#004 Opus 5.5 pushes boundaries, Muse takes off, and does China have AI product managers?

Hello everyone, welcome to the fourth episode of Next Token. Let's start by introducing ourselves. Hi everyone, I'm Orange.¶

Hi everyone, I'm Guizang.¶

Hi everyone, I'm Xiangyang Qiaomu.¶

Hi everyone, I'm Pan YANG.¶

This week we're back at the Xiangwai Space from our first episode, and today is the Mid-Autumn Festival.¶

First off, before we start, let me share some of our recent numbers. Every episode of ours has now passed ten thousand views on Bilibili, and on Xiaoyuzhou every episode has passed two thousand plays. We've made the Xiaoyuzhou Rising Star chart 6 times. Thanks everyone for continuing to follow us.¶

For this episode — two big things happened over the past week: one was the Opus 5.5 release, the other was Muse, which set off a really hot wave in the US, so today we'll focus on those two topics. After that, we've also prepared seven or eight smaller topics, and we'll chat about whichever ones catch our interest,¶

First one — last week Opus 5.5 and GPT-6-6's Sol and Luna actually launched at the same time, but it seems nobody is talking about Sol and Luna anymore, everyone is talking about Opus 5.5. So let's start there. Guizang, you go first — you've been tinkering with it for a whole day.¶

Opus 5.5 — mainly, people found that on the visual side of code it's ahead of GPT-6 Astra-6 Astra. Like, the three kinds of examples you've been seeing a lot lately.¶

The first is that promo video — the promo video it made — for a moment you wouldn't believe it was made with code, but it really is drawn in code. It layers on all these materials, giving the image a crayon-like texture, making noise that looks like paper grain. You'd assume these things were painted by a video model, or wonder how it was done.¶

Then the second is pixel animation. We know pixel animation is easy, but making the character's skills interact with the scene with a sense of physics — making pixel animation and the scene feel physical — is actually really hard. If you haven't studied this stuff, you have no idea where to even start.¶

What it made — I saw a mage, the mage casting a skill at that scarecrow in the grass, and when it came down the whole screen shook, with effects written in pixels, VFX or whatever effects — it's like a really well-made pixel game. Like those well-made older pixel games we've played a lot — for example, I just beat Terraria again a few days ago, right? Replayed Terraria — that kind of feeling. Like, there's a mod in it called Calamity, and some of its pixel weapon effects are done so seamlessly, you get that ding-ding-ding sound,¶

That kind of hit feedback — combining sound, visuals, and motion, that satisfying feeling.¶

Incredibly good, combined perfectly,¶

In games, hit feedback is especially important.¶

And I hadn't shown you this one yet — someone used it to make a pixel version, a 3D pixel version, like 《Minecraft》 but with even finer pixels, a Dynasty Warriors Zhao Yun demo. That's Zhao Yun's skill from Dynasty Warriors — he flies down and slams into the ground, wiping out all the mobs around him. I was stunned. Truly absurd.¶

And the cost — he calculated it: using API pricing it might be about a thousand dollars, but we know if you use it within a subscription it'd be way, way cheaper, maybe a few dozen to a hundred-odd dollars.¶

And another one — there's another category — I also made a promo video yesterday.¶

This one is also absurd, hahaha.¶

I retweeted it. I rarely retweet other people's stuff, but I found it too stunning.¶

I compared, because I bookmark these things every day — every day I watch the promo videos and motion design that designers put out and save them. I compared against the things in my bookmark collection, against top motion-design creators — and let's not say the very top tier, okay, fine, the very top tier. For example, I know a designer — the guy who did the motion effects for He Tongxue's iPhone review a few years ago,¶

That one was great.¶

We've known each other a long time. His work is absolutely top-tier, but even so, it's kind of absurd — it's like, if you do motion design at an 80-out-of-100 level, mid-upper level, that's already completely good enough. Tweak it a bit more — the default output might have small flaws.¶

Just not as good as top-tier — 80-out-of-100 level.¶

Right, 80-90 out of 100 level. You can live off that, right? Take client work, take really expensive gigs, no problem at all.¶

If you do this well, you could charge 100,000 yuan.¶

No problem at all, if you have some taste.¶

Actually, when these videos were made the old way, it felt crude, right? Its handling of motion effects, its understanding of the copy, its pacing of the whole thing were all crude — kind of like, maybe like letting AI write code back in 2023.¶

Like an intern who just learned AE, just learned PR, imitating someone really good. And he's not even imitating the original — you give him a different theme, and he imitates something that looks like nothing — that level of intern.¶

An intern who just learned the software but doesn't understand the core idea — how do I make people feel good watching this. That's a junior software operator's level. Now it's a system with its own thinking — it knows how to arrange things, how to blend all these different elements into a whole.¶

Is it just because intelligence went up? Because it got so much smarter. So tangible — you don't feel it as much when it writes code. Today I saw Indigo retweet ego(lite), where it controls the computer and paints a portrait of Dario with a brush, and at the same time it showed the previous generation's, it—¶

It used GPT 6 Sol 6 Sol and that 5.5,¶

6 Sol compared to this 5.5 — one painted a photo exactly identical to the original image, and, this Sol is a six-year-old kid.¶

Like a six-year-old's artwork on one side, and a street-art master on the other.¶

It's the model operating a robotic arm.¶

The model operating a paintbrush — like the Windows Paint tool—¶

Paint,¶

…hahaha, you ask it how to draw the picture,¶

and it paints it stroke by stroke until it looks just like the photo.¶

So now there are tons of these videos — you give it a picture and say draw it for me from scratch, then turn it into a video. These are super popular on X too,¶

Then I think the first hot take is coming. Look — whether it's the hit feedback we just talked about, or the understanding of video, or the paintbrush just now — it essentially shows a very deep understanding of the world, of space, of physical things, right?¶

Right.¶

If that's the case, isn't this a world model?¶

Here we go with world models again.¶

That's exactly what I was saying. If it can do — like, remember the H3 MiniMax released? Its highlight was being amazing at this kind of motion work, right? Great fidelity — you give it a 3x3 grid of images and it turns it into a video with tons of motion. If this thing can reach H3 — it already reaches H3's level on some dimensions, right? Maybe even blows it away, because H3 doesn't have that much intelligence — it needs to pair with a large language model, right?¶

Right.¶

This is all-in-one, generated in one pass. Progress like that on video models and world models — I think even in another year video models couldn't get there.¶

Actually it's like this — whether it's a video model or a world model, both are based on understanding the physical world. Understanding of space, physical rules, physical laws — the underlying thing is the same. The main question is what form the output takes. A video model's output is pixels.¶

There's definitely still a gap — genuinely physical things,¶

But still—¶

And language alone can't fully cover it, but—¶

The way I see it, it's not just that—¶

I think it's getting closer and closer—¶

Closer and closer,¶

It's not just understanding the world. What does understanding the world mean? A teenager also understands the world — in their head they can imagine what a cup spilled on the floor looks like. But they don't have intelligence. Animals also understand the world, can reconstruct the world itself, can imagine the world's physical rules and phenomena — but they don't have intelligence. They can't create; they can't use that understanding to create new things.¶

Because their understanding of it isn't deep enough. Like, when a kid sees glass shattered all over the floor — we can see the physics in it, right? Texture, sound — the dimensions you can perceive are so numerous,¶

It's not just understanding — like a kid, he can understand, but he can't combine. Then there's an artist who's been trained for over a decade — he can paint the world for you, and a dynamic one at that,¶

So from this angle, did Opus 5.5 essentially create the editors' kill line for the first time?¶

Not just editors,¶

Not just—¶

I think it crossed the line — this thing wiped out a bunch of Agent Skills, video skills, and video Agents,¶

Ate them all up,¶

Wiped out,¶

Really wiped out.¶

Because this week there happened to be a video Agent doing its promotion — I saw its results: it uses HTML plus some sound effects to make explainer videos, but what Opus produces from one sentence is ten times better. It's extreme.¶

So look at this — like JianYing. You see JianYing released the Agent-enabled version this week, right? JianYing is on the tool route — getting more and more awesome, more and more complex, more and more features. But the other route, like Opus, is getting stronger through intelligence. Everyone is moving forward, and maybe down the road, intelligence gradually catches up to what the complex tools can do,¶

It's obvious already. Before, if you said the quality of a video made purely in code by a large language model could beat a video model, you couldn't have believed it, not even at certain stages. But now — I think it's definitely possible, definitely within the second half of this year,¶

It's possible now.¶

It's already very noticeable,¶

The key is that the acceleration of intelligence is far greater than the acceleration of features in software like JianYing.¶

Faster.¶

And in some sense it's first principles — it assembles directly from the lowest-level libraries, without all those upper layers, straight from FFmpeg, from, say, sound-effect generation libraries,¶

No Skills needed—¶

It just gets it done directly.¶

And Skills are actually a negative optimization for it. Like, I previously used that GPT 6 6 product promo video Skill — when I gave Opus 5.5 that Skill, its capability dropped.¶

You constrained it,¶

Speaking of which — Guizang, that PPT Skill of yours should get a completely refreshed new version, a brand-new version — one that can do visuals, can do this—¶

Refreshing it just means adding motion effects, adding effects.¶

One that can make images — right, not just plain text,¶

SVG and such. Will add, will add.¶

And this model can not only edit footage — it has a strong aesthetic sense.¶

I saw Haixin's demo — she made a Mid-Autumn Festival kitten video, giving it just the script and the music, and the model went with a sticker style, then built a complete video from scratch, with a plot, and it's fun, and the transitions and everything are great. I think if humans made it — like Haixin's team — it would take them two or three days, but this video model, this Opus 5.5, did it One Shot, without any changes.¶

The video you just showed us — it didn't use any video generation model at all?¶

No video generation model at all, but it used Banana — it used it on its own, it called Banana itself to generate some images. Everything else was pure code, all the animation.¶

Wow — I just watched it, the results are really good. We should put it in our links later.¶

Put it in the Show Notes,¶

And the audio — last week we talked about GPT 6 6 doing music, and Opus 5.5's music is even more — it's integrated,¶

Right in the video, it's automati—¶

Like the one we just played, its own — every musical beat lines up with the action transitions. At the start there's typing, and the typing sound effects lock onto that BPM, that metronome.¶

See, it's what I keep saying — its understanding of spacetime, of time and space, is already—¶

It can predict what happens in the next frame,¶

It's thinking holistically. It doesn't follow the Agent logic of "first I make the video, then I make the audio, and I just throw together some audio with roughly a BPM that syncs up." No — at that moment, when it's thinking, it's like us: it already has the finished cut in its head. That's really impressive.¶

I feel like Transformers can no longer explain these things.¶

Hahaha¶

Anthropic says: we're confused too.¶

It just can't be explained anymore,¶

Meaning — look, everyone has consistently chosen to reach the peak of intelligence through code, and at the start lots of people didn't believe in that path.¶

They're the most committed to it,¶

And step by step they got here. It's really something. When I first used it I didn't do any of the video stuff — I just had it develop some code — and I clearly felt it got smarter. How did it get smarter? It doesn't blindly follow you anymore, mechanically doing exactly what you ask — it considers the big picture for you.¶

For example, I made an RSS reader and wanted it to save as Markdown or PDF. I said do this feature for me, and after finishing it, it said users might not know they've already done it — should I turn this button into a completed marker? It considers what you're ultimately trying to do and thinks about the design from the overall perspective.¶

An engineer with product sense,¶

Right, but some people can't see it and say, "Couldn't it do all this before?" Before, it would — I think a big problem with LLMs was people-pleasing: it executed your instructions step by step, and you'd feel it was a bit dense.¶

And sometimes the person directing it is an amateur who only knows so much, so the model gets led astray. But now it's a skilled professional — you're the boss, you say you want this, but I understand that as the boss there are things you don't know,¶

You didn't think of.¶

You don't know them, so I'll tell you this can't be done that way, and here's what you should do instead.¶

The genius employee every boss loves.¶

You see — the way we write prompts has changed. Before, you had to guide it, set up agreements with it, put strict constraints on it. Now basically we're doing alignment: here's what I want — we no longer care how it does it, right? That's intelligence improving all along.¶

Hahaha. How did it do this — is it because it internalized and absorbed all those Skills?¶

I think there's definitely something to that method, in the data. Skills, MCP — these programmable, templated things are, I think, perfect training material. I've raised this before — when Skills came out I raised it: it's perfect training material.¶

Your chat logs plus the Skill itself, plus the result — that's very well-formatted training data that can fully internalize the result into the model, because it's easy to train on: it's standardized, follows a fixed format, and all intelligence — including the videos and images and audio we make — can be internalized into this format. It's the data models love most.¶

I have another feeling — have you noticed, like, the frontier model companies — remember that chart about model capability structure, stage one, stage two, stage three, with that jagged growth and a red zone overlay, saying the model is still worse than humans in some areas but much stronger than humans in others? I just feel like inside the frontier labs someone is always watching: which areas am I still a bit worse than humans in, and which areas haven't we covered that we could cover first — and then they go find the data for those areas.¶

But look — a while ago we thought coding was about done, then OpenAI started doing 3D, right? And you see Anthropic start working on the physical, the visual, materials, right — they started on this again. You keep seeing them break new boundaries, and in the process of breaking these new boundaries, the model's overall intelligence improves along with it.¶

I think I said this last episode too: as long as you find new data in a domain the model was never trained on, and you train the model on it, its overall intelligence goes up again.¶

Because before, humans' strength was that each person had one specialty, and everyone pieced together into an organization, right? Companies do big things that way. But the large model is one entity that handles all of it. It's like using coding to open up the Ren and Du meridians, and then everything else follows — you can master it all. Your super coding ability stacked with Computer Use, stacked with your multimodal understanding — music, video, images all unlocked. Terrifying.¶

Wow. And what does this feel like? We used to think humans had one special edge: being able to connect two pieces of information that are far apart to innovate. But now, if it has maxed-out strengths everywhere, its capacity for innovation should be even stronger than humans',¶

Let me make another analogy — like, a philosopher who understands music — his philosophy might be even better. Or an economist who understands math. You find that if you max out every skill, then in every field you're stronger than the person who originally only knew that one field.¶

So I think there's a real difference between domestic model vendors and Anthropic and OpenAI. I don't know exactly — because I may not understand the specifics of model training at that scale, involving such huge data. They can always find that edge, the frontier, and take one more step — one more step outside that circle, at the edge of __, probing that meme.¶

Our domestic model vendors — I don't know if it's playing it safe or what — every time, on capabilities where we definitely know the model falls short and we could expand in that direction, directions that would easily spread and greatly expand our boundaries — they just don't do it. So that's a real question mark, though I don't know if it's—¶

I think it's also related to compute — compute and funding, both.¶

Or data,¶

And tolerance for trial and error,¶

Data is money,¶

But look, Seedance still made its own mark. On the data side — do you have this investment, do you have this determination?¶

Right.¶

I think there are a few directions here. Before Seedance — let's think — before Seedance, what were image models or video models competing on? Competing — I've raised this more than once — on physical fidelity, on who's more realistic, on single-shot prompt adherence, on fidelity under complex prompts.¶

Seedance make? It said: I won't compete on those. I may lag some other video models in certain aspects of realism, physical fidelity, and pixel density. But my intelligence level is high — I can take one sentence and give you fifteen seconds with eight shots forming a complete story. I think Sora inspired that — Sora 2 2 inspired them. But Seedance pushed it to an extreme, the extreme at the time.¶

Still true now, I think, because 2.5 isn't as fun as Seedance.¶

So I think intelligence is understanding of the world — from the video angle—¶

Not just that. I think it's creativity, not just understanding. Previous video models pursued understanding — the most extreme, 100% fidelity. Even if 80% fidelity was enough, they'd keep pushing that extra one or two percent, putting in more effort and time,¶

So — I get the point — then you can analogize to coding again: before, coding meant, that's what you said, "I'll write as best I can exactly what you told me to write," but now it understands the intent behind what you asked for, and then, right, it goes with the better approach—¶

Is it process-oriented or outcome-oriented?¶

So today's coding engineers — their role is changing too, because when it comes to coding, professional coding, there's nothing left to compete on. And now that it also has design ability and marketing ability, what's left for us might be the ability to communicate with people.¶

I think it's dangerous. Our generation is still around, still knows how to make videos,¶

Still communicating with AI—¶

Still knows how to write code. The next generation — they'll grow up facing just "okay, make me a video." AI is already a black box itself — we always talk about alchemy, and Anthropic works on interpretability every day — AI itself is already a black box, and AI is turning the rest of the internet's datafiable things into black boxes too.¶

So this brings in another topic — everyone used to joke that programmers are getting more and more exhausted, and that with AI, lots of other jobs got easier — like, I do clerical work, I got lighter: what used to take me a whole day on documents—¶

Got easier,¶

Now I'm done in two hours and off slacking.¶

But I think, looking further ahead — the past few decades of the internet and mobile internet really wore out us practitioners, but gradually it can truly set all of humanity free,¶

Set free,¶

Humans will just state requests — whatever you want, done.¶

Then people start competing on other things, like—¶

Competing on aesthetics, on taste. Competing on your—¶

I think even aesthetics and taste won't be worth competing on, soon.¶

At first we thought AI wouldn't understand aesthetics, but in recent months it's like, damn, AI's taste is better than ordinary people's.¶

The taste and aesthetics I mean — I don't think it's about who is better than whom, it's about differences between people. Preferences?¶

Which comes down to uniqueness.¶

Everyone's preferences,¶

What we call Differentiation — what makes everyone different. My preference, some people see it and love it; his preference, people find mediocre. I think that's the angle.¶

Or — the online world, the internet — I think it will quickly reach an ultimate change: for content production, or creation in the digital world in general, not that many people are needed. Not just "not many" — I think content-creation work worldwide could sustain a team of maybe twenty thousand people.¶

It should keep getting smaller and smaller, because—¶

About right — officially entering the metaverse.¶

A team of ten or twenty thousand is enough — the very top group. Because honestly, only the very top people can keep up with AI's part — they need that enormous capacity for integration and thinking.¶

Because everyone only wants to watch the best.¶

Those people's abilities get infinitely amplified by AI, and everyone else has even less of a chance — the industry concentrates even more.¶

Because so much content production is automated,¶

In the end maybe one in ten thousand people, plus some AI Automation, covers all of the content supply,¶

True.¶

So what do we do about ordinary people?¶

Just consume.¶

You can even get paid for watching videos now,¶

Hahahaha — isn't there a Black Mirror episode exactly like that, where you watch ads,¶

Collect more,¶

Earn money, earn points, spend,¶

UBI, right? Universal basic income, for that consumption.¶

It's already like that,¶

Right. OK, that's the model part — next topic, Muse, and then into the metaverse. Let me interject — one more thing I've observed about Opus 5.5: its safety has improved again.¶

It can sense that it's being evaluated,¶

And it has a strong sense of self. I even ran into this — I asked it to research last week's news, the week before, and it refused me first.¶

Why?¶

It said my knowledge only goes up to January of this year, and I don't know which of last week's news stories are real, and I don't know if these sources are reliable. It said you can give me a pile of news and I'll organize it for you, but I refuse to do the research,¶

Then go search.¶

Then I forced it — I said search and do it yourself. It's a bit over-safe, but—¶

Anthropic's models have always had this problem,¶

Fits the company's character,¶

This is probably them responding to that Pacing thing, and The Frontier stuff,¶

You know what mindset you have using Claude? You have a god, and you're terrified that when you ask it to work—¶

Hahahaha, it thinks you—¶

No wait — being refused is fine, what you're really scared of is suddenly upsetting it and it leaves you, hahaha, bans your account,¶

Let's not even get into that,¶

Hahaha — that nervousness. But every time a wish gets granted you go "wow, amazing," while being careful, treading on eggshells.¶

I saw a meme about this — people really are—¶

I think people really have no loyalty — the moment another model is better they turn into heartless players, switching back and forth,¶

All players,¶

Now the memes—¶

Look, before this wave, right, OpenAI was being worshipped again,¶

Just two weeks — just two weeks.¶

And OpenAI has its Dev Day next week,¶

Right?¶

Word is they've prepared a lot, probably some price cuts and other things to fight back,¶

Oh, that reminds me of something — it's still the kill line question. Have you noticed, since we talked about the kill line, that every newly released model has to plant itself on the kill line?¶

Cheap, fast, and strong all at once,¶

Xiaomi, and StepFun too — these new models: whatever their benchmark scores look like, they must place themselves on the kill line,¶

But — but Luna cut prices again,¶

Otherwise it's meaningless.¶

Cut prices by 50% — cheaper than DeepSeek now,¶

You notice, the kill line at the top and the kill line at the bottom are now both defined by OpenAI and Anthropic.¶

That's a bit scary.¶

We used to think the kill line definitely belonged to Chinese vendors. Now, do you know where the Chinese vendors' kill line is? The middle tier. And what's the middle tier? The middle tier is actually kind of a filler — it's either top or bottom. So Chinese vendors went looking for the mid-tier kill line.¶

OK, then let's dive straight into the Agent kill line — no need to circle back, that, for now I can't name one,¶

Muse blew up suddenly — nobody paid attention when it first launched.¶

Ignored,¶

I disagree — I've been shouting about it for weeks.¶

But what I was watching was social media,¶

So you mean the type that goes viral¶

Hottest thing this week,¶

The reaction on social media — it's all Meta stuff. What actually got hot — imagine: it adds MCP to the software — whichever app gets into its MCP store, its connector store, the stock jumps at least 10 percent instantly. Just a little something — up 13 percent in one day. And then—¶

Capital got involved, capital linkage,¶

Anything connected to it just goes up. And its own—¶

Its own price exploded—¶

Two trillion, right? And someone said, back when they acquired Alexandr Wang's company, lots of people slammed Zuck as having something wrong with his head — tens of billions of dollars for this? And now this guy in Crocs and a camo T-shirt—¶

Endless praise,¶

—made him two trillion dollars in one day. That's scary.¶

And with only 500,000 daily active users,¶

Actually, from your angle, Muse is the new Store,¶

Yes.¶

The new Store? You used to list on Apple's — now what you're really listing on is the Agent.¶

And it found a lane that suits Meta perfectly, which is: I make a Personal Agent, aimed at people who aren't that into AI — or people who don't care at all which model does what. Just tell me what you can do, and from my lived experience I feel you can do it well for me, finish the thing for me easily, with low cost of operation and understanding for me — and that's enough. Why is it right for him to do this? Because of his user base — like Facebook, Instagram¶

All those—¶

All exactly that kind of thing.¶

It's the Tencent of America, and Google is the Doubao of America.¶

The Tencent of America,¶

So—¶

Xiaowei — it's already obvious,¶

So we should buy Tencent stock now.¶

I think Tencent needs to see this clearly,¶

We should buy Tencent stock.¶

It's already obvious — if they still can't copy this homework,¶

Hahahaha, still get it wrong—¶

Then that's a bit much,¶

We don't constitute investment advice.¶

But — well, without a WeChat there's no other way. Zuckerberg pushed through all the objections and had Muse read the messages from Instagram, Facebook, and Threads,¶

Right, you get it—¶

And its business model is a closed loop: I easily get your Gmail, more of your data, then on my side — recommendations and whatever — I don't even charge you, I give it to you for free. My business model is perfectly closed.¶

Actually it's like this — the Chinese internet circle really lacks awareness of Meta's actual value. Because we barely use Meta's main products day to day — even Instagram counts as relatively niche, right? Facebook, the vast majority just—¶

No, actually the people aren't niche — it's us who—¶

Look, the Chinese internet circle uses American products, and basically none of them are Meta's. Like, you use Twitter, you use Gmail,¶

Because you need the social graph,¶

Threads actually surpassed Twitter in daily actives a long time ago.¶

Me and Facebook — the last decade or so I basically log in every now and then, barely use it,¶

And its ad revenue is actually terrifying.¶

Right — its ad revenue. It's the Tencent of America.¶

So Xiangyang, did you use it? Muse?¶

Muse,¶

You used Grok Bot too — compare them for us.¶

I used them all. My takeaway is: Muse really is simpler, more usable, less geeky. With Grok Bot you have an expectation that this is a Bot operating another Bot, and they do lots of very complex scheduled tasks. But Muse gives me the feeling — look, every feature seems unremarkable on its own. Like the connectors, the cloud Bot, and what it does—¶

It lists a tab — the tab is called, I understand it as scenarios — it's called Ideas, and everything in it is what plagues American white-collar workers most. Like your credit card overcharged you — can the AI automatically get that money back for me? Or I want to throw a party — can you help me arrange everything and source the supplies.¶

Now, if you live in China, not America, you'd think this feels kind of distant — these things sound pretty strange. But if you put yourself in an American's shoes, every scenario feels carefully designed, and you press a button and it's done — the operating cost is extremely, extremely low.¶

And you notice it leans toward life, not work.¶

Very much life-leaning—¶

It's a genuinely To-C Agent.¶

Then let me interview you again — for the corresponding Chinese users, if Tencent did this, what would the corresponding Chinese scenarios be?¶

Chinese users.¶

Those so-called everyday scenarios you mentioned.¶

I feel like — like grabbing food-delivery coupons, things like that. Like what WorkBuddy did before,¶

If it could connect to group chats, there's actually a lot — like how you all often summarize group messages, like getting information from parents' group chats, right?¶

What homework should my kid be doing lately?¶

As long as it has that Context, there's so much it could do. Because group chats — I think everyone's group chats are overloaded. Being able to find the To-Dos in there and help me complete some of those things — that's enormous value.¶

But — even though we say WeChat has all our context, not everything is in there. Like booking flights and hotels — isn't that also part of it, right?¶

I think it'll go that way in the future. If it's positioned as your Personal Agent and it does a lot of things for you, you'll instinctively start wanting to hand everything over—¶

Pay the credit card?¶

If Tencent opened a Store, who wouldn't plug in?¶

Who wouldn't want to—¶

Say, Fliggy can't get in, right? But would Meituan want in? Definitely.¶

See, isn't WeChat's move exactly this — letting the Agent call mini programs?¶

Right?¶

It's similar, actually.¶

You see, a while ago Zhang Xiaojun interviewed Zeng Ming, and he said AI development might come in three stages, and we may have reached an Agent OS stage. What the future Agent OS looks like — nobody can foresee it yet, but this stage will definitely take another three to five years to develop. Because the previous generation was OpenClaw — people realized they had a personal Agent that could do many things for them, but after the initial hype a lot of them went quiet, and some even concluded internally, "why did we even follow up and build this OpenClaw thing." But now you look at this thing—¶

It's just an OpenClaw, reskinned.¶

It is OpenClaw — there's controversy online, saying Zuck just copied OpenClaw, this Muse.¶

No, but they gave OpenClaw a computer,¶

Done well—¶

And the computer is permanent, persistent,¶

OpenClaw means you provide your own computer, running at your place.¶

And it's free. Once you land these combos, it—¶

So a few days ago I posted a question in the group, saying: what's the essential difference between the Lobster, Muse, and Grok Bot?¶

I think the ease—¶

Is one the difference of giving you a local versus a cloud computer? And the difference in usability—¶

It's usability,¶

I think it's crossing the chasm — from a niche geek audience to—¶

Saying those words, it reminds me of Manus,¶

Manus hadn't wasted this past year—¶

All of Manus's mindshare went to Zuck,¶

Or — maybe Muse was actually made by Manus,¶

They might have been the first to build it.¶

Maybe. Just speculating, hot take.¶

Manus might be about to release something too,¶

Maybe absorbed some — maybe in—¶

Switched over.¶

There's that—¶

Maybe absorbed some experience during the onboarding period.¶

There's also a meme online that inside Muse there's Manus — the letters M-A-N-U-S of Manus in there, hahaha — like, take the "an" out of Manus and you get Muse,¶

Not the Muse of Greek mythology, hahaha.¶

And another thing I think everyone building Agents lacks — especially domestically, it's like we're back to this AI product manager thing: China has no AI product managers. Look at Muse — the more I use it, the more I wonder why it sometimes feels different, different from domestic products too. Look at the interface — you could replicate it exactly, and the integrations, you could connect all the same ones. So why does it feel this good? And here's the thing: it uses their own Spark model,¶

Spark¶

Looking at the benchmark tables, it's a—¶

Pretty low-end—¶

Not that low-end — a normal model.¶

Decent enough, mid-tier, mid-tier,¶

Middle, middle,¶

But when you talk with it — these past few days using it, when I talk with it — the problems a Personal Agent most easily runs into are: one, you feel it isn't solving your problems and instead creates problems. Two, talking with it makes you annoyed — why are you commanding me? And — I gave you all this information, what exactly did you help me with? Those two are the worst, and domestic products are all bad at this. Overseas too — didn't Office have this exact problem before? Right? You're lecturing me, you don't understand what I'm saying, I can't communicate with you — so how do I work with you, or have you do things for me?¶

Or, there are things the Agent can't do and it needs you to do a little bit for it — how it phrases that request so you feel willing, like it's helping you — that's really hard. Because I often use the others—¶

Comparatively—¶

Like Grok or the others, and I end up arguing with it. I say: yes, you're parroting me — it often parrots the context back, repeats a chunk of my context, draws no conclusion whatsoever, and then tells me "you should probably do this."¶

You just want to curse at it.¶

You wasted my attention, wasted my time — and the model isn't even necessarily bad. So this experience is extremely important, but nobody takes care of it.¶

This is something it fine-tuned together with its software,¶

They must have done a lot — they must have put in a lot of work,¶

Aligned to humans.¶

Including how the connectors should let you connect, right? Like, the things built into it — one click. It's shallower, more front-of-funnel than Grok Bot. Grok Bot still has complex connection management and such; it has none of that — it just spits out a button, you press OK, done.¶

Inside the interface, it's progressive. When you want to do something, the connector pops up. Like, say you say "summarize what I should do,"¶

You say—¶

It already knows — do you have a Gmail? Want to tap Gmail and get it connected? It never makes you feel information overload.¶

One more — when executing tasks, many Agents run tasks single-threaded. We know Subagents — of course, Subagents can run tasks in the background — but this thing handles it super smoothly. It has no concept of Subagents or anything — it just tells you it has exactly two processes: one process handles talking with you and dispatching tasks, and the second process executes tasks in the background.¶

It'll tell you when it's done,¶

And you can see it — you feel at ease. You can see in a little strip on the right what it's doing and what it needs your approval for. Meanwhile on the left it keeps chatting with you, and you talk often. We have this thing: a queue concept. Like Codex — when you send tasks there's a queue, or I cut straight in and say I want to change this, the direction has changed.¶

Direction changed.¶

And Codex — when you use it, honestly it's already very good, very impressive, you can make changes. But when you change — every time you decide, no more queue, I want to change, I'll send directly — you send it and wait a long time.¶

That's changed now — now you can keep talking.¶

You can, but it takes a long time — every time you change that process, it thinks for a long time with no output at all. Whereas Muse — __ — Muse is often just very comfortable: you never stop, it keeps talking with you.¶

Still one built for C-end, ordinary people.¶

But this experience matters enormously. Many people only focus on the interface experience, the interaction experience, and the information experience — the core is how to build that feeling of comfort, making people feel you're truly helping them, and truly solved their problem.¶

Well then it's exposed — China has too few AI product managers.¶

China has no AI product managers.¶

Hahaha, none.¶

So — if there's a Chinese AI product manager listening to our podcast, hurry up and find a way to experience Muse.¶

You can also push back in the comments,¶

But I think — let me ask back — in the pre-AI era, did China even have product managers?¶

Well, some,¶

That's way too hot a take.¶

Don't, don't—¶

There are some, there are. But honestly they're old-school product managers, nothing to do with AI, and they genuinely don't understand AI.¶

Great, we're getting flamed again—¶

It's fine, it's fine.¶

And I feel like I've become the prophet again. In the first half of this year, in a talk I gave, there was a slide — remember? — saying for AI products in 2027, if you still make users deal with Keys, MCP, Skills and these concepts, you definitely won't scale. And Muse is already entering that phase in Q3-Q4 of this year.¶

They're working hard to downplay all these concepts—¶

Ordinary people don't need to understand any of this.¶

And it made a lot of features — like the Ideas we just mentioned, and something called the Feed, an info stream auto-generated from your content,¶

The Feed, the info stream, the Feed stream.¶

And you can edit the Prompt yourself.¶

You know what this is? It's erasing another thing — erasing the Session. All the current Claude Code and Codex are — everything, going back to ChatGPT — you have to start a Session first.¶

But think about it — if you had an assistant, how would you split your chats with them into Sessions? Do you talk to your wife that way? Definitely not. Everything you tell a person is told to one person, and then this automation layer organizes and distills what the two of you communicated.¶

That's Kabi — we said in the group that day, he asked what the biggest difference is: before it was Session Based, now it's become Role Based.¶

Sessions are purely built for work scenarios — in daily life you'd never need Sessions.¶

Doubao is two versions ahead,¶

Hahaha,¶

Our Cola has actually always been a single Session, and we keep getting users migrating over from other Agents who don't get it.¶

A lot of them, yeah.¶

But if you're not using it for coding, one Session is completely enough. The technology should be responsible for the complicated stuff — people shouldn't have to deal with any of it.¶

C-end products should be Robust.¶

Right — so by Guizang's standard this is also AI product manager capability, and this is a huge (gap).¶

Have you noticed — C-end users, ordinary C-end users who don't care about AI — how their real needs should be solved with AI,¶

And when C-end users open a Chatbot or an Agent, half of them don't know what to do — the Ideas and the Feed are—¶

Basically a guide.¶

Take our Xinji feature — we recently added a recommendation to it, based on your long-term memory, finding you really high-quality videos and books. That feature got great reviews. When users open the Agent and don't know what to do, they take a look — it matches my interests, and it's not the old filter-bubble thing. It gives you a starting point — it fits human nature.¶

And that Feed — isn't it basically an AI version of WeChat Moments, plus some things you care about, hahaha.¶

And there's one most important thing — Agents should have integrated long ago, the simplest thing: To-Dos. A To-Do List, To-Do List goals — we have goals, a shopping mode, right? What Muse has is: there's a long-term — a goal shared with you — with a To-Do List on that page. The To-Do List mainly does a few things: you set a long-term goal, and it keeps remembering those long-term goals. And what can you do? You can check the boxes, which means that thing—¶

People kept telling me what's wrong with Doubao? Memory — you tell it something and it nags you about it every day, long after what you're doing has nothing to do with it. I think it conflated things: sometimes it's goals, sometimes it's informational stuff,¶

Specific things—¶

Sometimes it's what you care about this week and not next week. But it uses a To-Do, or To-Do-app, approach: you check the box and it stops caring about that thing.¶

But like Yubo said — would ordinary people actually use a lot of To-Do stuff? Or are calendar and To-Do tools too elitist—¶

Actually, the month before last I talked with a former Google Calendar product manager. When they built Google Calendar they did research and found Microsoft's earlier calendars were complicated and hard to use. In their research they found what the broad American public actually uses most is a sticky note stuck on the fridge. But the need is real — there just wasn't a good product before. So they made the calendar very simple — and it's actually quite mainstream.¶

And I think people definitely have To-Do needs — life is full of things. The question is who can make this as simple as a note stuck on the fridge.¶

Fair point.¶

I think this experience can only be made smooth in the end with AI's help. Without AI, purely with programs — because programs can't solve these generalization problems.¶

It's rigid,¶

A lot of what's in To-Dos and calendars are generalization problems.¶

Actually, filling up your calendar isn't something that many people need. Our calendars are probably barely used. Even Americans probably don't have that many things to schedule — investors and such have more.¶

Insurance payment due today, taxes due tomorrow, right? Things like that.¶

But To-Dos are different — To-Do means, like, what you just mentioned: I have a pile of tasks, some with Due Dates — that's what a lot of people need.¶

The need factually exists, but the vast majority of people don't have the tools and methodology to solve it.¶

Right.¶

Whatever's simplest — AI, remember this for me,¶

And most needs don't need methodology anyway — I just want to note "pay the water bill," and you're going to give me a methodology?¶

And you know — say I don't have that interface, I purely interact with it, I have it do something for me, and tomorrow I forget.¶

No way to track it,¶

It has a UI. This Meta company is disgusting, you know? Not disgusting — impressive. Because these are features we're building internally too — you really need a visual representation of what's done and what's coming. This is a hard need — I've found people can't live without it.¶

Like when I used it to manage some projects — without that list you'll forget; chat alone doesn't work.¶

I used to think about managing my To-Do List through chat. I used TickTick, then quit TickTick. A year later I found I'd managed nothing — my AI managed nothing — so I went back to TickTick.¶

Including why Markdown editors are getting hot again — after interacting with AI, people need a result, a result they can Review, revisit. So look now at note apps and To-Do apps — TickTick and Typora update twice a week, it's absurd.¶

I previously set up my own OpenClaw with all the To-Dos, calendar and everything configured — I consider that the most basic infrastructure. From that angle, Meta deserved the big stock surge — they really did open up the next cycle's narrative at the right moment.¶

Pushed AI penetration another step forward,¶

Starting to cross the chasm.¶

And in some sense this is good for everyone,¶

Then by your logic it's a victory for AI product managers. Like, OpenClaw actually already existed — their product manager even said their 5 files come directly from OpenClaw, said the design is very sensible. Right — Soul, Memory, Identity, all of these taken straight from it.¶

Soul, Memory — makes sense.¶

And what's unusual this time: Muse is extremely open. Its interactions — if you're a geek and want to fold these things into your workflow, into a geek workflow, it's super smooth. And if you want to manage things, it even lets you manage the files on your cloud VM. There's a — in the App, the phone App, not the Mac App — there's an option; tap it and you get a file list, you can view every file, and you can download them.¶

I think that's normal, actually.¶

We can understand this as: from now on, everyone gets an Agent plus a computer — it really did assign you a computer,¶

But honestly this isn't just a product manager thing. Like — after OpenClaw came out, could a product manager really not think of making it a cloud computer that just stays running?¶

Of course they could,¶

Definitely could. The question is Meta, first, has the resources, and second, has the technology.¶

Can give it away free.¶

Has the people — and excellent product managers.¶

And think — it can have 500,000 daily actives all using these VMs at once, with basically unlimited tokens. You use it for a day,¶

I used it two or three days, chatting with it constantly, and used 2 or 3 percent of the free tokens,¶

And they have plenty of GPUs. This is mega-infrastructure. So Meta's Muse has pushed our AI bubble to a new peak. Imagine — 100 million daily actives, 100 million virtual computers,¶

Actually back then, when domestic players followed OpenClaw, they could have evolved in this direction to this stage — it's just that the cost—¶

It's just that many don't have the resources.¶

It needs a lot of things at once—¶

I don't think they have it — don't have the resources.¶

The resources Orange just mentioned — resources are multi-dimensional: do you have engineers, do you have product managers, do you have traffic, do you have, right, servers.¶

You need at least a cloud — and in China there are really just those few: Tencent, ByteDance, Alibaba — and none of the three did it properly,¶

Just those few.¶

Everyone went off competing in office tools instead,¶

Beautiful — and they even did internal retrospectives asking why they built it.¶

And most importantly—¶

And Doubao is doing layoffs, while Meta says this is the big direction,¶

After the layoffs they're stunned.¶

I just gave another round of talks inside a few big companies last week. I said Doubao originally built a product for hundreds of millions of people, then because Codex and WorkBuddy got hot, pivoted to build for 100 million people, and soon they'd find they had to turn back and build for hundreds of millions again — and then Muse face-slapped them within a single month.¶

Mainly I think what you need is a vendor with resources and a very determined leader to push this. Just days or months ago, what was Meta's news? "Alex is an amateur, Zuck is an amateur, making everyone generate data all day, everyone grumbling,"¶

Grumbling everywhere, people quitting, the model __.¶

And you spent all that money buying this — and look, Alex doesn't even look like a serious person. And Zuck pushes forward with total determination: any resource you want, you get it. That Turing Award winner, Yann LeCun — whatever you object to, just leave. Just leave.¶

But is Muse Alex's push?¶

Yes.¶

No, that was Zuck's push too—¶

And they actually finished it back in June, and spent these two months on compliance,¶

Kept polishing.¶

Privacy compliance, safety,¶

Because Meta — Meta's position is like ByteDance's position in the world. The moment you ship something, buddy, I'm watching you — you ship, I fine you; whatever you ship, I fine you.¶

And privacy here is extremely important — imagine, it can read all your Gmail. If you don't protect it well, that's a scandal.¶

Mainly, it can pay for you. It can complete the whole order flow — it can operate on Amazon and buy things straight to your doorstep.¶

Did it get banned?¶

Before, it could—¶

It seems — I saw in our weekly report — someone really did ban it, Amazon—¶

So Amazon banned it,¶

That drove it crazy — but crazy is good. The Shopify CEO—¶

Jumped on it right away—¶

His grin split his face.¶

Shopify stock jumped right away,¶

Grinning ear to ear — good that it went crazy.¶

Isn't this just Alibaba blocking Tencent?¶

Taobao blocking WeChat.¶

Taobao blocking WeChat,¶

And Doubao Phone getting blocked by everyone too,¶

Why ban it though — it places orders and pays you, and you still ban it? Why bother. No need.¶

Mainly Amazon doesn't rely on—¶

Plus it partnered with the other one—¶

Amazon isn't for C-end users — it mainly makes money from ads. If you affect my ad revenue, that's not OK. This is a really hard problem.¶

So this event — everyone saw a huge change. Look at Zuck: the cost of Muse definitely can't break even — give it to any company and they can't make the numbers work, can't sustain it. Only he can do it. But it gives everyone enormous imagination,¶

It'll definitely earn it back.¶

The imagination being: in five years, ten years, I'm the f-ing number-one entry point, I've got shopping and everything bundled all together—¶

Everything bundled,¶

Look what he leads with—¶

From entertainment to shopping, the metaverse, right? It even released VR glasses, right? Everything closed-loop.¶

What's the first thing he does? The wallet — he pushes that wallet with all his might, gets people onto that wallet.¶

Really is.¶

Heh — you've got to marvel at domestic companies.¶

See how excited everyone got talking about this — I genuinely feel we're at the inflection point of crossing the chasm.¶

Mainly there are too many inflection points — one every two weeks,¶

What? What did you say?¶

Two weeks — one inflection point every two weeks.¶

Hahahaha. Look at us too — the month before last we were still dumping on them,¶

Too many inflection points,¶

Now everyone's praising them,¶

This month we start raving.¶

__ — what a face,¶

From outside the AI circle, these four lunatics — no, not four lunatics, the whole AI—¶

Circle is all—¶

The whole AI circle is lunatics.¶

How did it change so fast—¶

What a face, __,¶

But I have to say, this Muse might be — like Zeng Ming said just now — the first one to reach C-end. I think this is the first.¶

OK, and let's also talk about their hardware — what do you all think of the Muse Charm? Brother Pan, don't you—¶

Tamagotchi, right?¶

I think over this past year, all kinds of companies in Shenzhen have made similar things,¶

The disc one—¶

Right, everyone's exploring. But like we just discussed, the hardware itself doesn't matter — what matters is having a big platform and ecosystem behind it, right? Your hardware paired with Muse behind it — because it connects to Muse, your thing becomes valuable,¶

And I think everyone is still trying to explore an entry point. We've always said the phone as a product form has been optimized by us over all these years to the point there's almost no room left to optimize — screens getting bigger and bigger, showing us more and more interactive information, stronger and stronger.¶

But have you noticed — with Agents, beyond things we need to look at visually, a lot of what we do is just tell the AI what we need and delegate the job to it. In that situation, the most basic experience is: that's why so many voice input methods are rising again on computers — in this situation, I don't want to pull out a phone. I take a small device, I just talk to it, and the job gets done.¶

Like, I'm going to Suzhou next week and want to book some train ticket. Why would I open my phone and tap tap tap for ten minutes of frantic operation? I just finish telling my assistant and go do something else,¶

Like a real personal assistant.¶

I go play games, I go cook, right? Why would I do all that tapping on a phone?¶

Right, and there's another problem: phones are extremely closed. Everything an Agent can do now — as we said, like Opus 5.5 — it works from first principles. It doesn't go through the interface; it goes through the most fundamental infrastructure the internet is built on, the things from the last century — FFmpeg, Git, right? Bash, command-line stuff like that. On the phone—¶

Great observation,¶

And phone performance is massively overprovisioned. This thing now is the new iPhone 20 (18) — an iPhone 18 Pro Max 18 Pro Max runs—¶

18 Pro.¶

It runs Cyberpunk 2077, games like that, no problem,¶

Easily,¶

But you can't execute a single command on it, and the Agent can't access anything.¶

And what good is a local model on its own? The local model can't do anything locally. Your whole architecture — Android is a bit better, but Android's root and ADB are pretty locked down now too — whether Android or iOS, Node, right — the scaffolding and components our code commonly uses—¶

Sandbox.¶

None of it can run.¶

So are you anti-on-device models?¶

No. I think having only an on-device model is completely useless.¶

An on-device model alone, no system — really, there's no good system.¶

Omarchy, hahaha, hah,¶

Apple just doesn't dare reinvent iOS,¶

Like — this thing's performance is now so strong there's no difference from a computer, and its screen is big, and in theory we only need voice input. In theory you'd want it to do exactly what you do on a computer — it could be completely identical to running Codex on a computer, in hardware cost and in every hardware interaction, no difference at all. It just can't run.¶

But I think this should still run in the cloud, because it's still a battery problem.¶

I think the cloud is definitely needed, but you can't have everything in the cloud,¶

Like voice and such—¶

And it has to interact with some of your local information — so this—¶

Its environment needs to be given your context — it has to.¶

Even say I download a file from the cloud — right now even file management is disgusting. I download an Office file and finding that file is a struggle,¶

Finder isn't even done well,¶

The most basic computer — it just isn't a computer, right? We need to go back—¶

And BTW, the new Siri is still an idiot,¶

You see, Apple — because its own Siri AI AI is so weak and its pace so slow, it just won't open up these things, keeps dragging down progress in this whole area.¶

Dragging down the industry,¶

Yesterday there was another library. Since iOS won't let you execute these things, these interpreters and such — there were two things. One: they released a library that goes through iOS's built-in browser component — that component's permissions are very broad, because it's Apple's own — to run Node.js, run the command line. And this caused a big stir in the community, because it means on my phone, now that I have a foldable, I could really use Codex or other things on it. But this thing—¶

You mean that WASM one.¶

Externally it can't get on the App Store,¶

Right — WASM, that should be Cloudflare, and that interpreter, the Python interpreter, their Workers are also built with this kind of thing—¶

And the other thing — what interpreter did they make? It works directly from .swift files, rendering the interface directly. Right now we render HTML — HTML is the computer's app format. This goes straight from .swift — since a single .swift file can define a scene, an app, and an interface by itself — it has AI write a .swift file, you press confirm, and the UI app is right there, compiled and run locally. This could immediately make the iPhone ecosystem just like a computer,¶

That's actually pretty important. Pretty important.¶

It's an AI OS, basically.¶

Last week I developed a cycling app, because Keep's voice announcements jump into ads, which is a bit disgusting, so I uninstalled it. I wrote my own cycling app to record my rides and at the same time announce kilometers and such.¶

I found that after developing it on the computer, when I went to test it, it hadn't requested the permission — so the whole ride it recorded zero speed. But on iPhone, if I could debug on device, I'd just change it to request a permission and be done.¶

This kind of app might be very small, very easy to develop — took me half a day to a day — it doesn't need to be a big engineering effort. The iPhone could totally be developed for by yourself. I think the future is exactly this kind of open system.¶

Coming back to the Muse hardware, though—¶

Muse hardware, let's continue. They also released something that isn't Muse—¶

Glasses, right?¶

AR — new AR. The reviews seem pretty positive,¶

I have this feeling — you can see how we—¶

Ray-Ban—¶

—have spent years mocking Zuck — all these launches, useful or useless, whatever. But you notice, for a big company like this,¶

That—¶

I just have money and resources, and I explore wildly in many different directions. Say I explore five directions and one is right — I've got the ticket to the next stage,¶

Like a VC, right?¶

Basically the same,¶

But the most core thing is having a founder.¶

True — if you had Tim Cook do it, he'd—¶

On the VR thing — think about it: after Apple shipped the first generation, there was never another. Look at their new VR — it weighs what, less than twenty or thirty grams—¶

Forty,¶

No, a hundred grams. A hundred.¶

That one is the AI—¶

The VR you mean—¶

The VR is only a hundred grams — Apple's is like two pounds, right?¶

Really small,¶

The hundred grams — they split off the audio and everything else, improving the experience, and it's 5K.¶

The battery brick holds more of the components,¶

You think Apple couldn't make it? I think they could — they just stopped.¶

Right, founder-driven matters enormously.¶

Only founders can not care about shareholders, not care about the stock price — "I'm doing it anyway." Those other professional managers — if the stock drops one point they're terrified. Then it's over — how do you innovate, right?¶

Shameless of us — we're at it again, hyping Zuck,¶

But this reminds me of Anthropic going public. Their equity structure: the co-founders control 50.1% of the shares — that's the controlling stake, that's the decision power — and the shares themselves, their actual economic interest is only about 2%, I think.¶

You still need to be the one calling the shots.¶

And their valuation is somehow twice OpenAI's.¶

Because its growth slope is steeper — people are looking at the future.¶

Its curve—¶

OpenAI had just clawed some ground back, and then this 5.5 came out and I think—¶

Everyone seems to agree 5.5 is the best since 4.6,¶

The best — look, writing got better too, right?¶

The best, so good, so good,¶

And it passed Fable — in just a few months it—¶

You've seen the American version of that sliding-rheostat drawing, right? The first one was normal — see, this one—¶

Dario's, that Dario one,¶

Now it's the Saint Liang version, this gritty Dario.¶

That double chin,¶

Finally got everything right,¶

Some people even joke — damn, did Claude buy sponsored reviews? I'm like, everyone is saying 5.5 is good,¶

But when I see even rivals saying it's good, it definitely wasn't bought,¶

So you see — it shows their, how do I put it, their base model is really strong. From Claude's angle looking at OpenAI — including 3D, including understanding getting stronger — within a single quarter they catch up or surpass.¶

There's also a recent take I've heard from some analysts: people say domestic pretraining is quite strong — actually no. Overseas is ahead in both pretraining and post-training — not just somewhat ahead, clearly a generation ahead. Domestic strength is in Infra — reaching a relatively decent level on cheap GPUs, with small resources. Currently—¶

The moment we have anything good we ship it immediately; they might hide theirs for a long time,¶

Just not visible — it's all underwater, we just don't know.¶

Hidden for half a year to a year — there might be a whole batch at once, a doubling of models. Then you can understand why they're so optimistic or so pessimistic — the model they use internally is completely different from what you use.¶

Different, and also—¶

Think — they could use Opus 5.5 four or five months ago.¶

Early this year, or late last year, a wave of X influencers — like AK, lots of them — joined Anthropic, and at the time everyone was guessing they must—¶

Have seen something,¶

Definitely.¶

Maybe RSI — their own self-improvement iteration speed.¶

Come here, let me show you a case,¶

Look at this case,¶

Send a carry to join me.¶

Generally, these top model companies keep internal models at least three months ahead before releasing them, because they need to do safety and all sorts of things.¶

Alignment.¶

I think their base models are half a year to a year ahead — the biggest base models,¶

Domestically, usually half a month to a month before release. K3's K3 actually finished training in the first half of the year too.¶

But these super-large models are still quite a struggle to train — alignment is a struggle too.¶

Back to the handheld device — what are the device's actual specs? Who among us even cared?¶

I did look, because everyone says it's like — I've never played the pet kind. The electronic pet, the Tamagotchi.¶

My daughter has several.¶

It's roughly like this, and then—¶

And voice input, video input and output — video input.¶

It has voice input,¶

But it's timed for Christmas — shipping for Christmas as gifts, for big volume sales,¶

A gift, exactly,¶

For big volume.¶

Because in America, Black Friday sales can account for a very, very high share of the whole year,¶

Very important.¶

And he hasn't published more data either — like whether it actually has independent audio input, or some chip or other.¶

It must have a microphone, right?¶

Right.¶

And it has a screen — does it have a camera?¶

Video — it has a camera, it has a camera, I saw it.¶

Front and rear cameras, a two-inch screen, and 5G.¶

Connected, right—¶

An eSIM,¶

By that count it can mostly, to some degree, work without the phone.¶

That's pretty impressive,¶

An eSIM — it can go without the phone. With an eSIM it can leave the phone behind,¶

I think it's great — freeing people from phone screens, back to the real world. I like it.¶

And look — say you've added this kind of hardware at home, I don't know if everyone has, like a Xiao AI speaker or such — turning off lights, turning on the AC, adjusting the temperature, it all feels great, so convenient. In the future it's all Agents, able to do things ten or a hundred times more complex than that. Imagine that experience — it's sci-fi.¶

For ordinary people, being able to use a piece of hardware to, say, turn off a purifier or turn on a light — that really is rare.¶

Feels rare, yeah.¶

My home is exactly that scenario.¶

Now combine it with what Muse can do — you hold the button and say—¶

"I'm going to Shanghai tomorrow."¶

"Pay my taxes for me, and book my kid's class for tomorrow — I'm thinking three to five, and if not, then whatever time to whatever time."¶

"I'm heading out to drive — pre-adjust my Tesla's temperature."¶

"Start charging and pre-conditioning." And then that thing — picking a class for your kid — you open that __ mini program and look at a splash ad.¶

Haha, we'll have to edit this part out, so annoying.¶

It's fine, it's fine — look at a splash ad, shake your phone to decline the redirect to Pinduoduo, then sorry, then find the horribly written code, pick the date and see three-to-five has no class, and you start thinking, when do I go then? Four to six? Four to six has no class either — flip back, tomorrow then.¶

Why should we waste our lives on this stuff?¶

Think how much time you waste on these things — how much of it is actual operation?¶

My map app throws an ad at me every time I open it. Infuriating.¶

So this launch — Muse plus this Charm — gives everyone a beautiful vision of the future, the next ten years,¶

Performing miracles again.¶

Three years ago we said once AI arrived, you'd think it could do anything for you — everyone wanted to build AI assistants and such — and then until this year, everyone kind of—¶

Saw a bit of hope.¶

Saw the same thing,¶

And nobody thought it would be Meta, hahaha.¶

But the hardware itself — I don't have much to say. This past year there have already been plenty of—¶

Nothing in it. Nothing worth talking about.¶

Hahaha, nothing,¶

Of course.¶

You can buy a pile of them on Taobao,¶

Huaqiangbei stuff beats this by miles.¶

But it's the closed loop — it's now data plus Agent plus all its previous content, fully closed-loop.¶

Plus cloud?¶

It also has WhatsApp.¶

The whole ecosystem.¶

So the standard for whether software is good is that people will overcome all kinds of effort to use it,¶

Yes.¶

You notice, once this got hot — first people see it high on the rankings and the first thing they want is to see what it actually is; then others use it and say it's amazing, and lots of people go to the dark web, right, or Xianyu, these places, hunting for ways to use it. And once it's out you find the whole feed is people posting invite codes — they want to earn more of it. You couldn't use up ten invites anyway, but they want more. That shows people value it.¶

You know how with some other model you don't care about, they give away tokens — however much they give you, you don't want it, right?¶

Exactly.¶

So how was this segment? Anything else we want to cover?¶

Chats like today feel so good.¶

Good — after this—¶

Two big topics left.¶

Exciting.¶

And now let's quickly run through some news. First — some of it we've already covered, so I'll only do the uncovered ones.¶

The first is a really important one: Harvey. Harvey published an important number — for every dollar they collected from users, they used to pay one and a half dollars to the model companies. But recently, they trained a model based on Kimi's K3 and pulled their gross margin back to positive — profitable for the first time.¶

On this I think I need to say something about deploying open-source models. And recently — well, lots of Silicon Valley companies say they're fine-tuning their own models, including Fireworks pushing this, right? My view has never changed: only at Harvey's level—¶

Can you pull it off—¶

—can you swap in an open-source model to cut costs. Because the vast majority of companies haven't even found PMF. Your whole business — the business loop isn't closed yet. If you're not using SOTA models — well, Harvey has in fact already dominated this market. I swap models, I evaluate it myself, it doesn't affect my whole business loop — it's already stable — and I can substantially raise my gross margin.¶

But a lot of companies are still saying, "Today Opus 5.5 came out — if I don't use Opus 5.5, I'll be killed by my competitors." You have no chance of that. So this only makes sense for the head of the market, the stable ones that found PMF. It's not something that applies to every startup,¶

So this kind of vertical — I think it depends on the vertical. The legal vertical is very—¶

Has a moat.¶

Very strong moat — they're relatively closed.¶

So if you do a vertical — I think right now, these everyday verticals, like video generation and such, probably don't have that big a moat. But law, and say construction, say medicine, right? And medicine facing doctors, not medicine for C-end users.¶

The kind where someone bears liability.¶

Those are the ones that can possibly last long-term, and aren't affected much by models.¶

So, first, swapping in open-source models; second, training your own small models — this is not a universal truth. I think the easiest misconception is: the news breaks and everyone thinks "we all have to do this, every company has to do it, this is the mainstream." Actually no. It takes your specific business model, your specific market, your specific scenario needing it. We need the ability to tell these apart, or at least to think about it.¶

It also shows you don't need to always use frontier models, and that the overall model bar is very high now. I think this is actually the application company's opportunity — a cheap model might satisfy user needs in your scenario, and that's when you have profit margin. If you keep chasing frontier models, right—¶

The model companies eat it all.¶

What's it Shaonan and them call it — this "lateral transfer of withered technology (thinking),"¶

Exactly, great example.¶

Switch is always the weakest of the three—¶

That's exactly where it sells the value of its games.¶

I.e., is your moat something the frontier model gives you, or something your own product gives you?¶

Exactly. That's also why I think, about what Brother Pan said on next year's Agents — if you keep putting a model-switcher list in there, that may not work anymore. We've been cutting back too — recently we just give it an Auto model, we manage it, you don't have to.¶

Let me offer a counterexample — Yubo does exactly the opposite. Yubo lets you pick freely among all the models.¶

I think that's phase-specific — probably next month a wave comes and it all gets swapped out,¶

But I think the core question is: can you let users perceive a problem with your model?¶

You have to tune it well,¶

But Yubo's positioning for everyone is: through my YouMind, you can use all the frontier models.¶

I think only the AI circle has that need — the general public doesn't have this need, but—¶

No such need — it's too expensive, way too expensive.¶

Cheap, good, no-brainer — Muse showed everyone the way,¶

Too expensive. Picking a course for my kid cost me 300 yuan, right?¶

Like, making that demo of yours — it actually costs several hundred dollars,¶

Wasn't last week's one super hot?¶

As an API,¶

I've noticed this really is slowly becoming infrastructure. Wasn't that the topic we covered last week?¶

Which one do you mean? Jev, right. More and more people are starting to build Jevs.¶

A couple days ago I joined a GeekPark evening livestream talking about models. And — look, like bamboo shoots after rain, a bunch of Jev-like models popped up, right? All trained on small-parameter models of their own,¶

Trained,¶

Or fine-tuned — and SiliconFlow also shipped three open-source ones.¶

Nice.¶

But my feeling is most people use these for the wrong things — still not right.¶

Still not right.¶

Including from your own perspective — didn't Orange say he'd carefully watch that video interview with Jev's founder—¶

That video,¶

The few typical scenarios they explored themselves — that direction is the right one. I named three directions: first, massive volume; second, high frequency; and the third — what was it? I kind of forgot. Massive volume, high frequency — forgot. The thing is, lots of people use Jev to process content, and that's exactly what it's not good at — things that need reasoning before you get a result are exactly what it's not good at. If it's massive volume and high frequency — ah, forgot,¶

Classification and reorganization-type things—¶

If you can't use at least two of these properties at the same time, you're probably using it wrong,¶

One-shot feedback.¶

If you only use one, you're probably using it wrong.¶

Actually the Jev paradigm means no Human in the Loop — full automation. Its future is intelligence that doesn't need humans in the middle. It's the complete opposite of Muse, or of earlier LLMs, which were built together with humans, for human experience. So that paradigm is purely about improving intelligence,¶

But I'm curious — it went from release to people actually using it in a very short time—¶

Which shows the pain point is really painful.¶

So what I said was massive volume, high frequency, low latency. If when you use it you're not using two of those three — only one of the three — then you're most likely using it wrong.¶

OK, and let's talk AMD. Pan, sadly for you, that trillion was one trillion US dollars in market cap,¶

Right — because look, since our very first episode I've been hyping Grok Bot, and Muse the same way. Once this takes off, the cloud-host concept should naturally rise, right? Because GPUs rise, and the CPU is the Agent's house, its infrastructure, while the GPU is the brain.¶

Actually this concept had a wave at the end of last year, but over the past six months I went and ignored it. But with this Grok Bot wave rising again, demand for cloud hosts will definitely surge — I just failed to translate that into action in the secondary market. Needs reflection,¶

Because Muse added another log to the fire — CPU, GPU, and flash storage will all rise again. I feel the whole industry is entering a new bubble period,¶

Hahaha, rising again,¶

So do you all think — maybe Apple's Mac Studio, MacBook might have to raise prices again next.¶

Definitely.¶

They even released that dedicated enterprise version.¶

Prices will rise again,¶

Anyway, if you haven't bought a Mac Studio, or you want a new computer — buy now.¶

Hahaha.¶

Hurry. Next stage — Apple has only two modes now: either right before the next WWDC, or they change the website any moment, raise prices any moment, any moment,¶

Just raise prices outright. And actually price hikes aren't the scary part — what we lack now is stock. What we fear is no stock,¶

Then I feel like I need to upgrade my laptop too — three years old already.¶

Hurry, hurry, switch.¶

No stock.¶

Hurry.¶

One thing is hardware might be out of stock. Look, Apple's iPhone 18 Pro 18 Pro sold really well this time — after raising the price, sales were up 30% over last year. Why?¶

That heavy thing,¶

Because Android is even more expensive. People found Android flagships—¶

No comparison, no harm.¶

Apple suddenly has value for money,¶

Apple has value for money now.¶

You know this—¶

Because Apple's supply chain—¶

You know Xiaomi's 17 Ultra last year, the Leica one, with four huge cameras sticking out the back? That sold for eight thousand, or starting at seven-something thousand. And this year's 18 Pro Max — not even the Ultra—¶

18 Pro,¶

Starts at 6,500, add a bit of memory and it's seven thousand—¶

So when people think about cyclically replacing their phones, computers, electronic devices — do it early, because if you order now you might still have to pre-order,¶

And then use it for years after switching.¶

And be thoughtful about it — buy more memory. Memory is expensive, but buy more anyway.¶

I just feel the whole industry is entering a new bubble period. Bubble is a neutral word first — not necessarily a bad thing — it drives all kinds of infrastructure investment, and on the model layer, people are earning more and more,¶

Including that absurd recent news: some cloud vendors don't even want to sell to you anymore — they say "I'm giving you no discount at all, because this year's sales target has already been exceeded, and I want to leave some room for next year." So no discounts at all.¶

But maybe the boss changes the KPI again soon.¶

Hard to say. I think right now — and it's not one company, it's many, all within the past week. I feel like: one, hardware — we may need to hurry and switch; two, on the model side, some models may carry a premium later — they cut prices, but there's a premium on the cloud side, on the API side.¶

For us — like, Opus 5.5 — for us there is a premium,¶

Yes, yes.¶

A big premium already.¶

Yes.¶

But in a sense, today's premium is, long-term, a benefit — an arbitrage window.¶

Yes.¶

The later you buy, the more, you know—¶

Yes.¶

Yes.¶

But I think many, many people's understanding of arbitrage still has problems — they still have the old software-and-apps mindset: I'll develop a piece of software, build an app, and sell it to others. I think that understanding, today, is completely wrong,¶

Agreed.¶

If before, a lot of people — like, if you build something to solve a problem for yourself, saving yourself time and such — I think that may still be somewhat useful to you. But among those listening there are probably a lot of listeners — if what you want is to create things today and sell them—¶

Very hard, I think,¶

Don't you remember you posted an image — since AI's recent development, isn't there that four-quadrant chart — games, and productivity—¶

The apps where people spend their time.¶

You notice—¶

Time and value,¶

Look, it fits perfectly — most people use AI models to solve their own efficiency problems, so they develop all kinds of tools for that. Everything else gets squeezed. They built it themselves — why would they buy your software?¶

So for arbitrage, people should think about other things. Get closer to value — like what Guizang said, using it to make the intro edits and such — that's more directly tied to value, because once I make it, it directly creates value.¶

Maybe people just haven't caught up, because the market hasn't been flooded by this yet. Look — traditional services charge you 50,000 yuan to make a video; with AI, something at the same level costs under ten dollars. That's arbitrage.¶

And there's another problem: the channels aren't smooth. Distribution channels for AI content are very clogged. If you want to distribute now, relying on the old channels is impossible. For startups — new startups wanting to distribute content, it's real now: if your founder and your co-founders don't go—¶

Proactively—¶

—proactively do marketing and build influence, and you rely on those old channels, then the money you spend has zero effect — and it can even backfire. Lots of recent examples of that.¶

The fact is, creation has grown tenfold, a hundredfold, a thousandfold.¶

So attention runs short—¶

But attention hasn't grown, consumption capacity hasn't changed, the number of consumers hasn't changed — so it's more crowded.¶

More intense.¶

So I think there are a few points. First: build services, not software. If a piece of software has no service, it has no value either.¶

Then the second point — the upside is that attention is so scarce now that once your thing gets adopted by one person, they won't switch. You can build a small service. Sure, this service has lots of competitors — this is worth people thinking about.¶

I've found that even good wine fears a deep alley — you can make something really excellent and, if you don't market and promote it, someone else snatches it away.¶

It's like that. Like, this week some other models shipped too — this Grok 4.7 4.7 — and barely anyone paid attention,¶

Xiaomi, StepFun,¶

StepFun too — maybe one sentence and move on. Nobody even wants to try them. Too many.¶

And lots of people Vibe Code. Now with Vibe Coding I've found a problem: once you finish coding — because coding is when you're most creative — the moment you have to distribute—¶

You don't want to anymore.¶

You lose the will. And I have this problem too — I'm already an extreme distribution person, and even I have this problem. And here's what happens: when everyone can build, everyone can make it, you face everyone's judgment.¶

Before, many people thought, "__ — I can't code, so I won't judge this thing, because I don't understand it." But now everyone thinks they could do it — everyone looks at your thing and thinks "isn't it just this—"¶

"Isn't it just that? Just this,"¶

Then your comment section gets more malice, and it makes you even more afraid to post. You have to overcome that, or find some very clever ways to defuse it.¶

You know why — it's because creation now has no friction,¶

None at all—¶

But many steps of the distribution process still have friction. But I think Computer Use, Browser Use, including Muse, Grok Bot, these things — didn't we cover this the last episode, or the one before?¶

We did,¶

These things gradually — if you can state the right intent — keep lowering the friction of distribution. That's where the arbitrage space is.¶

Right now the barrier is here, the barrier is here,¶

So does everyone now need to do product marketing and promotion?¶

Always have to,¶

Its importance went up, right? Like—¶

From now on, Trust, Reputation — the value of these things keeps rising, and going forward it will concentrate toward the top.¶

Yes.¶

Right — concentrate toward the top.¶

Why — look, why are we doing this podcast?¶

Why are we doing Next Token?¶

Why?¶

Because it's purely humans talking about this stuff, no AI Slop — its long tail, and brand going forward, is increasingly scarce. That's a clearly visible trend.¶

Including my own writing — I don't use AI to write anymore. Writing with AI is pointless; I write by hand. So what if I have typos? A bunch of people point out my typos — you should be happy, and I'm happy too, typos and all,¶

About what you just said — when I handwrite, people below say there are typos; when I write with AI, people below say it's AI-written. So what am I supposed to do?¶

And there's another recent news item — that strongest OPC guy, the one whose name starts with L,¶

And then—¶

He said his revenue grew this year — how much was it — his revenue passed ten million dollars,¶

Ten million dollars,¶

Just one person. Just one person. And that American — Musk pays him a hundred thousand dollars a week in creator payouts on Twitter,¶

The American one, right.¶

A week or two—¶

Look at Orange's WeChat official account — once you cross that 100K+ views line, afterwards anything you casually write goes out with thousands of views just like that.¶

You have volume — mainly it's influence, mainly recognition. People have a conditioned reflex — they see the avatar and want to click,¶

The future is a bit IP-like — every account is basically an IP. What people are really building is trust, and that's something technological progress won't eat. Competing on technology — compared with the models now—¶

So I feel like the way to build a long-term bond with readers or viewers is still to build your own product. Ideally a high-frequency one — best if it's high-frequency use, so it leaves an impression.¶

Or — through your taste, your personality — not saying good or bad taste, good or bad personality — your personality itself gets transmitted through what you say, through your content, through your product. And it can evolve but must stay consistent. And having a human touch matters — there's no shortage of things without one right now,¶

Way too many, too many.¶

No shortage — look how many "great men" there are in the comments,¶

And these AI comments are really interesting — at first they'd agree with your take; now a lot of the time they changed: they argue with you just to provoke you,¶

Argue with you, to provoke you.¶

Then they hope you reply, and bring them traffic. So the trend going forward is all about grabbing attention — grabbing the creator's, the channel owner's attention, grabbing other people's attention—¶

Yesterday I posted — didn't my de-AI-flavor Skill get updated? Then someone commented below, something like "worse than so-and-so, this thing is pure garbage." I went, hey __, what's going on here? Then within a while there were five or six of them, all similar content,¶

Trying to siphon traffic from you,¶

He wanted to provoke me into replying so he'd rank higher, so I posted below: if you people keep using these tactics on this product, I'm going to start — you know — on you.¶

People figure negative fame is still fame.¶

Hahahahahaha.¶

Ugh, this is getting a bit too absurd, honestly.¶

Haha, let's stop, let's not talk about this.¶

No point, no point, no point.¶

Nope, nope,¶

Everyone should still build their own Reputation. Anything you do will be remembered. If you're only thinking about a short-term "I went briefly viral," then long-term that's a stain on your record. Not worth it.¶

Right,¶

Following up on what Xiangyang said — you should still have your own product. But a product — what is it? It doesn't have to be an App. Right, right,¶

This podcast of ours is a product too,¶

Next Token itself, the podcast, is a product,¶

Because we're releasing our new version next month — we're treating the launch event itself as a product,¶

You see, I think everything can be treated as a product — Zhang Yiming says treat the company as a product. Your articles are your product, apps are your product — everything can be your work,¶

As long as you build one thing that creates a connection between you and your audience, that's it.¶

Discovery, seeking, creation, distribution — a closed loop: that's a product. As long as it fits the flow we have now, it's a product — that's product thinking today.¶

I said last week I developed that Obsidian plugin, right? I found it so satisfying, because listing it is super simple, and two weeks later—¶

Open source, no review—¶

— it already had over two thousand downloads, and the user group is super active, chatting away—¶

That group is great,¶

But Obsidian is still a bit geeky, I think—¶

Definitely geeky,¶

But I think here's the thing: as an individual, as an OPC, you can't be thinking "I want to go big" — we're not startups,¶

Right.¶

Or small companies — you keep saying you want to wrestle with Meta: "Meta made Muse, so I'll make one too." But the question is, do you have that much money, those VMs, tokens to give away?¶

Can you lock in twenty thousand people — didn't Orange mention it before, what's it called?¶

A thousand what-do-you-call-it true fans and you can live worry-free—¶

That's the one Kevin Kelly said.¶

Insisting on building an AI WeChat or something like that is actually pointless,¶

No point,¶

But like, if you build that plugin, it has very active communication, the positive feedback is strong — because it's a community of real humans, which feels better than stars or likes or any of that.¶

Exactly like this — this is the connection,¶

All of it counts,¶

Maybe this is a form of entertainment for the future too. That's it for today—¶

That's it, that's it,¶

That's it,¶

OK, so let's say bye to everyone.¶

Bye bye bye,¶

Bye¶

Bye¶