I find? These models — other than LLMs, the ones related to multimodality and content production — are way too much of a hassle to use.¶

I think so too.¶

At the very least, even in a Coding Plan — or give me a multimodal Coding Plan so I can call it quickly, or at least try it out. But there's nowhere to use it. Digging through an ocean of documentation to find the model's name — oh my god.¶

They're selling it B2B style.¶

Hey — if folks from the vendors are listening to our podcast, take note right now: we've already covered so many good opportunities and ideas in this episode.¶

When we partnered with cloud vendors, we found the cost of integrating one model was just too high. After finally finishing the integration, we discovered the first-token latency was 100x someone else's — terrifying.¶

So I think these middleman relay services will always have opportunities across industries.¶

Exactly — the relays do it better than the cloud vendors themselves.¶

But the biggest problem with relays is that most of the people running them are short-sighted — they never seriously make the relay itself a good product.¶

Some have done it well.¶

Done well.¶

Right, didn't I mention one last time — a YC-backed one, a relay for speech models.¶

Hmph.¶

See, isn't Pika doing that now too…¶

Pika did it too.¶

They picked a great direction: multimodal — covering, besides LLMs, all the video models and image models.¶

Pika was founded by a Chinese woman, remember.¶

Right.¶

They've been quiet for a long time too.¶

They've started selling a wrapper app now.¶

They pivoted.¶

Pivoted.¶

See, Orange's choice was the right one. Orange, walk everyone through your business logic.¶

No, wait, I…¶

I've said it before: these days, don't be afraid of people calling you a token seller. In the AI era, if you don't sell tokens, what do you sell?¶

Yes, yes — same for me.¶

Look at Mian-shen — Mian-shen is so far ahead, and he still gets flamed by everyone. I don't know what those investors and self-media types are on about.¶

Right — it only clicked for me this past month. I made a round through the Greater Bay Area, talked with the founders of several startups, and everyone's like, how to put it — if you're in the AI business, whether you're on the application layer or whatever, don't be shy about positioning yourself as selling tokens. We're all working for the token sellers.¶

Don't treat your engineering capability as a moat.¶

There is no moat.¶

Everything we do up front is to get users to buy tokens from us.¶

Nowadays a single person…¶

You use tokens through me — say the cloud vendors can't get you connected, and I provide you one your Agent can call directly. Isn't that an advantage?¶

This is what you call experience arbitrage.¶

Exactly, experience arbitrage — the cloud vendors are just too hard to use; the individual products are actually easier.¶

Let me add a recent take of mine here. Earlier we said nobody has time to try that many Agents and that many models anymore. I now have this feeling,¶

A model vendor's model and its Agent are a combo, and these days I like using them as combos. If I'm not interested in the model, I probably won't use its Agent; if I'm interested in the Agent, I'll probably try the model along the way.¶

These two things, slowly — think about it — going forward people may say less and less… In other words, there are two kinds: Agents made by the vendors, and neutral third-party Agents. I'm increasingly leaning toward picking the vendor's native Agent plus the vendor's model.¶

And not just because joint training makes the quality better,¶

Is it just "love me, love my dog"?¶

Just yesterday people in a group chat were discussing: in ZCode, using DeepSeek, the cache hit rate can reach 99%. Then someone else said running DSH inside DSH is 100%.¶

Setting aside who's right, I wonder whether consumers will also form this kind of habit going forward.¶

Definitely, because you…¶

"The broth that cooked the dumplings digests the dumplings."¶

And another thing: its API experience, or the optimizations in the product itself — building an Agent looks simple now, I think. But once you use one long enough, you find it's really hard. That's why everyone uses Codex so much — it really is good to use. It's hard to imagine it being able to…¶

Very smooth.¶

In every respect, it's top-tier.¶

Let me tell you one thing: where Codex is far, far ahead is its self-awareness. Codex has dedicated tools — you can pull up all its Projects and Sessions and check their progress. Through conversation with Codex, you can tell it, for example: "My branches and worktrees are a total mess — I keep opening new conversations. Go through everything in this project and clean it all up for me."¶

It will go through the Sessions one by one itself. I've even told it, within one conversation, to open a new session on another Project, or to close one of them — it can do all of that. And this feature has existed for ages, while Claude Code only shipped something similar in the last 2 weeks.¶

And it doesn't do it well either. Yesterday there was another situation, about its self-awareness. Codex has an old problem: some long conversations keep reconnecting, but new ones don't.¶

I mentioned that some long conversations had reconnections, and it pinpointed that exact conversation and read only that one. Because reconnection events aren't… there's no way, they're hard to search. They're things stored in logs; it's a network state.¶

It found it and said, "this is the one." I looked — yep, that's the one. Then it dug into why the reconnections were happening: conversation too long or something, or maybe an IP issue.¶

When they designed this architecture, they actually had model training in mind from the start — everything gets stored.¶

And it's been iterating for so long. There's another thing: constant rapid iteration. Compare the iteration speed of domestic Agents with its — in one day that update prompt in the bottom-left corner pops up eight hundred times.¶

Tons of it.¶

And I don't think their engineering team is even that big — very few people.¶

And it never breaks from iterating either.¶

You can clearly feel Claude Code has way more of a garbage pile of code than Codex.¶

Way more. And Claude Code is CLI-driven, which in theory should make it easier to iterate, yet it's now much more of a hassle than Codex.¶

ZCode iterates fast too.¶

It does.¶

Right, my daily driver now is probably ZCode — ZCode has become my…¶

ZCode.¶

Alternative. But ZCode has this problem: the other day I used Niulai to try building something like a Three.js multimodal thing, to see if it could be done. Multimodality is pretty token-hungry, and its multimodal part burns tokens fast, so your quota runs short. I'm on the mid-tier Token Plan.¶

In its Token Plan, aren't the model prices multiplied several times over? More expensive than the API.¶

The way I understand it, it's a bundled thing, so they have to balance the quota across every model.¶

Or maybe they should set it for mid-tier users. One more thing: the subscription lock-in of the vendor's own Agent.¶

Still convenient, very convenient.¶

And there's another thing,¶

Like the Omarchy Linux mentioned below — after installing it I found, damn, I couldn't handle any of the configuration, and at that point the network wasn't even set up. So I installed ZCode and had it do everything for me, since it connects really smoothly from China.¶

Speaking of which, I really don't get it — this Omarchy thing, why did it suddenly blow up?¶

It took off with 4.0. You see, the Ruby on Rails guy — the author and all that — and they promoted it really hard, and honestly… 4.0 — I never paid attention to 3.0. Last year lots of people installed that 3.0 version; some praised it, some trashed it. But 4.0 is genuinely much more polished. A lot of the people who trashed it last year now feel this can be a daily driver. And its design is Agent-first — how is it Agent-first? After install, almost every Agent you can think of is built into the system, from—¶

A whole pile of CLIs stuffed right in.¶

Claude Code — everything's added in.¶

You get in there and set up a Token Plan.¶

Then I've got to try it.¶

Another really interesting point: you know how Linux crashes easily if you don't know what you're doing. So it ships its own CLI, and hands control of what can and can't be changed in the system to the Agent. So wherever the interface crashes — even a black screen — the Agent can fix it and rewrite it for you on its own.¶

You just bring up a Terminal, and as long as you can launch it and type, it can repair things.¶

OK, and there's a philosophy behind it I never quite got. You know how we normally open tons of windows — you often need one window on the left, one on the right, even windows stacked on windows. In that system, everything you open is a new window; even in Obsidian, click Settings and boom, another window pops out. I really wasn't used to it. Later I asked why it's designed this way, and the idea is that people shouldn't spend a lot of time resizing windows — so everything is, in effect, precomputed.¶

You mean everything goes fullscreen.¶

By default, one window is fullscreen, 2 splits the screen, 3 becomes a three-way split, then automatically four-way.¶

Auto tiling.¶

To use one of them, press F and it swaps to fullscreen; press again and it goes back. That's how it cuts switching costs — and everything is keyboard shortcuts.¶

Tiling.¶

I'll go try it out then.¶

One more thing: because it self-evolves, writing software on it is super convenient. Why do so many people play with it? Because it writes a lot of its own stuff, and writing on Linux is simple.¶

You know how Macs have multiple desktops? With multiple desktops you never even know which one you've switched to. It starts out normally listed 1, 2, 3, 4, but you can tell the Agent to name each desktop for me, and it goes and changes them all for you — extremely customizable.¶

Very low-level, plus AI Native.¶

An AI OS.¶

AI Native — a very AI Native OS.¶

But people will laugh at it too, saying: with all those distros out there, I just add an Agent and isn't that the same?¶

Actually it's not quite the same — they've done quite a lot.¶

It really isn't the same — it's way too fat.¶

To experience it natively.¶

Actually from this angle, it is itself a Harness. A giant Harness.¶

The thing is, not being able to install Linux on a Mac is annoying. If a Mac Mini could run this system, that would be amazing.¶

You don't need to buy a Mac Mini — just buy a barebones machine, there are plenty.¶

Windows machines are indeed a bit worse on price-performance.¶

But the experience has its pain points, since support in China isn't great. He said they're about to launch a China official site plus domestic mirrors. When I first installed it — who was it that said they'd do this?¶

Domestic mirrors.¶

The foundation.¶

Because in China lots of people, you know…¶

A domestic Linux has very good odds — enterprises and all that, government and state-owned firms.¶

After installing, the first thing was setting up an input method, but there wasn't our usual one, like Squirrel or anything. So it was bug after bug after install, and only AI could fix them. Eventually Squirrel worked, and it felt pretty good. And WeChat installs too.¶

It's a bit like that OpenClaw period — back then everyone had to fix things too.¶

Eventually it installed.¶

Kind of that same state. But if you like very high freedom and Agent-first control, I'd say this is the first genuinely appealing system. And the interface is really… I think design matters a lot here — whoever made it probably has good taste. It has a built-in…¶

Quite pretty.¶

There are some design principles behind it. That split-screen you mentioned — it's like the Mac split screen: all those windows look great, but for regular people it's painful in practice. Adjusting it… at my company a bunch of people can't adjust it at all — can't use shortcuts to tile the screen or half-screen and so on.¶

I can't either, I… I can't either.¶

I just use shortcuts directly. OK, let's recap. On the models section — see, we've barely touched Hunyuan, and we talked little about Fable 5.1 too…¶

Specifically.¶

Hunyuan is decent.¶

Hunyuan.¶

Their own GPUs aren't even enough for themselves — at 8 a.m. I was queued at number four-thousand-something. So why do they iterate so fast? Last month they just released Hunyuan 3, and this month it's already Hunyuan 4 — one version a month.¶

The architectures are probably completely different between the two. OK.¶

Different parameters — then the pretrained model is definitely different too.¶

Got it.¶

I think they're probably just trying to catch up to the table first — let's not talk about the rest.¶

I've also noticed something: model vendors now release Preview first, afraid of a reputation collapse — that's one pattern. Another is that Qwen, going forward, won't even have version numbers. See, this is 0902 — Qwen shipped 0902 — saying Qwen will roll out continuously. Doubao Seed is the same; Seed too: no more version numbers from now on, continuous online updates.¶

The evolving edition — the evolving edition.¶

There'll be a testing edition that's always new — just assume it updates several times a week.¶

Niulai updates too — new architecture and all; that's how it works, it can keep going.¶

Niulai, from initial release to version number, has been giving us this all along.¶

So that's constantly folding post-training into releases. No — but it still puts out a version number externally; this is about never publishing version numbers again.¶

Qwen's model release interval has already shrunk to half a month, 20 days — that's about the limit, right? No, not Qwen — this one.¶

Not that important. It's, in a way, flipping the table — going rogue.¶

Everyone's like this now, I think everyone is.¶

But too frequent becomes a problem: people run out of attention. If you release too often, nobody cares anymore. Fable 5.1 is an example — honestly nobody really paid full attention to it.¶

Right.¶

I think Orange has a point. On one hand, doing it this way makes benchmarking impossible. But not being able to benchmark cuts both ways: first, you can't call it bad; but conversely, you can't really call it good either.¶

This comes back to the earlier question. I think frequent releases happen because you need frequent iteration and feedback collection. The core issue is you need a closed beta.¶

Take red-teaming — isn't red-teaming mostly about safety? On the capability side, I think we may also need paid evaluation.¶

But how will people pin things down going forward? People used to say such-and-such version of a model was strong or not, but if you keep rolling updates continuously, I think that whole notion gets diluted.¶

You shouldn't release to the public at all — you should have a closed beta, say released to a few thousand people. Like having core members of the core Agent community do paid testing.¶

I raised a KOL point the other day too: when you look for a good KOL, you're not after their prose — it's not that this KOL writes well, it's that this KOL's cases are good. When you pay them to post, you're really paying out money for their Case. So you might as well pay directly for Cases.¶

I think the mindset needs to shift: when evaluating, find the right evaluators, have them test, and pay them what they're owed.¶

How should I understand this? Recruit more people to test, then sell the good cases and showcase them — is that it?¶

Or rather, when you join the test group, your Cases get archived by default — meaning you're getting paid while you test.¶

So it's about archiving — spend the money finding good, third-party independent benchmarks. Even if the results never go public, you hold them.¶

Qwen found a sponsor, AI — they run 10 benchmarks for you, and you look great on every one of them.¶

Independent. I think that one is quite meaningful.¶

Quite interesting — whether it's meaningful, who knows.¶

Because public benchmarks already feel pretty meaningless.¶

Meaningless.¶

Let's move on to Agents — let me gather this week's, well, not exactly Agents, let's say product-related stuff.¶

One is — I think it was yesterday — WorkBuddy held a launch event, pulling together its whole big WorkBuddy ecosystem, software and hardware, must have been dozens of partners, all joining the event together.¶

Also, this week Qoder released a New Qoder. It felt like — people used to say Alibaba's Qoder was an IDE, right? Now it may have turned completely into an Agent — though I haven't tried it in depth.¶

And there's one drowned-out story: OpenClaw 2.0, released this week — there's barely any… news coverage is already very thin, and what I did see was people still complaining about it.¶

The news I saw was that Meituan is making everyone company-wide reflect on why they caught the Lobster.¶

We can talk about that a bit later.¶

And they forced it before. Now they're saying it's all on you guys.¶

That one's probably not great to dig into in detail — later.¶

Then there's DeepSeek Harness. The hype has passed, but what I see is it keeps iterating versions. In my group chat, remember, some people were confused about what the versioning rule for DeepSeek Harness actually is. In fact there is a rule — releases follow Semantic Versioning — it just keeps shipping new versions.¶

So this is also why people can't really get going with the DeepSeek Harness — its plugin system. When I was building that client of mine, its plugin system…¶

Updates constantly.¶

It's — I think to some degree they've shifted the complexity onto developers, and with such frequent updates, my plugins can never stay compatible with yours.¶

It's breaking changes.¶

Don't worry — it hasn't stabilized yet,¶

No, it can't stabilize — I don't think it ever will.¶

Lots of people ask why my library isn't updated. I tell them I can't keep up.¶

So you end up making your own release.¶

No — it's just like OpenClaw. Back in the day, every OpenClaw update broke everything for everyone, and it didn't care. Like I said, I can't keep up with it.¶

That's fine — it was never made for you in the first place.¶

No difference.¶

You're just too emotionally invested.¶

It'll be usable once AGI arrives.¶

Also for this — just let the model self-iterate and be done with it. No — wait, no. Make it into a Skill, and the Skill just…¶

No.¶

It updates, you update.¶

No — the Skill just reads the gap between two versions and then fixes whatever you already had…¶

No.¶

That's a bit idealistic.¶

No, it can't be fixed — there are so many bugs. It has bugs itself; they can't even guarantee their own quality, that's the main thing.¶

Your base itself has bugs, and your plugins on top of that — you're saying it can rely on reading the gap? Impossible.¶

It's a beta to begin with — a perpetual beta.¶

No, it's a test preview — a developer preview.¶

One minute Alpha, next minute RC.¶

Just slap a Preview on everything.¶

Slap a Preview on everything.¶

Everything's Preview.¶

OpenClaw — right, I saw the complaints about OpenClaw: OpenClaw is no longer the OpenClaw it was; it's even built in capabilities like Feishu. This is really two sides converging: people noticed that after OpenClaw launched, our old Codex, including ZCode, has actually been absorbing a lot of OpenClaw's features. And conversely, OpenClaw, and this Hermes and the like, are also absorbing some of Codex's capabilities. It's really converging in both directions,¶

Anyway, when it shipped, they said the experience and everything had improved, these problems were fixed, we have an interface now, better experience. Then a while later you refresh and there's a wall of complaints: still garbage, still unusable.¶

It's even been consistently worse than Hermes.¶

Hermes at least is run by people who make products, right?¶

Let me offer a different view, from my own experience, because I find it so hard to understand: everyone says their OpenClaw keeps crashing, but I've used it for half a year and it's never crashed for me — I don't know why.¶

And the second way my experience differs from everyone's: I switched to Hermes last month, and afterward I felt a huge gap versus my original OpenClaw. Let me give one detail: with OpenClaw I go continuously — sometimes when I hand over a task I need to send several messages, and OpenClaw naturally understands it as one thing, without stopping in the middle or interrupting me. Hermes, on the other hand — you're talking to it, you're talking to it, say I usually send a photo and then explain what I want done based on the photo, and it just — it frequently stops. I found it extremely choppy. And I've never had to repair my OpenClaw; it's been running, running stable all along, and I don't know why.¶

You still use it?¶

Some tasks — here's my setup: I've hooked it to my WeChat, my WeChat Bot, and sometimes have it complete special things, ha.¶

If you're just having it look up info for your weekly report and such, you won't — no real impact.¶

It won't go down either.¶

Your usage pattern is the difference — people used to keep adding models and tweaking model configs, and all that fiddling is what breaks things.¶

Changing configs is the easiest way to break it.¶

Those configs have zero protection.¶

Well then I just shouldn't fiddle with it.¶

You can wreck the gateway.¶

WorkBuddy — why does nobody seem interested? The WorkBuddy ecosystem looks like a big production — what does everyone think?¶

Mainly, there's nothing there to use.¶

Did you also leave out that Doubao Work?¶

The ecosystem is just…¶

Doubao Work probably didn't do much last week, right?¶

Doubao Work's problem is it looks too much like Doubao.¶

The week before last, right? The week before.¶

What I mean is, it's in a corner nobody watches.¶

No, listen — is it Doubao Work or Doubao Office? People keep writing it wrong. A couple days ago I did a livestream on the Wu Xiaobo channel, and after the stream they wrote a news piece and wrote Doubao Work as Doubao Office. I told them to fix it. The name just hasn't claimed any mindshare. Trust me, plenty of people are going to keep calling Doubao Work "Doubao Office" from now on.¶

Does it matter? People even get WorkBuddy wrong.¶

WorkBuddy.¶

See, both words are wrong.¶

The one with the person in it.¶

Nicknamed "Outsourcing."¶

Outsourcing.¶

As if it even deserves to be called outsourcing.¶

WorkBuddy — people write the "zou" as in walking, and the Body as in the body of your body.¶

Yes — the walking body.¶

Work.¶

Maybe because I did a lot of ecosystem work before, I'm sensitive to ecosystem matters.¶

The way I see it, at Doubao Work's scale, judging by the splash — there's no splash.¶

I think maybe they're still holding back a big move. Among these players, its gap in pacing is the biggest. Qwen moved fast and quickly pulled off some team and product consolidation; on the Doubao side, I think they dawdled for quite a while.¶

Their moves were fast though — they merged a lot, like that TraeWork, and Coze, right?¶

Pretty fast — Coze had just shipped an update, then immediately got merged, and the update got rolled back.¶

Immediately. So their internal org probably still needs some time to settle, because everyone's moves are…¶

The moves are quite decisive though.¶

Execution is fast.¶

I think the Qwen side is pretty fast.¶

Even more absurd — last week didn't I attend that TraeWork event, and on the event badge…¶

The badge said TraeWork.¶

TraeWork.¶

Bleep that out.¶

They removed it.¶

And it got scrubbed off.¶

Hilarious.¶

It was pretty sudden.¶

Nobody's interested in this.¶

Let's move on.¶

Actually there's one I didn't write down, because it had no news last week, but I feel it's stayed hot: Grok — Grok Bot.¶

Should be.¶

That one's pretty good.¶

That Grok Bot is quite impressive.¶

Grok Bot released an Android version this morning.¶

It's an Android client, right?¶

Android client. Wasn't it iOS-only all this time? But it still lacks scenarios — good use-case examples. I think it's still missing those.¶

The way I see it, Grok Bot is, in my mind, the form of the future.¶

I think its potential is huge.¶

Think about how many people still have a desktop at home, versus those left with just laptops. How many people keep their laptop on forever? You're always having to close it, close it.¶

But going forward, AI in our lives really needs to be online 24 hours a day. So where does it live? In the cloud, obviously…¶

So—¶

And that cloud has to be persistent. When I use it, you must give me a 100% guarantee that everything on it will never be lost.¶

Never lost.¶

It even happened before with a certain vendor's Agent — you're using it and everything's just gone. No names.¶

All those virtual machines.¶

A lot of VMs these days shut down like that, or are dynamically allocated.¶

Lots of them are VMs — remember that original one.¶

Manus VM — but the Manus VM doesn't shut down.¶

That was me! Really, credit to me on this one. Seriously — didn't Manus say giving the Agent its own computer was the more sensible take? You should dig that up — a prediction that's been ahead for so long.¶

I'd say ahead by the better part of a year, at least.¶

A year.¶

The idea was ahead, but it never achieved permanence or persistence — it's one computer per task, without persistence.¶

If we're going there — who here has used that GPT Cloud?¶

GPT.¶

ChatGPT Cloud actually also gives you 1 machine.¶

Chat—¶

GPT Work's Cloud — also 1 machine.¶

Work. Yes, it has one — that machine's specs are huge, 12 gigs of RAM.¶

But see, Grok Bot gave it a name — a whole product sitting right there. Whereas what we just mentioned, the Cloud inside ChatGPT Work, is already buried three layers deep,¶

Didn't YouWare do that thing of theirs? And Bloome was doing this kind of thing very early too.¶

I think the core is experience and cost.¶

I think Grok Bot is just so simple — it looks simple, and the interaction is excellent.¶

Like I said just now: before, it was ChatGPT — why did everyone pick it up so easily? You just chat. Then it became Agent interfaces, and look at today's Agent interfaces—¶

So complicated.¶

Right — open today's Doubao Work: connectors, a whole pile of skills, my god, you don't know what any of the words mean. Nobody recognizes any of the concepts. Grok Bot is just—¶

Chatting. There's nothing else — just chatting.¶

Nothing there, but all the functions are inside.¶

And I think once you have that 1 computer, you discover it can do lots of black-magic things. For instance, I said "go test the top ten products on Product Hunt one by one," and it literally opens a browser.¶

Installs.¶

What OS is that computer running?¶

Linux — looks like Linux to me.¶

Linux. But it can hand it over to you anytime — I log in with my account, and once my Google account is logged in, it effectively has my Google account's permissions. From then on, everything it does, it signs in with that Google account to test.¶

And it really went site to site, signing in with my Google account and entering test cases.¶

See, our whole journey of using the internet, mobile internet, AI is one of gradually releasing our privacy and various permissions. It's all traded out.¶

A process of concession.¶

Nobody cares about that.¶

That doesn't matter — you trade it out to get the features and experience.¶

Don't care.¶

And it's convenient.¶

And its flow is really good too. Like you said just now, it needs your Google account. With what we call connector MCPs, you have to connect them manually — if you haven't, and this is how many domestic Agents are — you tell it and it says "I don't have that." It doesn't even know it has that connector; its self-perception is broken.¶

Whereas with Grok Bot, you say "pull up my Google mailbox, let me check my email," it asks which emails, and then Google pops up a card right there.¶

So you log in.¶

You click authorize — okay, authorized.¶

It's progressive — it only surfaces when needed.¶

This is something I really want to rant about.¶

For example, I tell any of my Agents: help me plan a route. Say I'm going to Shanghai this week, meeting a few people, each at a specific time and place — plan it for me. It needs to call the Amap MCP, so for this one task I have to go register an API key with Amap, come back and configure it. Way too much hassle. Isn't this exactly the thing you should be handling for me?¶

So on this — on Memory, on context acquisition — why did products like this appear overseas first? Because look: even Google, with its full family bucket of apps, gives you a really convenient MCP for all of it.¶

APIs.¶

In China—¶

Isn't WorkBuddy already working on this in China? Isn't it being worked on?¶

What I mean is — Qiaomu said "progressive" just now. When I have this need, you should handle the infrastructure connection with Amap for me, the authorization, even the payment is fine. You should solve it for me, not make me go deal with this stuff.¶

Even if it burns more of my credits — at that moment, if you tell me it costs extra, I don't mind.¶

Deduct my credits, whatever — tokens are tokens. These things, I think, these products absolutely have to solve them in the coming year. And I might use this once within a year — you can't possibly make me go configure something for that.¶

Didn't I share a take a while ago — that in a '27 product, if the product still has MCPs, Skills, and what else, including Plugins, Connectors, Keys,¶

Skills — those pages, dedicated pages.¶

Exposed to the user, right.¶

None of these should exist in a '27 product.¶

Agents are just too complicated right now. Going forward they need to simplify all of this — it's still not smart enough.¶

And I think the trend is pretty clear now: everyone's building Grok Bot-like things. But the core, I think, is that China is slower, because our—¶

The APIs don't connect.¶

Context is the — context is the most serious problem.¶

Openness is just too poor.¶

So the AI era pushes back on all this, forcing these ecosystems to open up.¶

They've already given ground once. See how before, hardly any CLI let you read and write freely, all that Feishu mess. If they don't do this, usage will just shrink further.¶

I saw a news item,¶

It said one of WorkBuddy's biggest use cases is stock trading. It hooked up some stock-data provider, and that data can only be read by an Agent inside WorkBuddy — nowhere else.¶

Other AIs aren't allowed to read it.¶

So people swarmed into WorkBuddy to use this feature. And when Doubao found out, it plugged into one too.¶

Right, connectors — in Doubao's connectors there isn't even a category; switch to stocks or switch to finance and you're instantly several horses richer. Wow, it's fierce.¶

The point is, copy the conclusions.¶

These are the real users.¶

Didn't they publish stats on queries before? It was stock trading, fortune-telling, stuff like that.¶

Fortune-telling, right.¶

But that's not work — that's not office.¶

All the office Agents.¶

Not Office — yet everyone's building Office.¶

Actually, let me flip it: in office scenarios, what is there that you need accelerated to the point where the efficiency gain produces actual returns? Take lots of white-collar work, all that paperwork — weekly reports and such. Speed it up 100x — what good does that do? Honestly, none.¶

Think about what office scenarios mean — what comes to mind is converting PPTs and PDFs back and forth, handling documents.¶

I did some research recently, you know — programmers are more tired after using AI, but non-programmers, the PPT writers, the doc writers, all started slacking off after adopting AI. They used to work 8 hours; now it's one or two.¶

Mainly because the documents written before were the same old boilerplate anyway, and now they're still boilerplate — doesn't matter, no change.¶

But coders, you know — the more they write, the more tired they get.¶

It's bugs — the more you write, the more bugs.¶

It's not that the writing tires you — it's that you've opened too many projects, you feel you can handle them all, and monitoring the AI yourself is what's exhausting.¶

And more bugs too.¶

Right, the bugs.¶

Let's see what other topics are left — let's talk hardware. On hardware, I've rounded up a few hot items from the past 7 days: one is Hugging Face's MicroDuck, that little duck robot; also Apple's newly released Mac Mini M6 and Mac Studio — not strictly last week's news, but the high-end configs have climbed to tens of thousands, even over a hundred thousand [yuan], with memory going straight to 128, 256, 512; and Plaud released new earbuds too — they hadn't shipped a new product in ages.¶

Look at this MicroDuck — it blew up overnight, but it's actually made by a Chinese manufacturer called Seeed.¶

Don't a lot of trade shows have these little robots — the blinking-eyes kind?¶

But its design is clearly good, and the demo is great. The cost would be too high otherwise — this duck can pick things up with its beak, whereas if an ordinary little robot did it with hands, it'd feel unremarkable. I think this thing is, one, cheap enough, and two, fun. Compared to humans it's not actually that useful.¶

Fully custom, top-tier. Before, it was all backflips and this and that—¶

But my take is: it went viral because Hugging Face released it. Try having someone else release it.¶

Not necessarily, though — not necessarily.¶

No — it was released by a company Hugging Face acquired, and that company's earlier products were already hits before the acquisition.¶

The earlier one was that desktop mini robot, right?¶

What are these toys? One: fun. Two: useless. It has to be useless enough to be fun.¶

And it's cheap — at this price…¶

Not cheap.¶

Pretty cheap.¶

Over 300 US dollars.¶

You know what other…¶

Chinese robot prices.¶

No comparison there.¶

In terms of technical substance it… ah, forget it, never mind.¶

No, everyone's tech content is about the same—¶

At least this one can do backflips on its own, while Chinese robots still need a remote control.¶

But there will be more and more of these little gadgets. The phenomenon itself — I just look at Xiaohongshu and feel like an era of everyday inventors has arrived. Anyone who wants to build something, these days…¶

That ESP32 dev board has basically become the new LEGO. Sold out every day, never enough stock.¶

Look at Wang Le and their stuff — it's all…¶

What he's using isn't even a particularly high-spec version. That thing of Wang Le's — next episode we'll bring it in and show everyone; it's really nice.¶

That hardware — mainly, people treat this thing as a toy.¶

Basically LEGO.¶

It also makes a great gift, the design is pretty. I think if you go to a robotics expo, this is the best-looking design there.¶

You spot it at a glance.¶

Did any of you go to that World Robot Conference?¶

No.¶

No.¶

Not much to it.¶

I went to take a look anyway — I thought it was nothing much.¶

Everything's way too homogeneous.¶

So before…¶

But wasn't there a video clip online that went viral, of a robot's running gait — the kind that looks like a wom—¶

So before — that day I saw the Tsinghua one, let me think, that demo we covered.¶

That was the sports meet, that was the…¶

Held at the same time.¶

I think they're seriously underrated. Before, who was it… robot models — I looked at the local model, the one I recommended the other day, the one Qiaomu introduced earlier, made by a Tsinghua team.¶

This was the first time I'd seen something that truly needs no remote, can really communicate with me, and can be embedded in any robot. The most absurd part: embedded in a motorized wheelchair. That wheelchair isn't a robot — it has no brain at all. Strap a local GPU's compute onto it, plug it in, and it can automatically go…¶

And a robot buying me coffee, right?¶

Now that's a brain.¶

An embodied brain.¶

And I think this point matters — also a local model: in a corner nobody watches, Perplexity released a local inference engine.¶

Perplexity.¶

After you install Perplexity Comet, it automatically installs a local model on your machine, tunes it itself, then lets the local model handle certain specific tasks. That way, the vendor's overall costs come down.¶

That makes a lot of sense.¶

Chrome has a Nano — Gemini Nano.¶

That one's way too small.¶

A small model you can download online.¶

That model is too small. It downloads one of its own, probably something like a 6B or a 27B model,¶

Once it's deployed — there seemed to be some other optimizations too; I didn't look into it closely.¶

Let me ask everyone: do you currently have, say, a Mac Studio or anything, and any plans to buy one?¶

Right now I have a Mac Mini and a Mac Studio.¶

How much memory does yours have?¶

And another… The Studio I bought is just the regular 32G. I don't want to run local models, I…¶

32G. Then what's the difference in positioning between a 32G Mac Studio and a Mac Mini?¶

Mac Mini still doesn't cut it — cooling and performance genuinely aren't enough; it easily melts down running engineering workloads. Like opening dev environments and such — open too many and it still chokes.¶

At our company we run some tests on it — you can only have 2 going, two pipelines.¶

Once I had a memory leak and it went down. A Mac Mini with a memory leak can hold out a while longer.¶

Anyway, the one I have — I was really lucky, I ordered before this year's Spring Festival, and…¶

Same here.¶

It's appreciated in value.¶

Yours appreciated — 128G of memory for just 29,900 [yuan].¶

Actually its CPU isn't the most important part — it's the memory.¶

You know what the felt experience is after using such a machine? You stop feeling the computer's existence. Everything is buttery smooth; all you feel is software — you can't feel the hardware.¶

The hardware has no bottleneck; you have no perception of it — no sense of the hardware acting up or anything.¶

That part is glorious.¶

And mainly it's always online. My Studio hasn't been shut down once since the day I bought it.¶

Same for me, never shut it down. So you two — any plans to buy a Studio in the coming year?¶

I have a Mac Mini; I don't have that need for now. I'm thinking of switching to a better laptop with more memory.¶

Not for me, for now.¶

I feel this connects to what we discussed — say, Grok Bot. It's really just like having one in the cloud; the trend may simply be the cloud.¶

You want one thing.¶

But I think maintaining local memory and local context is still really important. I feel there will definitely be products for this — quickly keeping that stuff of yours always there. But the current ones do it badly.¶

You need a sync mechanism.¶

Or you…¶

But you can't just — as you said — give up on privacy and go all cloud.¶

But the problem is the AI landscape shifts sometimes: you gave it your stuff, then you can't get it out, and meanwhile something else really is better, and you, you can't switch. So you still need something local, a—¶

I heard the new Coze version is basically a cloud Coze that operates your computer — once it's connected you can control it from your phone. That's pretty nice.¶

And locally there's also the network-drive angle. Cursor is about to release that new Agent too — one that can be hosted locally, the Cursor Agent, then connected to the cloud. Similar to what you said.¶

Grok Bot is made by the customer-support side, isn't it?¶

Same family as that, but a different product line.¶

It should be the trend.¶

Actually, an Agent product that manages my local hosted environments for me and connects my local setup to the cloud as part of it — that would be great too.¶

So I think on the NAS question — a NAS can do this too.¶

Getting geekier and geekier.¶

Did any of you buy the Xiaomi NAS?¶

I bought one.¶

I bought mine during the beta.¶

Oh, they'd definitely buy one.¶

How is it?¶

It's just a regular NAS — it doesn't even have Docker,¶

I used Xiaomi's old router-style NAS — it was actually decent, Xiaomi's UI and all.¶

How about compared to UGREEN, and what's that other company called?¶

UGREEN. There's nothing to tinker with on it — the Xiaomi one just has nothing to play with, not great. It doesn't even have Docker, its specs are terrible, it's just a normal one — use it to move files and that's it.¶

And there's Synology.¶

That one — Synology, people tinker with that a lot.¶

Synology is just too hard to use.¶

You've used it?¶

I gave mine to the company — way too hard to use.¶

I was thinking, now that I have the Xiaomi one, I'd write my own mounting thing. My local memory is really a filesystem — databases and documents. I write another mount and it syncs automatically; I'm not syncing to that private GitHub thing anyway.¶

Same for me now — syncing my own memory to GitHub is too… just enormous, way too many files.¶

And there's the multimodal problem too: I play with images and videos a lot, and I can't upload multimodal images and videos to GitHub.¶

GitHub isn't that reliable either.¶

So I think the NAS is better, and the NAS—¶

This week I met up with AJ, and AJ said syncing all his various data has already worn out his drives.¶

Too much reading and writing.¶

AJ probably keeps all his raw multimodal data every single day.¶

Exactly, and I think this is important — cloud drives should work like this in the future.¶

Storage is getting more and more expensive.¶

So look at Xiaomi — I didn't write this one down — that Xiaomi chip called O3 and O100, right? The one Xiaomi is about to release.¶

The third one — O3, D100, and O100, 3 chips.¶

That thing, combined with a Xiaomi NAS, all sitting at home together.¶

HyperOS 4's Super Xiao Ai now has Memory. I mentioned this in a group chat the other day — what's its Memory? It means when you favorite something on Xiaohongshu or Douyin, that favorite automatically lands in Xiao Ai's Memory.¶

And there's this thing called the Xiaomi Sync Assistant. It has Mac and Windows versions, right? The current Mac version of the Sync Assistant has Super Xiao Ai built in — it's an AI — and the Super Xiao Ai on your phone can operate your Mac through this.¶

So you're effectively getting some of Xiaohongshu's data through the side door.¶

You're saying this is all implemented on Android.¶

Its own Xiaomi system, connected to your macOS and Windows this way, so you can operate your computer through the AI on your phone. Even if the computer has nothing to do with Xiaomi — just your Mac — and it holds all your photos, and in that app you can view all your photos.¶

Hey, don't I have an OPPO device too, that also lets you install something on the Mac. I just think the software quality of these vendors on the Mac is terrible — the experience… and the… quality. Setting aside that the features are actually useful.¶

But that 2.0 one — the one with Super, with MiMo — because Super should have fully migrated to MiMo, and the old MiMo Cloud, they dropped it all, so I think that one is decent.¶

Actually, since Xiaomi's Luo Fuli joined, their AI work has been decent. OPPO honestly isn't great,¶

Isn't the Doubao phone supposed to launch in September?¶

Those Nubia ones — this time they're not calling it the Doubao phone, right? Just calling it Nubia, going straight ZTE on it.¶

But that thing feels pointless now anyway — they've blocked pretty much all of it.¶

Lots of users.¶

Mainly it's too closed off.¶

I agree — with Android phones and AI, when we truly reach the mid-to-late AI era, I think all of this will get wiped out.¶

Because Android itself wasn't built for this — its sandboxing is too closed.¶

Let's see what other business news is left. I picked a few highlights: NVIDIA acquired Hugging Face. And there's another story — after SpaceX acquired Cursor, OpenAI announced it was terminating model supply. Then there's the interesting bit in Zhipu's financials, the ARR question: ARR on a monthly basis is — 16 US dollars, 1.6 billion US dollars — and ARR measured on a weekly basis has already reached 2 billion US dollars. My god.¶

1.6 billion.¶

Anthropic — how much is that? Anthropic is less than 10 billion US dollars.¶

Really? 10 billion (B) — just about at 10 billion (B).¶

I think so.¶

No way — Anthropic? No way. Less than 10 billion (B) US dollars. Not the forecast — not the one where they originally forecast 70–80 billion and it actually came in at just over 60 billion.¶

ARR,¶

Really?¶

But it's definitely above 500 — you can look it up.¶

The vibe they seemed to give off was…¶

Then there's the Claude Code 20x plan thing — people have been talking about it lately, saying this Claude Max 20x plan is actually…¶

2x.¶

2x in total; it's the window that's 20x.¶

20x per 5 hours.¶

And people are asking: all this time, why did nobody call this out?¶

It's been constant shrinking, shrinking, shrinking, down to this point.¶

Let's go through these stories first and see who's interested in what.¶

The Hugging Face acquisition.¶

I think it's only right that they're the one acquiring it.¶

I have a feeling — OpenRouter got acquired, Hugging Face got acquired — by our historical pattern, that's a signal that this industry has hit some kind of node.¶

I think people finally saw the value. Before, many were bearish on Hugging Face and OpenRouter — their presence kept fading. And lately the core driver is the Chinese open-source model situation; these two are the infrastructure and distribution layer for open-source models, which has boosted their value a lot, I think.¶

Their value has always been underrated in China, no? Weren't you going to write an article about it? You never did.¶

I was.¶

What's that angle? It'd make an easy article.¶

One sentence: what is OpenRouter?¶

No, it should be one sentence making clear why the others aren't OpenRouter.¶

Actually, the way I feel it — have you noticed NVIDIA is now using its own credibility to backstop the entire industry? Its acquiring Hugging Face is, from one angle, itself an anchor. Including how NVIDIA is now pledging its own credit as collateral for many neocloud vendors — to put it unprofessionally, NVIDIA promises that after so many years, NVIDIA will buy back your GPUs.¶

Financing, compute.¶

It's anchoring the whole market, because the one who least wants this market to crash is actually NVIDIA.¶

They're building the grand infrastructure for AGI.¶

So if Hugging Face ever hits hard times, or needs to unlock another layer of value, NVIDIA will definitely step in.¶

Or you could say it's sparing no effort to keep this market multi-polar.¶

And relatively healthy and stable at that.¶

Keeping the model vendors in balance, so that no single player dominates and turns around to squeeze everyone else.¶

Did you skip OpenAI training that… of their own?¶

That CPU — Jalapeno, right?¶

Actually I think that's related to this too.¶

OK right — what Orange meant just now is OpenAI making its own GPUs.¶

CPU. GPU.¶

Seems like every vendor is doing it themselves. In the future they all certainly will. Claude is doing it, Microsoft, Meta — and I hear Zhipu is too.¶

Really? They're all fabbed by TSMC and Samsung anyway — further down the chain the whole industry becomes extremely concentrated.¶

And with AI, the software layer becomes easy.¶

Design is easy, manufacturing capacity isn't — the fabs are already smoking.¶

That's the problem.¶

What do you all think about the bubble thing? Right now.¶

No bubble to speak of.¶

No bubble. I saw a take the other day saying 15 million people using tokens is already more than supply can handle — saying, give it to 150 million and 1.5 billion people.¶

They say the people burning tokens now are mainly those R&D folks — and how many is that,¶

And even that isn't enough supply.¶

During our Niulai testing, we were burning several billion tokens a day over there.¶

Take the gigawatts of global data centers — they say it can at most double in the next few years, infrastructure at most doubling — the IDC data center buildout.¶

It's a physical limit.¶

Of IDC data centers.¶

The gigawatts consumed.¶

And then new architectures might improve this by 10x, but it's still¶

not enough.¶

Definitely not enough.¶

No bubble — I think it's a state of insufficient supply.¶

Actually I think Stripe acquiring OpenRouter is fascinating too. Stripe used to handle cash transactions, handle money — but in the future, tokens are money. So it makes perfect sense: Token Is New Money.¶

Then tokens represent productivity — all the productivity on the internet comes chasing tokens.¶

See, when you go buy a Mac Studio or whatever, it's like printing your own money — just printing a bit slowly. Seriously, it's like going back to the old days when people minted their own coins: I dug up a gold mine, set up a mint — back to decentralized money issuance.¶

It's barter — we're in a barter-economy state right now.¶

Fascinating. Look — Flash and all kinds of models you can now run locally; essentially, you can mint your own coins.¶

See, you mentioned something just now — a real quotable line: "you can even deduct some of my Token Credits." What's behind that sentence is that it's literally currency.¶

See that Big Genius bar of theirs — once you're in, all kinds of models are free inside the bar.¶

Free to use. Use them all.¶

So maybe in the future there's something like: models like Astra or bigger, everyone can only run in the cloud, and those are extremely smart; meanwhile the local compute, running Flash-class models, is already very good — Opus 4.8-class is enough for everyone — and the middle tier is the one that gets killed.¶

Look at today's models — for most ordinary people, I feel they're already more than good enough.¶

Plenty good enough, plenty good enough,¶

I think 5.3 Flash is absolutely enough for this kind of, this kind of use.¶

No, I think it's enough — truly good enough. It's like that Omarchy we mentioned, right — that kind of thing, lots of software systems becoming AI-native. The local side is really more about serving those functions and experiences. And it's the real, real cloud large models that are responsible for creating complex systems and products — but but turning the mechanical experience of our old ordinary software into an AI-native experience, but but turning the mechanical experience of our old ordinary software into an AI-native experience — that will more likely be done by on-device or somewhat smaller models.¶

Small models,¶

More likely done by on-device or somewhat smaller models. Like you said, renaming windows and such, tidying up my machine — that kind of thing.¶

No need for a great one.¶

An ordinary small model is plenty.¶

And if you use Flash, the monthly cost is only ten-odd US dollars. Very cheap.¶

Mainly—¶

Most people can still afford it.¶

But if most people tried to use it, there still wouldn't be enough — you can't even buy it.¶

But I've noticed a strange thing: people always chase what's new.¶

The old models could theoretically handle it, but people will always want to use the new, the good.¶

I have a clear logic here: these cheap models, including open-source ones, won't affect SOTA models' pricing power or revenue growth. What should grow keeps growing — they're still used to create higher-value things.¶

The new cheap and open-source models are expanding — this is the Jevons paradox — expanding the number of consumers, the scenarios, and the experiences. They're not built to compete upward; they expand downward.¶

Structural penetration.¶

These are two dimensions of the pyramid.¶

And the people below will convert upward too — once they've been penetrated, they start consuming at the top as well.¶

I want to ask: is there a model now — like that Claude 3.5 moment we kept talking about last year, which made every Agent capability stronger? But nobody uses the earlier models anymore, unless there's a special need — like if you're writing prose you might use a pre-4.5 model — and these models even get delisted.¶

I think the core reason is that inference efficiency keeps improving, and as vendors chase efficiency, they delist old models.¶

Today, look — Niulai is this cheap and still far stronger than Claude 3.5.¶

It is stronger, far stronger than that, but.¶

Far stronger.¶

But people still say "I'll use it": if you genuinely need to solve a complex task, you still want a good model to nail it in one pass, with as few revisions as possible.¶

I think that — the people doing the creating are still a special group. Slowly, tokens become a form of consumption; it's not, not used for — creation is creation, consumption is consumption.¶

So at today's model prices, it feels like ordinary people could already use them at massive scale — but it hasn't happened yet.¶

It's that there's no demand, really.¶

No demand.¶

Or the demand turns into, turns into consumption — it's not creation.¶

If you look at ordinary people, they actually have no demand.¶

For us, much of the time it's for making things, for working — nobody else needs it.¶

Like you…¶

People with actual jobs are maybe a tenth of China.¶

10 million programmers — China has 10 million programmers, a bit over — and 120 million office white-collar workers, the ones with computers.¶

White-collar,¶

And that's the ceiling.¶

Plus maybe civil servants — everyone else is just…¶

That's all there is.¶

That's it.¶

That's all of it.¶

I thought about it later: when you're out relaxing, you don't deliberately look for something for AI to do for you — you just focus on doing what's in front of you well.¶

So AI hasn't entered consumption yet.¶

Consumption has entered, actually — short dramas. AI short dramas are growing fast.¶

Short dramas.¶

That's where the growth is.¶

That's Kill Time.¶

So that one's ferocious — precisely because it touches consumption, the consumption volume is massive.¶

Massive.¶

Actually, it's the H3 we discussed earlier that will open up a big market next — H3 is the DeepSeek Flash of that field, Flash-like.¶

Seedance is fierce right now. If you scroll Douyin or Hongguo, you'll find a breakout AI short drama almost every week, and the quality is very high.¶

Its view counts are staggering too — for example, last week's top AI short drama on Bilibili did over 10 million views in a week. And the quality is high. What will it become? I was chatting with someone the other day — they have those short-drama factories. But short-drama factories exist to fill the low-end short-drama volume; you can't view the exquisite kind of AI short drama through the short-drama lens.¶

It becomes like the very top novels in the old web-novel world — it has its own authorship and creativity; it's not that pure-gratification, gratification-first kind.¶

There are novels too.¶

I already scroll past lots of them now — scripts and visuals both exquisite.¶

Really well-made. Wasn't there that Journey to the West one? Did it come out?¶

The Hunan TV one — the super… what's it, "Later Journey to the West."¶

Hunan TV.¶

"Later Journey to the West."¶

Made entirely with AI.¶

Broadcast on a satellite TV channel.¶

It's a completely new medium now.¶

And have you noticed: at first, AI short dramas put you off — the generation is too AI-flavored. But then you click into one, find the story is decent, and you overlook the AI flavor. That effect of turning text and scripts into visualized video should widen the consumer base again.¶

A new medium.¶

It still comes down to whether you tell a good story. Lots of people could tell stories all along; once the barrier drops, they can tell them well. This is a very new group, and they can earn incredible returns.¶

If your story is excellent and serialized — say you guarantee an episode every 2 days — many people will pay to watch. Not the old short-drama gratification-style payment, but paying the way people pay for the very top web novels.¶

Exactly — premiumization. So as we said, there may be a short-drama platform like Qidian.¶

It would have something like a monthly-vote awards mechanism, and pay-per-usage mechanisms. Hongguo has some of this now, but Hongguo is still too institution-facing. To list on Hongguo, its traffic-buying schemes are all very institutionalized.¶

The bigger market next should be the world beyond Hongguo, right?¶

But the global market is held back mainly by how strongly overseas audiences resist this.¶

Overseas, it's mainly on TikTok right now.¶

Nah — as long as it's this good, sooner or later they'll come around. You say overseas resists it — they'll come around fast.¶

Last topic I want to discuss: there's this claim that the AI application market will rebound in the second half of the year. Any thoughts on that?¶

Mainly—¶

It seems like in the first half, nobody was investing in AI applications anymore.¶

I think the core is…¶

Let me interview Orange — are you guys an AI application?¶

I'd say we count, sort of.¶

So how do you see this AI application market rebound?¶

The most fundamental reason is that Chinese investors are too starved for cash, so everyone can only focus on one track — in the first half it was world models and embodied intelligence. World models were really the same thing: too concentrated, everyone piling in frantically, believing only in this one thing. But the first half seems to be over already — the big money has all been spent.¶

But AI applications have indeed been a bit underwhelming.¶

Mainly it comes back to selling tokens — not enough tokens. Token margins, right — one problem is there are no tokens to sell.¶

There's barely any profit — some vendors' margins are just too low. Someone came to me saying they could offer the OpenAI API at 20% off. I said: anything with gross margin below 50% — the baseline for doing business is a 40% gross margin — look at that…¶

Fine for your own use, but selling it makes no sense.¶

GPT, AWS, Microsoft — they can offer 25%, 20% off, and there's simply no business to be built on that. Doing this thing is selling at a loss; you can't run that business. Why do I like Niulai and the Flash models so much? Because they give you margin room — that didn't exist before.¶

So I think there's still no margin, no tokens to sell.¶

Only when your product is cheap enough do you have a little room. Take Seedance — it had no discounts, none at all. Now there's some, right? But very little.¶

And yet it sold like crazy.¶

Right — because for the short-drama crowd, there's no profit at your level. Once they turn it into short dramas, they have profit. So if you go buy tokens, buy Seedance tokens, you can only resell at a loss. Like the Chen Mian approach: package it into some Plan, then place some bet or other.¶

Actually, I think — you could even talk cash flow: my cash flow is earned; you pre-sold a 1-year plan, then you stack all 12 months into this month, break even, and this thing just needs investors to believe it — that's how it goes. I think that was exactly the state of the first half.¶

I think the business model has to innovate. We've accepted the selling-tokens thing, but squeezing profit out of selling tokens is hard. And there's another problem, as we just said: no tokens to sell — GLM 5.3 Flash supply is actually insufficient.¶

Seedance 2.5 supply is insufficient too. The reason it can't give you a discount is that it doesn't have enough capacity, can't even meet demand — so it can't give you a discount.¶

But this isn't going to change for quite a long time.¶

So we need a new financial model, a new paradigm in the token economy, to break this open.¶

Actually, model development as a whole — I'd say the infrastructure part hasn't matured enough yet. Think about it now: it's like the years when bandwidth was supply-starved — saying you'd build Web 2.0 stuff on top of the network was laughable.¶

It's like the internet isn't even hooked up yet.¶

So we still have to wait. Or there isn't enough traffic.¶

But I think there's another thing: the first-half consensus was too high,¶

The consensus has all collapsed onto WorkBuddy and Codex, so as for the other AI applications, nobody may be paying attention anymore.¶

I think mainly many founders are just too slow. On Codex and WorkBuddy — frankly, you can't outrun OpenAI.¶

There's no way to beat it on that dimension — the office space, this thing.¶

I know this isn't just about models — models aside, on the product side, iteration is too slow. From spotting the early signal to shipping a product. Domestic startups — you guys count as very fast — but the pace of domestic startups and big companies is slow to an eerie degree.¶

But could it be that with AI, big companies have actually gotten faster?¶

Not faster. WorkBuddy is fast; the others aren't, I think.¶

WorkBuddy and QClaw — those came out of small teams.¶

Slow.¶

Codex is fast.¶

See — all of Meituan hasn't produced a single one.¶

Everyone else is slow. Go look at the domestic vendors' Agent products now — still missing things, the experience still poor.¶

As a big company, can you not use the best models? Can your programmers not use the best models? Or can your product managers not experience the best products? None of that. So why are you still slow?¶

I saw WorkBuddy launched a mobile version — after Dong installed it, he found there were no Skills.¶

Doubao supports Skills now. After selecting one, the moment you tap the input box, the Skill's selected state disappears — so I can't even tell I'm using a Skill. And once I install a Skill, it's installed — you can't delete it.¶

Doubao's Skills have another problem: after installing, you need to restart Doubao for it to take effect, before you can use the Skill.¶

That's at least 6 months behind.¶

Doubao is exactly 6 months behind right now.¶

No hot reload.¶

Qwen's scheduled tasks — no matter how much you pay, you can only have 2 scheduled tasks at the same time.¶

Doubao's scheduled tasks only exist in the Session that created them. This complaint is the most valuable one, because in a new Session you can't find your scheduled tasks — you have to dig through that original pile of a hundred Sessions.¶