Weekly #002 | iPhone Duo launches; does Astra using a computer count as AGI?

Hello to our friends online and offline. Welcome to episode two of Next Token | Beyond Tokens. First, let us hosts introduce ourselves—I'm Xiangyang Qiaomu.¶

Hi everyone, I'm Guizang.¶

Hi everyone, I'm Orange.¶

Hi everyone, I'm Pan YANG.¶

This is our second episode. First, thanks to Da Congming and Nan Qiao for inviting us to record at the AGI Bar here in Shanghai. Let us report last episode's numbers: we uploaded it to Xiaoyuzhou, Bilibili, and various other platforms. On Xiaoyuzhou we got over four hundred subscribers on launch day and over three thousand plays so far; on Bilibili we're at 10,000 views now. This is our second episode, and we'll keep the same order as last time: first models, then Agents, then AI-related hardware, and at the end we just chat freely.¶

So first let's talk about last week's model, the one many of you have probably used, GPT-6 Astra. Guizang, how about you kick off with your own take on your experience with it.¶

Why does Guizang always get to start?¶

OK, so last week with Astra I mostly played with 3D and modeling. On the first episode we also talked about this problem: model capability evaluations and how to make ordinary consumer users actually perceive it. It's getting awkward now—all those Benchmarks, and look, GPT-6 hits 99%, even 100, on several Benchmarks. So people are increasingly one-upping each other on 3D, grinding on 3D stuff, because it's the most intuitive: everyone can perceive that it's complex, and they can also perceive the result. Because before, you saw them drawing a pelican riding a bicycle, getting fancier and fancier—they're no longer satisfied with SVG and animations, they've moved on to 3D and all kinds of other scenes. Naturally—I don't know if it's intentional—it's this reinforcement of 3D. Actually, I think it is intentional. Everyone's demos are playing with Blender; as soon as it launched, right, it was genuinely impressive.¶

That's really interesting too, everyone.¶

I messed around with it, and I even published a little 3D game. I said I used 3% (of the Plan), and a lot of people didn't believe me. Under my Xiaohongshu post it was all doubt, intense doubting. But I checked—I did use 3%. So I took that Roguelike game and extended it into something like Vampire Survivors with an endless map—it keeps spawning monsters, you keep leveling up, choosing which upgrades and weapons to take. That added another 2%, so 5% of the $200 quota in total. Just now I was discussing this with Huang Yu, and I think a lot of the time—I've also had lots of people coming to ask me on WeChat, saying, why is it that I... I burned through my $200 quota within a single week, and the thing I made doesn't even meet my requirements. I burned through my $200 quota in one week. My conclusion is that a lot of people have no concept of how things work. One issue is they ask for too much—I saw they even use those prompt templates, "You are a whatever," right? Whereas we're totally casual now, we don't write any of that.¶

That's already last era's way of doing things.¶

Writing way too much. The other thing is you still need a bit of an engineering mindset. Say he wants to make a planet, and wants it to model everything on the planet. But the game is actually implemented in an engineered way: for example the starfield and such—outside there's a planet, and when you go in, there's a loading animation, and then it loads the current area—it has boundaries. It's effectively a complete planet plus five or six scenes inside that together make up the game's levels. He thought it was really a planet's worth of area being modeled, so there's no way AI could build that for him—and honestly, neither could we humans. He didn't understand how things work either. So it's these problems that cause it—excessive consumption and—¶

So he just didn't explain things clearly, right?¶

Right, the game is actually implemented through engineering. So a lot of times it's these problems that cause the excessive consumption and disappointing results. Pretty broke, huh. But seriously, GPT-6 really is very strong in this area.¶

There's also the infrastructure Blender has built up over the past few years. At first I assumed it was using Computer Use, but I asked around carefully and found it wasn't—it used bpy, Blender's own Python API. So what you find is: what does it model with? It builds things itself, it's writing code the whole time. But nothing moves in the Blender UI—it's just the welcome screen the whole time. After a while it tells you it's done building; you open it up—and there it is, already done. So it uses tools to do the modeling, and with MCP on top—that second layer gets used very rarely; it might just take a screenshot to check the result.¶

That's actually extremely token-efficient.¶

So why it's used so little—and why everyone's been using it recently: for the game engines I work with, Godot Engine has similar things too. So it's the same: when you're done, it opens up and it's there—there's no process of operating a UI.¶

See, this is what I talked about in my blog last time: going forward, software needs one set of infrastructure built for maximizing the efficiency of Agent calls, and then another set of interfaces for humans to look at interactions and content, combining it all together.¶

Look, last week DaVinci, the editing software, also released an MCP, exposing everything it has as callable, and it can hook up to any model. It's interesting—remember the paying issue we mentioned before? When they opened things up to AI, it's the paid version that gets the MCP, the free version doesn't. So it's the reverse.¶

MCP.¶

But I think the approach is right—it didn't bind itself inside one AI, like "pay money and then I'll let you use it, let you use some other AI."¶

So give us an overall assessment of GPT-6.¶

I—there's also my own Agent, which I keep iterating, and I used GPT-6 with it. I don't think there's a gap versus Fable—I didn't feel any obvious gap, or at this capability level the gap isn't big. Anyway it does the same work, and bugs are relatively few. It can catch the problems that used to slip through with 5.6. It's about the same. But on multimodal and 3D, I think it's a new, huge opportunity, a really big one—it might reshape the whole river of AI content production going forward. For example, like I said last week, the AI video crowd and the AI Vibe Coding crowd were two separate groups doing their own things, with zero overlap. But now I think it's a great bridge connecting Vibe Coding with AI video and 3D.¶

I think your Cola has integrated it too—any findings from a user-usage perspective?¶

I found that our users aren't actually that into the model, because it's fairly expensive.¶

They will, and—¶

Too expensive. Still too expensive.¶

About the same as Fable in usage. The model isn't exceptionally good at coding—well, it's quite good, but not remarkably better.¶

But not—¶

This morning I just saw a new Benchmark comparing Astra and Fable. The two places Astra is clearly stronger: one is 3D, specifically trained for it; the other is Computer Use, for now far ahead of all previous models—so you can tell it got special extra training. One more thing: the model seems to be the first—not the first, but for the first time in a long while, GPT has started sounding human again, talking like a person, speaking human language. Which means it's no longer grinding away specifically on coding as a specialty, and really is starting to expand into more areas. This makes me feel really good, because the coding area has been squeezed dry already; grinding on it further—¶

Pretty much squeezed dry.¶

Basically, look, the newly released DeepSeek Flash is also good. And the point Guizang just made is really important: when a model does something now, its method may be something we don't know about, but it works even better. Back when I was developing my ColaMD, a Markdown compiler, previous models would always spin up an Electron version to test it. But with DeepSeek Flash, what it spins up is different from the others—sometimes it doesn't spin anything up and doesn't show me anything, it just finishes testing on its own. I have no idea how it tested, but it was pretty magical. I think these methods are beyond me, but it just... gets my requirements done. Even for such a small model—pretty magical.¶

Our podcast production involves a lot of automation. So, Teacher Pan YANG, which of Astra's capabilities did you use to get all this done? Tell us.¶

Because Astra happened to launch on the very day we recorded this podcast. So the entire release of our first episode, including the website we built, I did entirely with Astra. One takeaway: although it's expensive, it's actually token-efficient, extremely efficient—which in a sense buys back my time. For a task of the same level: I used to default to 5.6 Sol's high setting, which would think for ages before finishing a task. Now I use Astra's medium, and it just does it, straight away. There's a small detail here too, everyone: if you use Codex, when you set which model to use there's a mode called Default—if you don't pick anything, you'll see the tiers underneath go like this: it goes from Sol's—Sol medium, then up to Sol high, then to Astra's medium, they're continuous.¶

How should we understand that?¶

Sol has two tiers, and one tier below Sol's two tiers is Astra. It places Sol, Astra—and before them Terra, and before that Luna—on the same tier track. That is to say, Astra's positioning should be that Astra medium sits at the same level as Sol high. There's no Sol high by default. Anyway, my feeling after using it: it talks way less, but works really fast, and the results are excellent. Also, I think Computer Use is a massive boost to productivity. Because think about it—we were always locked into the coding thing; anything beyond writing code, a human had to go click on something. That way you're stuck in Human in the Loop—all the time we waste is wasted on human operations. Including right now, recording live: I launched a recording app and I completely... don't know how to use it. My usage now is to tell Codex, "help me start the recording," and it goes and clicks the mouse and does it. So I totally agree with what Orange said just now: this generation, if it can operate everything inside the computer—think about it from another angle: the computer and the Codex Agent are one integrated thing. It has, how to put it, its own operating space, able to accomplish everything humans originally did on a computer. Then there's only one thing left: you tell it what you want to do,¶

and the matter gets fully solved. So I think this is—one step, this is a huge step forward.¶

So they say they feel that this one—¶

Yes. You've essentially given the Agent—it used to run inside the computer but couldn't use the environment. Starting from Astra, the entire computer belongs to it.¶

From the standpoint of operating a computer, it's already AGI. Its computer-operating skill is far beyond ours. From the operating-a-computer standpoint, it's already AGI, its—¶

I recently wrote some iOS apps before, all for my own use, never thought about putting them on the store because the process seemed too much of a hassle—so much material to fill in. This time, after the release, I handed everything over to Computer Use, and it submitted them one by one, all completed—got two iOS apps and two Obsidian plugins listed.¶

Including our blog, too.¶

Our podcast went up on all platforms. Because previously, personally—I was the one who published our first episode—I had never used any of these platforms: Xiaoyuzhou, Bilibili, Spotify, Apple Podcasts, YouTube. All of these platforms were published to by Codex itself. I never even studied the features of these platforms; I just submitted requirements, and my only job was handling the prepared materials. When we were chatting before the show, we said there's a point here. Now that Codex can do Computer Use, a new arbitrage opportunity exists. Think about it—our previous arbitrage opportunities were programming, writing articles, right? What's today's new arbitrage opportunity? Before, if one person had to maintain five podcast platforms, that took a lot of time; today one sentence gets it done. Think about what other things you used to find super annoying to do by hand—all kinds of things—including our website: originally we wanted to build really complex features and just couldn't. Today we can build as much depth and breadth as we want. Another new arbitrage space has opened up here.¶

Especially our website's Transcript, with our IDs attached—the effect. A lot of SaaS products maybe spent a year, some software spent a year on a VIP tier, and still don't have that feature. We finished it in 2 days.¶

Everyone really should go check out our website, it's called nexttoken.tv,¶

We finished it in 2 days.¶

It's really our experimental playground, I think.¶

An experimental playground, I think. Take a blog, right, a platform to a—¶

Take a podcast, right, a platform all the way—¶

I think this is the best platform for presenting a podcast's information, bar none. All kinds of tags—any keyword you see in the text can be clicked through to a tag page. None of all the other podcasts, no platform anywhere, has done this. Ten years of podcasting for nothing.¶

Xiaoyuzhou could acquire us.¶

Xiaoyuzhou could, Xiaoyuzhou—¶

could acquire us.¶

Including what Xiangyang said—now he can operate platforms like Volcano Engine and Alibaba Cloud, finally figured them out. Couldn't before; it used to be a maze.¶

Couldn't operate them at all.¶

So their external marketing also emphasizes 3D and Computer Use. But why is their Computer Use so good while the others feel quite far behind? What's the reason?¶

I think last episode we mentioned one point: in '24 they acquired a small team specializing in visual recognition on computers, UI-control operation stuff. I'd guess that team has been working on it these two years. Also, does everyone remember the news last week that when the new Mac launched, GPT—I mean OpenAI—bought how many machines?¶

Bought 20,000. 20,000 machines.¶

OpenAI bought a bunch of machines, bought—¶

20,000, I think.¶

Bought 20,000, bought 20,000 Apples. I'd guess it's related to training Agents.¶

20,000 machines.¶

I think it's a matter of route choice. Like Orange just said, OpenAI believes Computer Use is a must-have capability for the next generation of AGI.¶

After coding, Computer Use is also a general capability.¶

They couldn't wait for everyone to retrofit software for Agents—so they said, fine, we'll just make it able to use old software.¶

Can't wait. Too slow, I think.¶

But I think it goes both ways. As we mentioned earlier, lots of software is also adapting on its own. I think these two paths should proceed in parallel.¶

Look, you say you use JianYing—JianYing hasn't adapted well at all, which is why I went back to FFmpeg.¶

Right, me—for editing videos. Everyone recommended JianYing, and I absolutely could not use it. I subscribed to an annual membership on the spot—paid for the annual membership right away—and I used it for 1 minute,¶

Annual membership.¶

and I wanted to chisel my computer apart, it was so awful. I went back to—what did I use? Codex. Codex plus FFmpeg. I edited with Codex. Look at my video from last episode, the first episode—all of it was edited with Codex.¶

Right, I think JianYing might be facing possible disruption now. Before they were unwilling—though they clearly could adapt, they didn't make a CLI.¶

It should be the newest among these editing apps—its history is the shortest. Hard to say.¶

Hard to say, you—¶

But do you all think GPT-6 counts as AGI? Can you sense AGI?¶

Depends on the definition. To some extent—to some extent, right now there are things I can do that it can't, but it can do things I can't. It's stronger than me.¶

By that standard, AI arrived a long time ago.¶

It's one step further. I think it's far stronger than 90% of white-collar people, the ones who work online, who depend on a computer for their work. No matter—¶

Speaking from my personal experience. The only thing blocking it now is the old guard. I won't name names. There's one platform—like I said, I published the podcast to 5 platforms, all successful. Except one platform, when I published with Codex, throttled me. Throttled me. I think that's the old guard. Without those—any website or app that doesn't actively block things: basically, everything a person can do, it can do.¶

So I think the core of moderation is: if you want to push back on low-quality AI content, you should moderate the content,¶

not the actions.¶

not the publisher or the publishing method. Like, the four of us are real people who recorded a video, and you throttle me just because I published it automatically—I think that's a ridiculous logic.¶

Anything else about Astra people particularly want to discuss? How about we compare it with Fable 5.1?¶

Nobody uses 5.1.¶

Great—nobody even wants to compare anymore.¶

Right, no one.¶

Nobody wants to compare anymore.¶

No use for it—my Fable now runs inside Cursor.¶

But people can perceive that recently, on writing: one, Gemini 3.8 Flash is quite good. And Fable 5.1—maybe like Claude in the middle, there's a bunch of models that aren't great, but 5.1's writing is clearly very good, possibly their best model right now.¶

One thing worth watching: Benchmarks. Benchmarks are all over the place now, all kinds, and every model is surely chasing the leaderboards with targeted prep. There's one—I suddenly can't remember the name—there's a Benchmark that released 2.0, 3.0, 4.0. I forget what it's called. Also coding-related, or something.¶

What does it test?¶

DeepSWE, maybe.¶

No, another one. You look, and everyone's models score great on 2.0 and 3.0, but on 4.0 they drop hard. Only Anthropic's models have stable scores across 2.0, 3.0, and 4.0—no cliff-edge swings.¶

That shows the questions are testing real capability.¶

Real capability. No matter what questions you throw at it.¶

I think it's Terminal-Bench. Oh, right, right.¶

Their questions used to be rotten. Then the new 4.0 came out. Fable leads by a cliff.¶

Both Fable and Astra should be 5T models, the latest generation, way ahead.¶

On writing, what's the current consensus? We joked before—"which model is best at writing?" A bunch of people in our group said Opus 4.6, someone said—¶

Opus 4.6.¶

Anything better? 4.5. Until now it seems things have shifted a bit, right? Gemini 3.8 Flash—Xiao Hu said the experience is excellent, he switched all his website content writing over to it, so it should be pretty good. Plus the Fable 5.1 Orange mentioned. Funny that the models good at writing are basically all overseas models. Are there any domestic ones that write well?¶

Pretty interesting—domestically, K3 is a bit better.¶

I—I think K3—¶

Domestically, K3 is a bit better.¶

Not bad, K3, yeah.¶

K3 a bit better. Writers—everyone's grinding on coding, nobody cares about this.¶

Nobody cares anymore. It's bad. They've been grinding on coding for a year already; it's saturated.¶

Nobody cares about it.¶

GPT is bad at writing—this deserves a special mention too.¶

But you know, on novel writing, since GPT 5.5 it can write novels too. But among the domestic models, not a single one can write a novel.¶

Can't. And it's been getting worse.¶

I wonder if it's related to logic and depth of thinking.¶

No idea, but it's strange. Writing Chinese web fiction, Opus and GPT can both do it now; the Chinese models still can't.¶

But I still want to strongly recommend DeepSeek Flash, paired with the DeepSeek Harness.¶

So let's move on to the DeepSeek Harness.¶

If you want to search for information in the Chinese-speaking world, use DeepSeek Flash with the DeepSeek Harness. In my usage it's simply ceiling-level. Whatever else you use—including the overseas models, the overseas tools we mentioned—doesn't work. My feeling is it can find basically everything findable in the Chinese-speaking world, in both depth and breadth.¶

Including full texts of WeChat official account articles.¶

What cases made you form that judgment?¶

Lots of them. I look up all kinds of data, all kinds of—¶

Why does it have to be paired with their Harness?¶

I don't know. Anyway I think the combination is just great.¶

It has some plugins built in that can get what the Qwen model can't, because WeChat has some blocking and such.¶

Next let's talk DeepSeek—DeepSeek V4.1 Flash. My first impression: on Twitter lots of people were saying it's fast, really fast.¶

Three hundred, even four hundred (TPS) tokens per second of output. One person's take was: Astra is amazing—ordinary people use it to execute, maybe 1 hour to produce a 90-point product. But a professional programmer can use 4.1 Flash to get a result fast. Because I have professional skills, I can judge, I can adjust. Maybe the first one-shot output is 60-70 points, but through my constant coaching I can get it to—70 points, but through my constant coaching I can get it to 95, maybe in just half an hour. So the reviews seem pretty positive. For professional developers using it to build things—4.1 Flash.¶

I think a clear trend has emerged recently. I believe the few of us have always been believers in SOTA models. But with this recent Flash wave, everyone's putting out Flash models, it feels like model capability has clearly crossed a line. Below that line, everything is OK—and SOTA models are for solving a different class of more complex tasks. You can feel it just from everyone releasing Flash models.¶

You know Flash's score in the Arena? Its frontend capability beats Fable 5.1. DeepSeek Flash, which just came out this morning, is weaker than Astra but beats Fable 5.1. That's how good it is.¶

DeepSeek Flash.¶

Really?¶

After hearing you all, I specially downloaded the Harness yesterday and updated it, because I installed it when it first came out, but it updates so often—I found that many of the plugins I'd installed mostly don't work anymore. As you know, it's built for—still for training their models, for producing data. The combination of its model and the Harness is really awesome. I remember Guizang built something himself before—made a GUI for it. Your feelings? Evaluate the model and the combination.¶

The combination is indeed awesome, but the core problem is it iterates too fast—it releases a dozen-plus versions a day, so your plugins—¶

Can't keep up.¶

Their design is really great, but the problem is they keep touching those things—every update touches the plugins' core,¶

What happened?¶

the underlying core—so all the plugins break, and you have to keep adapting yours to it, the adaptation cost is very high,¶

The core Guizang mentioned, which causes—¶

Especially once you build a GUI: move one thing and your GUI is dead—move a place and it's dead. Or some core logic. Or some core—¶

So you can only build MCP or CLI.¶

So you see even the official side doesn't promote a GUI.¶

Can't be done.¶

Can't be done.¶

Right, can't be done.¶

It's fine. It's very early.¶

Like Brother Pan said before, it's a bit like early Linux, right? If you want stability, you wait for someone else to package a distro for you.¶

A distro.¶

Wait for the distro.¶

If you're a geek who likes to tinker, then you follow the official releases.¶

But you know, its device—the model is really awesome. The Flash model, even inside Codex,¶

What do you mean?¶

Its cache hit rate can hit 99%. We don't have any partnership with them, but it's 99% here, which is awesome. In ZCode it's also a 99% cache rate.¶

I specially ran it today with their Harness—it displays token rate and cache rate at the bottom, and I saw the cache rate was 99.5%.¶

Yes, it's a percen—¶

It happens often—before, it was often 100%.¶

Rounding up, it's often 100%.¶

Because with other companies' models, ninety-plus is already good, and some models only manage seventy or eighty. So its cost is really impressive. And this generation optimized the KV Cache technology, a 4x improvement, while the model's parameters only grew 1x in size, so the price effectively dropped 50%.¶

Damn, so much black tech.¶

4.1 Flash—I saw the activated parameters are only 8B now.¶

It can also dynamically adjust the system prompt without affecting the cache.¶

Damn, what is this thing made of?¶

I haven't read the paper yet.¶

They should have published a paper on the KV Cache too. Who was it—Ma Dongxi has already finished reading it.¶

Ma Dongxi. His metaphors are pretty good, his comparisons pretty good.¶

Has 4.1 Flash formed a genuinely new kill line, or is the kill line still the Niulai models?¶

Our community's feedback: the model is just too fast. We never used to think speed mattered; now 10x faster feels damn important.¶

Niulai is still a bit slow, speed-wise. But basically—do you know whether Niulai's discount period is over? If the discount is over,¶

No.¶

it should still be a bit pricier than DeepSeek.¶

Even with the discount it's pricier than it. Flash cut its price, by 50%.¶

Right, also V4 Flash added vision, and this time it should be native vision, not a bolted-on patch. It's the complete package now.¶

And the Flash model has some independent thinking ability—it doesn't flatter you. Yesterday I was chatting with it about some things and it ripped into me, said what I said was wrong, and I spent forever trying to correct it. I said this way of writing is better, and it said "I still stand by my argument, but I can help you write it."¶

Time to do some realignment.¶

The thing is awesome. Look, Google still hasn't mastered the technique.¶

Someone says once you're fast, consumption invisibly goes up. Someone says with heavy use, with DeepSeek V4.1 you could spend 200 RMB in a day. With DeepSeek V4 4.1, one day you could—¶

Well, productivity went up too.¶

But I think it could cause some waste. Maybe—maybe you haven't reacted yet; before you even stopped it, it was already done, having gone off in a completely different direction.¶

Yes, really too fast. The current bottleneck in development is: it needs to launch an app to test, needs Computer Use or other things—packaging is too slow. Every release, going through Apple signing.¶

Don't even mention Apple. Apple is stu—don't even mention it.¶

Is Apple Apple? One pass takes 50 minutes minimum.¶

These are all obstacles, you know?¶

These are all the old guard.¶

You know development takes 10 minutes, but releasing takes half an hour—I can't even be bothered to release.¶

Mine takes 1 hour to release, because my cross-platform build needs notarization for each one.¶

Like my iOS ColaMD dev build—I can't even be bothered to release it. It just sits in the repo; whoever wants to install it can install it themselves.¶

I'm not packaging anymore, do it yourself.¶

Also, their lead Cui Tianyi said they'd route the old V4 Pro to V4.1 Flash. What kind of consideration is this usually? And why did they walk it back?¶

Actually this isn't the first time they've done this—they've been doing it all of '25. But only DeepSeek, of all companies—everyone noticed the model names are fixed. Remember when there was deepseek-chat, and another called deepseek-reasoner, I think—just two, one corresponding to R1 and one to V3. They'd swap any model version with no notice at all—just swap it out behind the same name. The model name didn't change, but the behavior you got was different.¶

Stopped working right.¶

This wave, a lot of people online complained: "my prompts were all written and tuned, and after you swapped the model, the behavior is different." In '25 I actually pretty much sided with that view. But honestly, in '26 I don't fully agree anymore. Because I think overall, the share that prompts occupy in our products today is, I feel, shrinking and shrinking.¶

A lot of it has been internalized into the models.¶

Right, internalized into the models, that's point one. Point two: they later did listen to public opinion, right? Imagine—'25-era DeepSeek would never have listened to any user feedback. DeepSeek's attitude was: here's the product we provide, just use it as it comes. Our goal is AGI, right? We'll keep moving forward and let you use it along the way. But today DeepSeek is listening to users and the market again. One clear signal: they have to become a commercial company, one that provides commercial services to the general public and enterprises, and everything has to change bit by bit. It also shows—including the IPO, that valuation of tens of billions of dollars—once you've reached this step, you've become a public service, and anything small you do affects a much wider range, so they have to behave.¶

But I think the nature is different. This time it's forcing one model ID to route to another model, which is a completely different nature from "we updated it." Different in nature, but the result is similar. And it was very brief—in a few days 4.1 goes live, and after 4.1 is up it'll route back. So during that one week, are you just stuck on it?¶

Then let me tell you a detail. What happened last night: the DeepSeek model name—normally it should be DeepSeek V4 Flash, right? There's no V4 anymore; it's called DeepSeek Flash.¶

Once it expires—ships and it's out. No legacy—¶

Going forward you'll find DeepSeek will have two series—it will go down that road: one called DeepSeek Flash, one DeepSeek Pro, with version numbers dropped. I think it'll still do rolling upgrades going forward. Didn't we talk about this last week? Everyone's models have started rolling now—Qwen is rolling forward, Doubao Seed will roll forward too,¶

Everyone's started rolling—Qwen is rolling forward. Rolling is fine.¶

I reckon DeepSeek will go down this road too.¶

Rolling is totally normal, because most of their cost now is post-training; they're training all the time. But it shouldn't—because first, your two models have different pricing. Yours is a to-B service, with 10,000 companies using your model—how do they coordinate pricing with you? A huge hassle, and you change it for three or four days then change it back. So annoying.¶

If DeepSeek doesn't want to earn the to-B commercial money, it's fine to let other vendors earn it. Whoever wants stable version numbers and stable models can use the other third parties.¶

They can't earn it—third parties' latency is 100x the official one. And cache hit rate is 90% of the official one.¶

Thinking about it, for ease of use, the best choice for DeepSeek is the official one.¶

Their Infra level is probably more than a year ahead of Alibaba Cloud, Volcano Engine, all of them.¶

I think, flipping it around, it's also an opportunity for everyone: if any vendor can offer speed and experience close to DeepSeek first-party, the user traffic will naturally flow there.¶

Let's watch SiliconFlow.¶

So WorkBuddy's 0.03x DeepSeek. Don't know if it's official. WorkBuddy.¶

Those two are deeply partnered. When DeepSeek Flash launched, it specifically mentioned WorkBuddy, working with Tencent Cloud too. Tencent invested in them, I think—Tencent invested in them.¶

So I can use it for free or at a discount—practically free.¶

Practically free, 0.03x consumption.¶

Last week we talked about Grok Bot, right? Meta also released one called Muse. Feels like, as we said last time, back to basics—everything started from Chat Bots, and now it feels like we're back to Chat Bots.¶

It's back. Everything is back.¶

Brother Pan said he really wants to discuss this. Go ahead.¶

Last week I felt the topic wasn't fully explored, but after this week I feel the consensus is even stronger. It sits on the edge of turning from non-consensus into consensus—and what is it? Does everyone know what Grok Bot essentially is? The essence is: it assigns you a computer in the cloud. Many of its tasks get done on that computer, of course through your authorizations and so on, installing whatever software. Think back: when we first used ChatGPT it could only chat. Then Manus came out and Agents arrived—Agents could write code and run a piece of code on my local machine to complete a task. The next step brings us to today: it has 1 computer—it can not only run code, it can have its own environment, install its own software to complete tasks, even open websites or whatever. I think we're immediately heading to the next step, including the Meta Muse just mentioned. And actually our domestic Doubao Work also gives you 1 computer in the cloud. We just spent all this time talking about Astra's Computer Use, calling it that last layer of window paper before AGI. But think about it: if that's the window paper before AGI, then an Agent without its own cloud computer is a whole generation behind the ones that have one. On the flip side, it massively increases vendors' costs. Early on, I think, to grab users they'll give it away free, but later, Agents with a computer versus Agents without,¶

will also have a gap in pricing.¶

Look, Grok Bot has no free quota at all—zero. If you come, you pay, because the moment you arrive it spins up a VM for you, and a VM is cost.¶

But Muse is actually more personal, I think. My understanding is it should be usable for free. Its positioning differs from Grok Bot—I think it's more like an American Doubao, that vibe.¶

Buying it will—¶

Lots of questions. Several already.¶

"American Doubao"—there are already several.¶

I don't mean that kind of Doubao. But think about it—even for Google, whom we love to rag on, Gemini's DAU is actually quite high, several million DAU.¶

So must products with huge user bases inevitably head toward this Gemini-and-Doubao kind of situation?¶

It has to go in two directions. One is simplicity—Agents are actually complex. Look at Doubao's features when you open it: I think 50% of people can't understand what that line of text at the bottom is, things like connectors—incomprehensible. On phones—so it absolutely has to be usable. Even for our Cola, it's hard to grow on desktop but clearly easy to grow on mobile.¶

So I want to stress my position again: I recommend everyone try an Agent that has a cloud computer. People who've used it versus people who haven't are, again, a whole generation apart in their understanding here. And I'll be quite aggressive in experimenting with this direction myself going forward. Of course, if you're not willing to put some of your stuff on a cloud computer, there are other options. I think Cursor also released something benchmarked against Grok Bot—it's enterprise-facing, and it has a feature that lets you turn your own computer into the so-called host that the Agent controls. In short, the idea is: for what we're going to do next, you must give your Agent a complete, dedicated runtime environment, and hand it tasks under those conditions. What it can accomplish will differ hugely from what we used to get with a Claude Code, with a Codex.¶

Devin just raised a new round too, valuation exploding. They're probably going to move this way too.¶

Probably heading in that direction.¶

Is it because they have the cloud-based assistant? Were they the first?¶

They were first. They—they inspired it—after they came out, Manus was built on that.¶

Wasn't that why it sold so expensively before—¶

Wasn't the first one Manus?¶

Wasn't the first one Manus?¶

Manus. But Manus was also among the earliest to give users a cloud computer.¶

That call was pretty sharp.¶

So the industry has several people a year ahead in cognition—Mian Shen counts as one, right? Manus counts as one, right?¶

Also, OpenAI released an Agent API. I haven't tried it.¶

Leaning to-B, right?¶

Look, it originally had no SDK. Is this essentially their outward-facing way of calling Agents? Right now everyone uses the local ones—for instance its CLI, integrating it into my own software works great. What's the intended use of this thing, the Agent API?¶

Haven't looked at it closely.¶

As everyone knows, Claude Code actually opened one—Claude Code originally had something that should be called Claude Code Server, I think. Between GUI and CLI—and behind both CLI and GUI, what actually runs is Claude's Codex—oops, I misspoke—Codex Server. So it's essentially OpenAI packaging their official Agent Harness completely for you. Meaning any team or company can use this stack and have a ready-made Agent Harness—you just call the Agent Harness and close the loop on everything an Agent can do. Look, in the first half of the year, wasn't every company, every team building their own Agent Harness? Have you noticed that starting in the second half, things seem to be converging? One direction of convergence is Agent APIs like the one OpenAI released, and the DeepSeek Harness is another. Once the DeepSeek Harness stabilizes, there'll be a market consensus around it. Meaning quite a lot of companies and teams will build on top of it and do secondary development.¶

So let me ask Guizang—didn't you develop Copilot before, CodePilot? How did you build the underlying framework in the earliest version? My question is: for a lot of small companies and teams, if I want to build my own Agent, what should I choose for the underlying framework? For example—¶

Right now everyone—I think everyone is a shell. People only need to build the shell; the core, like the loop part, nobody writes it themselves, because you definitely won't do it better than they did, and the model adaptation stuff is really hard too. So my earlier choice was to build a shell adapting to, say, Claude Code, right? Back then the consensus was Claude Code. But later that part took on a lot of Agent arbitrage activity: whichever framework or channel is cheaper and more capable, you have to adapt to it. So look at what problem these shells have now: they've adapted to everything—Cursor's there, Grok's there, Claude Code's there. But here's the problem: between all these different adaptations you end up with tons of redundancy. Because a lot of things—you originally only needed to handle the Claude Code part; things like the Agent SDK already did a lot well, and you only handled the UI, really only the UI and nothing else. But later you find you need to build a runtime router, routing user input to different CLIs, different Agent frameworks. Then there's much more to build. And these Agent frameworks mostly support only their own API formats—Codex supports OpenAI's stack, Claude Code is Anthropic's stack, while Cursor supports everything,¶

but then you have to build it all again.¶

Cherry Studio—isn't that the work it did?¶

Lots of them did. Like the super popular T3 Code, same thing—that guy has a filthy mouth, always trash-talking, but his user base is huge. So T3 Code.¶

What's it called? T3 Code.¶

Similar thing. It'll stay hot, because the person is hot. So with this stuff, the more you do, the more work there is. More and more—and you can't adapt to just one, because say Claude Code falls off today and everyone mocks it with the "A-livestock" nicknames or whatever—maybe in a few days it comes back, gets its act together, its quotas go up again, and you have to use it again.¶

But my view is the opposite—also something I mentioned on the last episode: I think vendors' Agent frameworks or Agent products will be bound ever more tightly to their own models. Going forward, something like OpenCode, which is neutral, completely two-way neutral, will be its own independent little track. But the bigger trend is: you use whoever's model, you use whoever's Agent; you use whoever's Agent, you use whoever's model.¶

So that's exactly what I meant: precisely because the binding gets tighter and tighter, your adaptation cost keeps rising. Because clearly, if you want Codex or Claude Code to support all kinds of Coding Plans, that's the most disgusting part, because every one is different. And some of it isn't even deliberate—it's just that they don't care about this stuff, like distribution, the API, the returned data. So you have to painstakingly adapt to all of it. Plus what I said about switching between different Agents, and switching models. And your product has to have some unique features, right? It's not like I just switch to Codex and become whatever Codex is, with zero differentiation—impossible. You'll always have some historical debt sitting there, and you'll need to do additional adaptation. So I think this type of product will get harder and harder to build, more and more painful.¶

One more point: the Agent Harness. In the simplest days, I'd build a loop and it would run; now it's getting more and more complex, with features added in, and—¶

Pi is supposedly already considered very minimal, and it's added a lot too.¶

More and more complex—to the point where once something gets complex enough, it needs to be encapsulated,¶

Looks like V2 is coming.¶

Building it from scratch yourself—that stage has already passed.¶

Encapsulated. If you start building it now, you've already missed the stage. So what's the trend? What's the trend?¶

Saying OpenAI and Anthropic will be the future Alibaba Cloud and Volcano Cloud—providing a service you just plug into, and you no longer build your own. And the trend is also spinning up a computer—they've got the computer all set up for you, you don't have to do anything,¶

Agent Harness.¶

And you stop calling it as a model—you just call it as a service.¶

That's a brilliant take. Imagine how we used to build a website or an internet service: you rented a host from AWS or Alibaba Cloud and deployed your program on it. Going forward, it'll be more like—you rent an Agent host, or even rent an Agent, an Agent environment. And a huge amount of business may end up being driven by this. The number of website and server hosts might become the small share.¶

It's because over the past year everyone treated Agents as products, but an Agent isn't a product—it's a technology. Later on, for a product—look at some of the overseas personal Agents that are hot lately: they don't even tell you which model they use anymore. To users it's all the same. So they just say: I provide my special service; the Agent stack behind it, you don't need to care—you don't even know it's an Agent, and you don't need to know. That's true productization.¶

A couple of days ago I was chatting with an Agent—an AI app developer—and together we sighed: what did everyone actually do in the first half of the year? Everyone was building Agent Harnesses, and today we find it was all useless, all wasted effort.¶

It was all learning, right? It was a learning—¶

It was a process. It was a process.¶

I remember once DeepSeek wanted to build a Harness and invited people to beta test, asking everyone to comment on what similar Harnesses or frameworks they'd built.¶

A few hundred people, right?¶

Wow, several hundred replies. Guizang replied too.¶

Everyone had built one. Today they all became test subjects for the model companies.¶

Everyone had built one, that Agent-in-a-shell kind of thing.¶

But it was everyone moving forward together as a process. All those people who built Harnesses—they can join the DeepSeek Harness. Aren't they hiring 150 people? 2 to 10 years of experience—hurry up and apply.¶

About the model-and-Agent-as-one thing—after we finished recording last episode, I went to Qoder to talk with them, and I gave them feedback as a user. I said the Qoder product might be a good product, but because I think the Qwen model isn't good, I don't use Qoder. But if some version of Qwen suddenly became really good—which Agent would I use to experience Qwen? I might install Qoder to try the Qwen model. So it's all mutually reinforcing. But really, for the vendors: if the model isn't good, the Agent can hardly carry it; if the model is good, carrying the Agent is relatively easy. Of course that's my personal view—I just like the original soup with the original dumplings.¶

Because everyone's Agents are pretty much the same in capability, so when you release a model you'll definitely subsidize—and subsidizing your own Agent is the right place to do it, because you can do a lot of work there, including the caching thing with the DeepSeek Harness we mentioned, and their own caching. So that's really how it is now.¶

Look at these features—connecting email, calendar, shopping and payments, health. Everyone can build these.¶

Which one doesn't have them.¶

You mean Muse, right?¶

Muse, just—¶

No.¶

Doesn't Muse just copy WorkBuddy?¶

No, WorkBuddy doesn't have them. No, not quite the same—¶

WorkBuddy can do all of it.¶

I think one of the better ones recently is A-Fu—it's done healthcare really well. Recently when my colleagues and I run into minor ailments, we all go ask A-Fu.¶

Why? Is it because it has the professional knowledge plus it's very—¶

Its professional knowledge really is good. Say you ask some other general Agent: "I've got a bump on my hand, what do I do?" It gives you some general answers. A-Fu is much more detailed, and even its medication recommendations are good.¶

So this shows what building AI apps is about: data, capabilities. Like we said last episode—WorkBuddy makes stock trading convenient,¶

It's going to come back, back to Shanghai to have fun.¶

Things that stock-trading software itself could never implement, WorkBuddy can, because vendors open up interfaces, open up capabilities, open up data. Those are the moats, the points of value.¶

Qwen has a problem, though. Qwen—for example if you ask it to shop on Taobao for you, it can't manage it.¶

That doesn't make sense.¶

Qwen can't manage it, but inside Taobao AI, shopping on Taobao works way better than it. Fliggy's AI handles travel and such really well, but ask Qwen to do it and it's not as good. Pretty strange right now.¶

That's department walls, isn't it.¶

No idea, huh.¶

No idea, but the facts are what they are. So right now it looks like A-Fu and Fliggy AI are nearly usable, while a general Agent like Qwen can still basically only do office work. I find it strange, and I really doubt whether a truly general Agent exists. Like Manus claims to be a general Agent—if you ask Manus to make a short drama, can it?¶

Could it possibly cover that, right?¶

Can't. There's really no general Agent.¶

If you put it that way, WeChat's Xiaowei actually has the biggest opportunity ahead, because—WeChat's Agent, it's not that—every mini program is one of its capability interfaces, every one is its MCP, its capability interface. Meaning, say Ctrip has a mini program, right?¶

If it can hook into Xiaowei, then I just place the order on Ctrip directly through Xiaowei.¶

Then Xiaowei places my order online directly.¶

See, yesterday we all tried Xiaowei. AJ, for instance, had Xiaowei find some group and send 1 message notification, and felt that in the future we'll all use Xiaowei to handle all kinds of things. When you invoke all kinds of capabilities, your binding with AI, with WeChat, will definitely get tighter and tighter. WeChat will surely move in this direction.¶

So look—software has basically lost its value, and an Agent is itself a technology, which also has no value. What has value, in the end, is the service. Only service has value forever.¶

Service—data. Interfaces and data are the service you're talking about.¶

The service you mean, right? Service. It all comes down to service.¶

And speaking of Xiaowei—Orange, remember we talked about this back in '24? I told you: an Agent isn't a chatbot sitting across from us; an Agent is beside us. Like Xiaowei—it really is beside us, like Xiaowei: she's embedded in the software. She doesn't have to go through a conversation where you tell her something, ask her to do something, and she goes and does it. Based on your context, based on the capabilities inside your software, she operates the software for you.¶

And appearing in the contact list in the UI matters a lot. Like Wang Dengke and their The One—it blew up immediately. I asked them, "Have you run into any problems?" He said the problem was that the OpenClaw QR code WeChat provides is unstable—about 10% of the time messages don't arrive.¶

Unstable.¶

I said why don't you build an App or a website? He said users just love the contact list—they can't leave the list.¶

I see—adding a friend inside WeChat, the Agent goes in there—¶

Which product of whose did you just say? Wang Dengke's what?¶

The One. It's not even an Agent, it's just a chat.¶

The One.¶

But just because you can scan the code and chat right inside WeChat, that improved conversion by 6x. Going from exposure to download, only about ten-odd percent of people convert, but scan-and-chat is a straight 100%.¶

I think WeChat is just too strong. Everyone's relationships and all kinds of data are here. For ordinary people, you don't need a particularly complex Agent to solve problems—a chatbot can handle that kind of thing.¶

So really, WeChat doesn't need a WorkBuddy. WorkBuddy, I think—along with Doubao Work and Qwen—are all intermediate forms. They exist to get these Computer Use data, to get data for training models. They won't make much money either.¶

I think as long as WeChat provides even a barely-usable feature, people are too lazy to use anything else. Like today, when I was fiddling with the Pocket lens, I didn't know how to do the lock. Normally I should go find some DeepSeek or ChatGPT, right? Instead I just opened WeChat's Xiaowei, because nowadays when you open WeChat you just tap the search box and the Xiaowei button pops up. For a really basic question like that I just asked Xiaowei, "How do I lock the lens on the Pocket 4?"—and it gave me an answer, and it was good enough.¶

So going back to the question—a lot of people actually don't work, have no work, don't need it—really, no use.¶

No need to keep tinkering.¶

And the number of people who need to work will keep shrinking. Don't need—¶

Also, Xiaohongshu's Diandian is really good to use.¶

I use it a lot too—I often look up game guides with it.¶

I think it's Xiaohongshu's best AI feature.¶

Actually Xiaohongshu's Diandian is just WeChat's Xiaowei—basically the same thing.¶

Right, the same. Same here.¶

But the data sources differ.¶

But Xiaohongshu's data quality is a bit better. But WeChat has Channels—same thing, the counterpart. Quality is a bit worse, but the volume is huge.¶

Mainly image-and-text can be—well, a lot of Channels content doesn't transfer well, hard to convert.¶

Including Xiaohongshu—say you want to buy a certain car, buy certain shoes, it'll summarize: 30% of the posts recommend this model. Wow—¶

You're traveling somewhere, you still check Xiaohongshu? It'll summarize it for you.¶

So accurate. You don't even need to read the posts anymore, image or text. It's killing its own feed.¶

The ratios are really valuable.¶

So for everyday life questions, go ask Xiaohongshu's Diandian.¶

The key platforms are super sensitive. The moment there's any automated action they ban you. I feel like we've just about covered Agents—actually quite important. iPhone released that iPhone Duo.¶

We're only getting to iPhone now?¶

Right. I used to stay up late to watch keynotes; this time I somehow didn't pay much attention.¶

I watched this one. I don't usually stay up, but this time I made a point of watching.¶

I just watch the 24-minute version the next day. The Verge only posts an edited cut. Efficient.¶

The foldable. Didn't you publish a podcast yesterday—Liu Fei walking through the whole history of foldables? Saying foldables existed seven years ago, and now that Apple has released one, it's as if they reinvented the foldable—this kind of sentiment. And lots of people talking about it; it really does make it easy to push sales. Especially when he unfolds it, that transparency—the visuals went viral on X, people feeling like, wow, this is—¶

Tens of millions of views, probably.¶

So what do you think of this?¶

Apple is an interaction company, that's all.¶

We call it UX Porn.¶

You know how every iOS update feels like fashion-level filigree, right? Each update adds some glass effect—more flour into the water, more water into the flour; too much glass, dial the blur back up; too much blur, tune the glass back up. But once they actually do something new, they wake up. The interaction refinements and the new adaptive designs on the iPhone Duo are pure genius.¶

How so?¶

Take the folding animation. Before—domestic phones actually did a lot before iPhone, whether narrow folds, wide ones, or tri-folds, they did a lot. But in adaptation, going from outer screen to inner screen when you unfold—¶

Nobody considered those things.¶

Nobody. It was always a stiff little rotation with a white edge, sometimes even a flash. Nobody cared about this area, or about adaptation. As for adaptation—what do people adapt? Domestic ones, lots of them—they just dump it wholesale on the vendors: "redo your UI to my proportions," or "let's partner—through some benefit exchange, through some means, get you to adapt the common apps for me with human labor." Still a very brute-force kind of adaptation. Their own systems actually do very little of it either, but that's mainly a problem of controlling the system. But look at iPhone adaptation—sure, the adaptation difficulty is high, and there's more to adapt, but the adaptation cost is visibly lower, I think, because they considered the whole interface approach. It's genius design: they put both the top bar and the bottom bar on the right,¶

Different folding states, right?¶

Because when you unfold, you're just adding a strip—a strip this long. Once you move them over, the content ratio on the left stays the same as before; you just show more vertically. That's how they free up the content, and at the same time your adaptation isn't much of a hassle. If you used the official components, you just move them over and that's it. Of course, if you customized too much to achieve flashy effects, too much of your own stuff, then there's no way around it—go handle it yourself. But most apps used iPhone's official components, and there are many devices, so once you move it over it basically just works, and you fine-tune a few details. Including that—¶

The three icons merged into one.¶

The Wi-Fi, battery, and signal animation went viral too.¶

The animation also passed 10 million views.¶

If you just look at the icon, you're lost—what is this? A Wi-Fi symbol inside a ring, with four dots below it? But the moment you see the animation, you instantly get it. You don't need a second look—one look and you understand what those three icons actually mean: the battery becomes a line wrapping around the Wi-Fi, and the signal below becomes four dots sitting beneath the Wi-Fi.¶

When the battery, Wi-Fi, and signal suddenly merge into one, my brain just collapsed. I watched it ten times, you know? It's incredible. Design—why has no one designed things like this for all these years? Someone said the iPhone, the iPhone 18 should merge them too. So elegant.¶

And its rounded corners line up exactly with the outer corners—its top-right corner, it matches.¶

The word you used earlier was spot on: interaction-design porn.¶

It really is something—you just feel great watching it.¶

And I think the iPhone Duo defined three things this time. First, it defined the foldable, because it's so impressive—foldables before. No wait, hear me out—before, foldables, like other vendors': none of them stuck with one model. They kept changing—some vertical folds, some tri-folds, some bi-folds, some narrower—always searching for a design. 1 over root 2. But Apple isn't like that: after making a machine, Apple iterates on it for several generations—meaning for the next few years, it's the ID. See, once it defines something, Xiaomi, OPPO, Huawei—everyone, everyone learns from it.¶

Now they know how to do it, right?¶

Because it defines the standard—like the MagSafe thing: everyone adapts to Apple, including Xiaomi power banks, right? Everyone fits the iPhone better. Isn't Xiaomi even harsher about it? Because it sets a standard, a hardware standard, and interaction too—the other vendors can't figure out how to do interaction. Apple's interaction—yesterday Master Cang's Xiaomi Duo, right? The interaction is explosive. Now everyone can finally copy it.¶

Now everyone can finally—¶

On interaction, I think this is Apple's biggest update in four or five years. The interaction is genuinely awesome, more than a year ahead of domestic phones—even copying will take a year. It defined the interaction, and it defined the third thing: your ecosystem. Why was adapting for foldables so hard for developers before? Because every one was different. Now they're all the same. It defined six places to adapt, six interfaces—including folded shut, opened up, half-folded. Adapt those six and, for the next five to ten years—I think at least five—you only need to adapt to those six screens. Interaction design aligns to this. And once you've adapted, Android gets it for free too; everyone will support this standard. Developers can finally do some foldable work.¶

So great companies define standards—that's what you mean, right?¶

The main thing is you can also follow the standard.¶

My goodness, Orange is such a hype man.¶

It's not that I personally wasn't into foldables—I just thought the things were heavy. But—because that day I listened to He Tongxue's 15-year retrospective on Cook. Actually, in the year Cook took over, they shipped Siri.¶

An introduction to an entire era.¶

In 2026 Siri is still an idiot. In between: Apple Car, 10 years and 2,000 people, never shipped. Apple Vision failed. So shipping an innovative product at Apple is not easy. Everyone thinks a big company, right, has money and taste. The foldable—I think it might genuinely be Apple's only successful new product in all these years—¶

One with hope of succeeding, right?¶

Really, it is.¶

I originally thought the iPhone Air was nice—big screen, feels amazing. I was thinking of waiting for the second gen. But now that the iPhone is out, I guess it probably won't come.¶

You mean the one that's super thin?¶

And there's a theory that the iPhone Air was made so thin partly to do some early technical exploration for the Duo.¶

Yes, a lot of people said that last year. It's true.¶

Let me give my take. I think there's something fundamental. Think about it: from the iPhone's launch until now, screens have kept getting bigger. Everyone's daily phone usage keeps getting longer. We're inevitably heading in this direction: our pursuit of how much content can be expressed on a given screen area is endless—the more the better. And it can no longer satisfy today's content consumption needs. So going forward, mainstream phones will keep making screens bigger. But no matter how big the screen, you can't have every person carrying a pad every day. So folding, I think, is a trend that's bound to arrive. Building on Orange's hype just now: once unfolded, we have more content to interact with and consume, so we'll definitely go down this road. And the next-five-years new trend Orange mentioned.¶

But on the AI side Apple really seems to have made no progress.¶

Didn't mention it this time either.¶

Don't get me started. Someone said—I asked it to record some workout note, and Siri said "Want me to write it down?" He said yes, write it down. Siri said "OK, I've created 1 note for you."¶

No memory whatsoever.¶

It's still three years ago.¶

Feels like it's still stuck before 2023.¶

Ugh, think about Siri—genuinely 15 years of iteration, from Cook taking over to stepping down, and it's been this bad the whole time. Absurd.¶

Part of it is model-related, I think. But also, Apple is a company extremely focused on privacy, and all its design—AI and privacy are deeply entangled. It has to think about local secure storage, encrypted cloud storage. Before it has the design figured out, it—¶

But it is, after all, still a consumer electronics company, not an AI model company.¶

So it won't jump in lightly.¶

Hasn't he already bought the Xiaomi Duo? Go ahead, share your thoughts.¶

My take: Android really is hard. The adaptation problem—adaptation is brutal. Because each of our apps is already really heavy now. Take those domestic super-apps—how many screens do they have? I don't think anyone knows; they themselves don't know. A lot of them are super-core interactions. When you adapt, once it becomes very wide—those six things we mentioned—it's terrifying, the adaptation volume. And with Android adaptation, let's be honest: screen ratios differ, widths and heights differ, displays differ, plus each system's own quirks. And domestic vendors love to fork and mod Android—every single one mods it, no matter what Android you are, they all mod. So no vendor can define—can claim the standard-setting we just talked about. No vendor can define it, so I can only follow Apple.¶

The root-2 thing—define the standard, and from then on everything executes to the standard.¶

Root 2. It has to be root 2. But don't you think—because of what Guizang said, so much to adapt—people are getting sloppier and sloppier about software these days, have you noticed? Nobody carefully polishes a refined product anymore, because the pace of society keeps speeding up. We used to spend a year or two refining an app's interactions; now it's like, whatever about adaptation, the vendors don't care anymore, nobody cares.¶

Now only one company still does this: Mideng Clock.¶

I often use domestic apps and often can't tap things—can't complete an order.¶

For me it's some domestic banking apps—can't go back, return—¶

Not just domestic banks—those top super-apps too. I can't place an order—something this important—can't pay. Doesn't work. Often a whole week without a fix.¶

So, done with iPhone—didn't this keynote also have a watch AI feature? I think it's—¶

Right, the watch AI is genuinely quite good.¶

I think a lot of companies making voice recorders and recording cards must be trembling.¶

It's a summary—only a summary, no—original text.¶

It can't do verbatim.¶

Privacy protection.¶

But look, before, lots of geeks developed recording apps for the watch—many people did. And there was a voice pushing back: "is this thing's recorded sound worse than a recording card?" Lots of people criticized that. But today with Apple doing it officially, I think all these problems will get solved. I don't know if the hardware changed too—maybe more microphones or processing. But there's another view: now that we have AI, things like recorded voice, noise, whatever—none of that is a problem anymore. None of it. AI can solve all of it.¶

It processes on-device, transcribes locally. And it has a really nice feature: if there's a sentence you didn't catch,¶

It must be tethered to the iPhone, tethered to the iPhone.¶

you press it and can replay 15 seconds, and all the text from those 15 seconds gets written out. Say you're at a restaurant ordering, the waiter says 15 seconds' worth of stuff you don't understand—one press and the text appears, and later there may be translation and such.¶

I've always felt that recording will gradually become society's default norm. Right now I still have to ask "may I record?", or people act uncomfortable when I record.¶

May I record?¶

It will slowly but surely move forward. From another angle: look at people now—almost everyone has ADHD, because there's too much information to process. Human memory—how to put it—certain memory abilities will degrade. More and more we'll depend on these electronic devices to do things for us. My Apple Watch hasn't been replaced for many generations; this time I'm thinking I might upgrade and try one because of this.¶

Health matters a lot. Apple already takes so much of your personal data—for instance this time they raised the heart-rate collection speed, the sampling frequency went up 60x.¶

5 seconds, every 5 seconds.¶

Right, 60x.¶

Sampling all day long. Very—they should be—I mean compared with what hospitals use, the precision is about the same now.¶

The official site even ran ads with cycling and swimming and such.¶

I remember someone—Xiang Shu or somebody—showed me a watch without a screen. He said the subscription fee is a year of—Google, I think it was that—¶

Google one.¶

Garmin released one, I think—¶

Not that one, not Garmin. He said its AI analysis results were really good—he just followed the results, and his whole deep-sleep time improved substantially. I think this health data collected on watches still hasn't been properly mined.¶

Oura ring too—renewal rates are really high.¶

Right now I only run one scheduled task: ChatGPT Work—health. Because every time I do strength training there's a text log. My coach logs it—he's the kind of—I've mentioned this before. My coach logs it—he's the kind of—I've told this story before: under my influence he uses it quite a lot, and he's gone full Agent now. I just pull my strength-training results from there; my daily cardio results are all on the Apple Watch, including how much the strength training burned—it can see that too. Every week—sleep matters a lot—every week he gives a report on the few metrics you care about, watching whether those key indicators have changed. For example, I used to think my headaches came from rhinitis, a nose problem. Turned out it wasn't rhinitis—it was my cervical spine: sitting too long gives me headaches, not a nose with rhinitis. So it found lots of problems like that. Overall your personal state improves, because a doctor will never go look at that much of your data,¶

That granular?¶

to go through that much for you. And doctors—many doctors aren't generalists; they have to look things up too. On these lifestyle matters—and frankly, you're not sick, you're just a bit uncomfortable, so you're not worth a doctor's time and they won't deal with that part of you. But every change [it makes] is significant. Given what you said earlier about A-Fu, I think these lifestyle Agents and context will matter more and more. Work actually isn't the most important—¶

We have a user who had an illness for ten years, saw all kinds of doctors, and none of them figured it out.¶

We have—there's an adapta—¶

He told Cola all the symptoms, all ten years of them, every episode he'd ever had, and Cola reasoned out that some medication might solve it. He took the suggestion to his doctor and asked, and it turned out—sure enough. The doctor said, give it a try—and it cured him. His illness was affecting his weight—he'd probably always stayed under 120 jin (60 kg), and after taking the medication he gained a dozen-plus jin and felt great.¶

Speaking of which, I have a recent example too. A month or two ago I got BPPV—otolith syndrome—from playing tennis. I don't know who's had it. Dizziness. I described my symptoms to Codex. The first time I described them, it made a judgment, and I also raised some directions I suspected. A few days later I had some new symptoms, gave it to him a second time, and he pointed me in a direction. What remained for me was, following the direction he pointed, to register at a hospital—because the final physical-world tests have to be done at a hospital. After the hospital tests, the doctor's diagnosis was exactly the same direction Codex gave. Including the later treatment strategy and everything—it all matched. For these things going forward, it's a real help to everyone.¶

AI plus health.¶

But what Guizang described as the closed loop is really personal data—still personal context.¶

context.¶

We're gradually expanding all kinds of our own context—the more there is, the more AI can help us.¶

So a lot of people now use the Feishu Anker recorder bean, recording as much as possible. A friend said his recorder bean's battery is about worn out—he's been using it over half a year.¶

AJ is right—AJ's, the recorder bean's, battery is almost dead. Too many home disk backups wore that out too.¶

If you use the recorder bean every day, it lasts about half a year. But you can go on Xianyu and buy a brand-new one, and it comes with a 6-month Feishu membership—exactly half a year.¶

So you replace one every six months?¶

No, it's just 500-something yuan. Pretty worth it.¶

So context, I think, is still core. That's why—remember we actually mentioned this last episode—why we think Grok Bot is OK: because overseas everything is opened up, all the overseas business models and MCPs are open. Why can we only use Diandian inside Xiaohongshu and Xiaowei inside WeChat? While it—¶

No one shares data with each other, no APIs.¶

Walled gardens all around.¶

We're about done with hardware. We have a few quick updates; I'll go through them briefly—see which ones interest you: Doubao Input Method released a Windows version. TRAE updated with Agent merging and enterprise capabilities. And two overseas model companies: Suno V6 shipped. ElevenLabs also did some partnerships with Universal Music.¶

TRAE is gone already?¶

No, TRAE and Work are two different things.¶

But their fates might be the same.¶

No—the product lines are already quite clearly divided now. The coding-oriented ones, right? I think it's already pretty clear: for coding you have Qoder on one side, TRAE over here; for office work there's Doubao Work over there, Qwen Office here. It's very clear now. Over there WorkBuddy, CodeBuddy—everyone's whole lineup is aligned.¶

So future updates get simpler too, right?¶

I heard a podcast yesterday saying the internal horse-racing at the big tech companies has reached a stage wrap-up, and they're starting to fight external competitors. The battlefield ahead is still huge.¶

And the coding track is done being squeezed, right? I don't think there's any growth left. Coding is IDEs—Qoder already switched, right?¶

And domestically everyone's pretty unified. Domestically nobody does coding—the ceiling is too low.¶

Already squeezed dry. 10 million programmers—gone.¶

10 million programmers, 100 million office workers, right?¶

About the music, I tried it. I didn't feel much change in audio quality.¶

V6 still has problems—on audio quality there's still that—¶

But isn't what it's solving the editing problem?¶

What it solves is: they added an Agent that you can interact with. What is it?¶

What is it?¶

Normally when we use AI to generate music, you might have it randomly generate a song, or give it lyrics to generate a song, and after generating, once you change something, you have to edit it yourself, or go edit in their Studio.¶

After this version upgrade, they added an Agent, and inside their Studio you can edit with natural language.¶

So the music industry is another year away from Mian Shen. Didn't video and images get Agent-based editing a year ago?¶

Nobody competes with them in music, so they can take it slow. They're already SOTA, right? Like Midjourney.¶

They're in no hurry.¶

And you need the top-tier membership to get this feature, the Agent.¶

They're in no hurry.¶

But since you mention it, GPT Image 2.5 is also—its version is really an editing-capability upgrade too.¶

Midjourney 2.5 too.¶

People around us who tried it said they couldn't see any difference.¶

It's like this, it's—¶

Honestly not much difference.¶

GPT Image 2.5 can achieve absolute stability: whatever you don't touch doesn't move at all, not a single bit. Before, you'd take a reference image, generate 1 person, generate ten-plus in a row, and you'd notice the person drifts slightly each time.¶

Slight drift—from this version on.¶

From this version on, say you edit eleven versions in a row—everything else doesn't move by a single pixel.¶

But it's the same old problem: the person doesn't move, but on everything else the artifacts and brokenness get worse and worse, more and more yellow. Old problem, unfixed. So still unusable.¶

2.5 was actually trained together with Astra. One of Astra's abilities: you give it a photo and say "repaint this for me, hand-paint it"—its hand-painting is strong. Lots of people posted,¶

Repaint it, very strong.¶

I just used my Codex and had it paint, and it didn't even use a drawing tool—it painted a picture with code. I was shocked, you know—code can even paint. 2.5 is the same—it's the click-here-change-that thing, trained together, with a very precise understanding of pixels. But it brings problems—on the, and also its 3D got stronger: ask it for some 3D images.¶

The aesthetics are still lacking. On aesthetics, yeah.¶

Because it's all 3D now—3D image generation is better.¶

It has a really fun demo: have it draw a cube, and each edit moves the cube by some angle. Before, you'd see the cube wobbling around—because when you move along one axis, in principle it shouldn't wobble,¶

Mainly the angle.¶

Because when you move along one axis, in principle it shouldn't wobble. The new one is normal—it stays continuous,¶

But it can't wobble.¶

The new one behaves. Say you edit 60 times in a row—it looks like a video, nothing moving.¶

It's gained the ability to predict the next frame—it could do video now, right?¶

So doesn't that mean it can make videos?¶

It could always make video—it just doesn't want to.¶

It can't. But with its earlier stuff, it wasn't precise—for example predicting your action's next frame wasn't precise, so your animation looked stuttery. But after this version it's very precise: just have it make one image containing, say, 20 frames, listing out every frame. Then use Codex to turn it into a video.¶

Another thing—didn't I make a Skill last week, an image Skill, for posters—turning travel photos into posters. Then I wanted to hook all these things together—¶

to mount them onto those design Agents. Those design Agents—everyone knows the ones. I found that on both sides there's what we always call user isolation, and on these Agents—the vertical Agents—the isolation persists. A lot of Agents—on design Agents it can't code. It has nothing at all. The model it hooks up to has coding ability, but on its Agent there's no coding ability at all.¶

So they don't let it code and handle it themselves instead.¶

Can't code. For example, I actually use a lot of tricks to guarantee the output quality. When doing layout, right, I first write an HTML that only handles layout, the text grid.¶

What you're saying is some Agents don't give you a runtime.¶

They restrict it—they actively restrict coding ability. The problem is your—¶

WorkBuddy restricts it. You just can't get away from it, can't get away.¶

Safety? None of it is for safety—it's that they didn't build it. This is the most basic thing: as an Agent loop—¶

None of that. They just didn't give it an environment.¶

Whether you're online or local, coding is the most basic ability—and they don't let you code, they don't let the model code.¶

So look at the market: several different generations of Agents exist at the same time. The most advanced we mentioned earlier, Grok Bot, right—it gives you a whole machine. And over here there are ones that let you run a runtime.¶

I have a bunch of reference images in there.¶

But aren't we moving too fast, for a lot of enterprises?¶

But saying we've advan—¶

prompt plus AI—they think this oldest way, not-even-an-Agent way, solves most problems.¶

To-B, or some traditional enterprises—I think, going back to the question, they can be forgiven or whatever; their pace is normal. But as an AI company, I think it's unforgivable. Including my prompt has a lot of image prompts: the final product uses four images; one prompt might be one to two thousand characters. The four images come from, respectively: the original image, the layout image generated with the frontend, a screenshot taken after the frontend generates it, plus a reference image—and there are three images. Three images, each from a different source. For layout reference—reference images—you need a filesystem. I uploaded and found images could be uploaded, uploaded them; Skills could be uploaded too. I uploaded, the images went up, the Skills went up, packaged and uploaded. But it can't call the images. So—I asked it: "Look, can you see the reference images in the folder?" It can see them. I said, call them. "I can't."¶

You know why? Because Skills were really made for Claude Code and Codex.¶

I think these are the most basic capabilities—for an Agent, whether vertical or not, you need these, because they complement each other. Including what we said about GPT-6 doing 3D modeling and Blender stuff—it models by writing code. These are things you should have, because we in design need them too. But they're just not there. Including VMs—those VMs we just mentioned. Look, Doubao Work has a VM too, but theirs is ephemeral—Manus before was the same.¶

No—long-term memory.¶

Manus, really?¶

Ephemeral, right? Right, ephemeral.¶

Every time you're not the same—right, ephemeral. Every time it spins up a different one, and different processes, different conversations spin up different VMs. So your context doesn't persist.¶

No persistence.¶

Kimi was like this before too.¶

They're all like this—domestically, all like this. So the cost is low, but low cost means the thing is useless.¶

So this isn't the same generation of product.¶

Useless.¶

Useless.¶

It effectively has no context.¶

Grok—right, Grok Bot never shuts down. From your start to your finish it stays in there,¶

It doesn't even have memory.¶

The Skills you installed, everything's still there.¶

Right, even your logged-in state stays.¶

So its cost is really high.¶

Which means right now we can't find a domestic competitor to Grok Bot—a true counterpart.¶

I don't think there is. Those VMs are all awful to use.¶

Doubao doesn't have one either.¶

But the cost is really high too.¶

Happycapy, is it? Happycapy—it gives a VM, I don't know, I have the impression it does.¶

Probably not either.¶

Doubao seems to, right? People at Doubao told me—no?¶

No. Try it, it's not.¶

Then I'll try it when I'm back.¶

Back to the core capability question again.¶

Input methods—does anyone still use voice input?¶

I do.¶

Which one?¶

In this week's news, NetEase Youdao released a free one claiming to benchmark against Typeless. Free forever. Because NetEase—remember UU's reputation took off a while back? Everyone operates their own remote-play products. This product probably isn't the same team as UU, but it's NetEase's, NetEase Youdao's—probably riding UU's traffic. They say NetEase has put out another product with a great experience for AI people. I haven't tried it yet, but their own reviews and comparisons go straight at Typeless. Right now I use Typeless and the one built into WeChat Input Method, both.¶

On my phone I use Doubao. But Doubao's problem on desktop is—I think it's that they haven't gotten the input-method part right: your dictionary, your suggestions. They should allow me to use only the voice. I'd use another input method for typing—it won't; it insists on being all bound together, so I can't use it. Every time I type, it comes out wrong. Miserable.¶

I get what you mean.¶

At that point you want to uninstall it. It's not that you don't want to use it.¶

Typeless isn't an input method, though. Typeless is just an app.¶

I only want the voice. When it mistypes four times in a row, you want to uninstall it first. You're miserable, you're furious.¶

The typing feel of an input method matters so much.¶

But WeChat Input Method's only problem now is you still can't customize the voice-input dictionary. If they fix that, I might not need the Tab key anymore.¶

If they fix it—well, this time he definitely won't drop it. But it's strange: you say, why can't big companies build such a basic feature?¶

They have the best voice. Like Doubao, they have—¶

I don't know if it's cost, or iteration being too slow.¶

Then it's management or content-moderation issues.¶

Definitely not cost.¶

Or maybe they figured the main push should be voice input first.¶

One more thing, phones. Guizang, are you on iPhone?¶

I won't admit it. I do have an iPhone.¶

iPhone's voice input is no good—you have to switch permissions, super annoying.¶

Doubao on iPhone—now that iOS changed, Doubao stays running all the time.¶

Changed—the new iOS changed it. iPhone's all got the new change. Disgusting—I can't accept that kind of thing.¶

No, it's disgusting too. Look, whether always-running or having to toggle permission every time, I can't accept either. So on my iPhone there's basically no usable voice input right now.¶

None of it works. A big reason I don't use iPhone: one is the battery, the second is I'm already used to voice input. Back on iPhone, out of habit I press the space bar to bring up input, and the moment I press it, it jumps away—and at that point I just toss the phone behind me and pick up my other phone.¶

The jump-away is fatal. People find that really hard to accept.¶

So now in the WeChat input box there's a microphone for direct voice input—that's my most-used thing right now.¶

Everyone uses WeChat the most. Even desktop WeChat added one: press and hold the mouse to input by voice directly—also great.¶

Press and hold the mouse and it appears? That awesome?¶

In the input box, press and hold the mouse and you can input—no input method needed at all.¶

Seems very intuitive, the interaction.¶

Very intuitive, but a bit annoying: that prompt keeps popping up, saying press and hold the mouse. A bit annoying—that prompt keeps coming up, press and hold the mouse.¶

But Typeless's core, I think, is its long-text optimization and cleanup—that I can't live without. Like Doubao and the others: their optimization is either overdone or absent. Doubao does none of it—only the most basic cleanup. It won't reformat for you or strip out the rambling, but Typeless does it beautifully. There was a period when it was pretty bad; now it's really good. And for people like us—when I write long things, it's basically all voice input.¶

WeChat Input Method handles long text fine.¶

Pretty decent.¶

OK, I'll try it.¶

But honestly there's no fundamental difference among them; they're all pretty good.¶

On transcription, everyone does it pretty well.¶

This topic interests me personally. Because I never figured out: why is a model super awesome and smart right at launch, and then after a while it gets dumb—everyone says "downgraded intelligence." I keep running into this dumbed-down thing—what's going on? From the model vendors' perspective, I get it: they must cut costs and boost efficiency; when usage grows they optimize, surely. But how do they optimize, specifically? I posted on X, and the guy with the most comments and likes gave four methods—I don't know if he's a researcher. He said the common one is routing you to a smaller model. If your question is very simple, they process it with a really simple model. But who sets the standard? Not anyone you'd pick. Something you think is important gets routed to a small model, and quality drops. Second, he said there's a value called the Juice value that can be tuned—also discovered by the community reverse-engineering GPT. For example when 5.5 just launched, it was apparently 700-something; after a while they dialed it down to 120-something, 6x lower. So its thinking tokens got fewer, so it got dumber. Then there's skipping thinking entirely and outputting directly, which also affects results. And then—you know multi-token prediction gives multiple candidate answers—they adjust the probabilities so the small model's result gets picked, saving compute themselves. I don't know, from a professional angle, or from—¶

Let me talk. Honestly, I think it's a bit conspiratorial. I think it's a bit conspiratorial. Two angles. First, personally: I rarely have this feeling. If you don't start out presuming the vendor is guilty from the get-go—I don't have that presumption—I rarely feel it. From the vendor's angle: as far as I know, the vast majority of vendors take internal evals of model capability, the assessment, very seriously. Including quantizing to save cost—they run evals, run assessments on how much it deviates, whether it's stable. People take this very seriously. Everyone takes their model's external reputation and brand seriously; they won't trade reputation for cost or speed. But the other angle: the whole LLM industry iterates extremely fast—a model's cycle might be one or two months, or a few months,¶

Or it'll—¶

so a lot of what happens in between might just be bugs. Even the vendor's own team doesn't know; users complain en masse—look, Claude and GPT have both had this happen several times this year, users complaining at scale; the vendor goes back to investigate, finds a bug caused it, fixes it, and things go back to normal. And then—also, I do think vendors may have policies, especially model vendors: for example, detecting your abnormal behavior, quota-farming or whatever, and deliberately dumbing you down—that's already product strategy, that's no longer—¶

But domestically the perception is very obvious. Domestically you get dumbed down easily.¶

My accounts have always been pretty healthy. Anyway I haven't hit it. Those are my points.¶

Also I think the core is: if they want to do it, they'll use a more conventional method. Like Claude, clearly: from 4.6 to 5.0, the models—Opus—have definitely become more and more inference-efficient. But the problem is you feel every model gets worse. Gets worse. They accomplish it through this normal model upgrading.¶

The user perception is real. Like Opus 4.6—it was actually the official side that dumbed it down first. The Amazon version to this day hasn't been dumbed down—it's pricier, but the AWS API has always had good results. The official one, on the Claude Code side, the dumbing-down is obvious. And then they stopped heading in that direction.¶

I still side with the vendors. It might really be bugs.¶

No—it doesn't matter. In practice, basically, community word-of-mouth is more accurate than Benchmarks now. Not that it doesn't matter—as a phenomenon—¶

Because, look, I was at SiliconFlow too. Both we and the model vendors we partnered with had these problems happen to each other—no deliberate malice, it just happened, got fixed, done. And about community word-of-mouth being important: honestly, when I choose or judge models I don't look at the various Benchmarks, or at the write-up articles—at most a glance. But I don't use them as my criterion, because I'm in a bunch of different community groups,¶

Watching people's real usage.¶

I check everyone's feedback roughly once a week. One is subjective impressions. The other, more objective: utilization, adoption—whether anyone is using it. If nobody uses it, say no more.¶

Like last time's bug, the Claude bug—if users hadn't pushed back, they'd never have fixed it in their lifetime. Took them over a month to fix it, right?¶

I think they fixed it because the reputation got too bad.¶

Actually I think this shows one thing: we're still at a very early stage. A model that's stably provided, low-latency, usable, with its intelligence in place—look closely at today's models, there are very few.¶

A model's lifecycle is just that long anyway—why put so much effort into it? It's only that long.¶

So the models worth using right now are either Astra or Flash; that pile in the middle is all awkward.¶

Like we said last time: anything inside the kill line is meaningless.¶

For users it really is like that. They themselves might be a bit—¶

There's also a European model. I remember using Mistral ages ago, two or three years back, when I was doing Agents. Recently they seem to have just raised a round—European. Foundation-something invested. No hope left, right?¶

Europe's sovereign LLM.¶

I think its reason to exist is entirely that Europe needs one. How good it is doesn't matter.¶

Everyone needs one. How good it is doesn't mat—¶

Their new model is a fine-tuned Gemini.¶

Right now it's really only China and America that have the ability to do this. Europe—well, right, whose model did they distill, who knows.¶

Didn't you say it was Kimi, Guizang?¶

Their latest is fine-tuned from Kimi 2.7.¶

Fine-tuning is enough. Their capability isn't there anyway.¶

The thing is K3 is already out and they're still fine-tuning Kimi 2.7.¶

But didn't you say yesterday Kimi released a really weird version number?¶

Kimi released a 2.8. Kimi 2.8. Judging by the version number, it's definitely not K3 architecture.¶

Definitely not.¶

K3 itself is very large and fairly power-hungry to run, right? Because it's big. Though its intelligence level is high. Look, everyone's shipping Flash models, so Kimi definitely needs to ship a Flash model.¶

Everyone's shipping Flash models, so they need to ship a Flash model.¶

But the strange thing is it isn't called Flash. Normally everyone's expectation would be it's called K3 Flash,¶

But the strange thing is it isn't called Flash. Normally you'd ex—¶

But I'd guess it's internal resource allocation or product cadence. I figure they couldn't produce a K3 Flash, but still needed to ship a lower-cost, faster model—so maybe the team chose to build it on the 2.x architecture instead.¶

It can't be made into a Flash—it's 1T. How would you make a Flash out of that?¶

The Flash naming is weird too. What's it actually based on? Why does everything coming out recently get called some kind of Flash model? And they're all multimodal.¶

Flash originally meant fast. Before that there was Turbo, right? Something—¶

Flash originally meant fast, and there was Turbo, right?¶

But I think it's now come to mean small-size flagship.¶

First of all, Flash is Gemini's greatest invention. They invented Flash.¶

Pro, Flash.¶

How so? What do you mean "greatest invention"?¶

Gemini invented the Flash tier. At first it just meant fast—capable of spitting out text at over 100 tokens per second.¶

But parameters—the actual positioning now is small parameters.¶

Not that small.¶

But now everyone has this what-do-you-call-it, the lower beating the upper, right? Gemini's Flash beats its Pro, DeepSeek too.¶

Why? Is it all because new architectures make them stronger?¶

New architecture, 500B—right, like 500B reaching 300 tokens per second.¶

I think it's related to post-training capability improving. And you'll notice the cycle is very short—suddenly everyone has, all at once, mastered strong post-training techniques.¶

Why is that?¶

No secrets in this circle. One team breaks through, and everyone else is classmates and friends, eating together every day.¶

One weekend party and everyone has it.¶

Smaller sizes make post-training easier, right?¶

But hasn't China been pressured into a bunch of black tech? Because look—¶

That point I especially agree with. China, precisely because compute is insufficient, has been pressured into black tech—and it turns out they're leading in reverse.¶

But overall, from a model standpoint, isn't there still a 4-to-6-month gap? Even if on some pre-training and post-training techniques—¶

I used to agree with that view of yours, but as I said last time, I think OpenAI and Anthropic are holding too much back. What we see is just the consumer side. What's released might lag by 4 to 6 months, but what's not released, I think, lags by more.¶

Anthropic's Model 2.¶

And those two have already entered the inflection point: using model intelligence to train models.¶

Right, ASI.¶

We haven't crossed it yet.¶

We haven't used it yet. I think Teacher Tang Jie said GLM 6 is—¶

Isn't everyone saying that now?¶

Next rung—GLM 6 should be arriving.¶

Models training models—haven't people been saying that for ages?¶

Saying it, yes—but only OpenAI and A-livestock have done it.¶

So from that angle it's pretty scary: America reaching the singularity ahead of you.¶

Then you can never catch up, that kind of—¶

That's honestly quite scary.¶

Once it's passed, there's no chance left.¶

Also, didn't an Anthropic employee resign?¶

And posted a whole pile of statements.¶

150 million impressions.¶

His view was that both OpenAI and Anthropic internally failed to rein in—failed to hold back AI—¶

development, right?¶

When I did an in-person talk back in '23, my last slide said exactly this. Because my personal belief, starting in '23, has been AGI within five years. That's my belief.¶

"AGI within five years—that's my belief."¶

The last page of my deck said: for the next five years, cherish the time with the people you love around you. Because what happens after five years? Two possibilities: one, things get even better; or—beyond description.¶

This brings up Anthropic's report. They published an employment report.¶

Everyone thought it was just a report.¶

It projected AI's impact on world GDP over the next 5 years. It said in the most optimistic case, GDP—sorry, GDP rises about 30%. But white-collar wages barely change, up only 0.5%; blue-collar wages rise 30%; returns on the capital side rise 80%.¶

Isn't that the same as now?¶

Capital only needs to combine itself with AI and robots—no people needed—and it can grow on its own.¶

Actually these past two years too: the biggest gains have all been taken by capital.¶

GDP and all that—the whole world is growing, but daily life hasn't changed much. And another change is—¶

Astra 3D. Astra 3D—I suddenly remembered: 3D. Once 3D modeling is done, the 3D printer—its usability suddenly jumps up a huge step. Really. Before, with a 3D printer you had to load filament, do the modeling—you could only use models other people in the community had made,¶

Couldn't operate it.¶

And other people's models don't fit my needs. Like one made before—turning an iPhone into an alarm clock, snapping in, charging fully. It was for iPhone 18, yours is iPhone 15—you can't modify his model, really can't. Now it's easy. Before, operating Bambu's software took some effort: how to load material, what to do about jamming, how to dig out the supports—now you don't have to care. Computer Use just—¶

Before, you talked things into existence in the digital world; now you talk things into existence in the physical world.¶

A few days ago I tried Rodin and it was great. We made Optimus Prime—with its MCP, I never even opened their website. I said, make me one.¶

With a 3D printer you could print your own Optimus Prime.¶

I don't know how faithful it is, but at least the model looks decent.¶

This Bar has one—print one.¶

Print one in a bit.¶

We can try it later.¶

There's a lot you can do in this area, and materials have broken through a lot recently too, so more and more—¶

It's starting. But yesterday we all watched Han Qing's concert, and you can't help feeling that embodied AI's development in the physical world is still a bit slow.¶

Slow has its benefits, right? Too fast isn't good either. Look at office work—it may be that in two more years there's no office work left.¶

No of—no need for office work. Office what? Nobody needs to work. About these robots lately—¶

What about the robots?¶

The robots—lately. Since Unitree went public, its market cap—¶

Basically represents a stage peak, I'd say. Time to squeeze the bubble.¶

Let's see if there's anything to add. Anything else to add?¶

Nothing left, right? Nothing—let's leave it here.¶

Nothing left, pretty much.¶

Let's wrap it here—I'm already hungry.¶

Wrap it here—you're hungry, sir.¶

About done. Hungry.¶

Nothing more to discuss, right, the talking—¶

Let's thank our venue.¶

Thanks to Shanghai AGI Bar. We're recording at Shanghai AGI Bar right now; there's one in Beijing too. I hope AGI Bar keeps growing better and better, records all over the country—record wherever you like. That's it for the recording of episode two of Next Token | Beyond Tokens. Thank you all for joining. Thank you.¶