# \#003 Unpacking Jev's new paradigm: do agents need an ultrawide?

Hosts of this episode: Guizang, Xiangyang Qiaomu, Orange, and Pan YANG.

> This transcript was text-cleaned by a user; speaker and time anchors come from Tencent Meeting transcription and have not yet been verified sentence by sentence against the audio.

[Read on the website](https://nexttoken.tv/weekly/003/transcript)

## 01 / Show intro

**Intro narrator：**

Hello everyone, and welcome to Next Token | Beyond Tokens. Next Token | Beyond Tokens is a video podcast covering AI technology, products, and their real-world impact, also available as audio. Our flagship show, Next Token Weekly, is a weekly roundtable with Pan YANG, Guizang, Orange, and Xiangyang Qiaomu, sharing observations, hands-on practice, and different judgments. Not just predicting the next token.

## 02 / Opening from the four hosts

**Guizang：**

Hello everyone, welcome to the recording of episode three of Next Token. I'm Guizang.

**Xiangyang Qiaomu：**

Hello everyone, I'm Xiangyang Qiaomu.

**Orange：**

Hello everyone, I'm Orange.

**Pan YANG：**

Hello everyone, I'm Pan YANG.

## 03 / Jev: no chit-chat, just structured judgments

**Guizang：**

Let's get started with this week's Next Token. First up — as usual, we'll follow last week's news and chat until we run out of things to say. The first one is, I think, the hottest new-new model of last week: the TypeSafe AI's Jev model. It's quite different from today's large language models, and its advantage is pretty obvious — you can treat it as a general-purpose classifier, and what it outputs is not what today's LLMs output, like chat or paragraphs of text. It only outputs things like program results. For example, you ask it to pick from a few things, it gives you a choice; you ask it to score something, it gives you a corresponding score; you ask it for a yes-or-no judgment, it gives you that. So it's extremely fast, it can process things in parallel, and it's very cheap. The official price right now is $0.042/MTok for input, and output is free, so for batch tasks the cost is much, much lower. Everyone has seen it — whether you've actually used it, who knows, but everyone should have seen it! What are your thoughts?

**Orange：**

Right, yes, I just used it. I had it go into my Cola and look at my 300-plus memories, score them, and see which ones it thought were important enough to keep and which could be dropped. It figured about half of the trivial stuff could be deleted — it seems to prefer abstract things — and it organized about 18 important ones for me. Yeah, pretty interesting.

**Guizang：**

It feels like it speeds things up for LLMs and for agents in a lot of ways. I saw another case where someone had the model pick Skills for them. Because they had a lot of Skills installed, it used to be slow to search through them and it ate up a lot of memory, and now they just have the model pick directly, and it returns results instantly.

## 04 / WeChat friend labels and high-frequency judgments

**Pan YANG：**

I have a case. I exported all my WeChat\-related data earlier, and there are 10,000 friends in there. Previously, for a feature, I had it read my chat history with friends, look at our shared groups, and tag these people. I have 20-plus tags myself — things like, this one's a CTO, this one's a founder, this one does such-and-such. I tried doing this with my local 27B or 35B small models before. First, it was really slow, and second, it was actually pretty inaccurate. What I'm trying now is having it do the tagging work for me. I proposed an approach: I said, after you finish, give me an enumeration or something. It said don't do it that way. I used GPT-6\-6 for this, and GPT-6\-6 said, here's what you do: have it evaluate each of your tags one by one, and compute whether that tag is yes or no. It's not three modes — one is yes/no,

**Guizang：**

A multiple choice.

**Pan YANG：**

One is enumeration, one is probability. So it told me to do it: for each tag, return a yes or no. Because tags are multi-select — if you ask it directly to multi-select, it can't do it. That's one scenario. Another scenario: my friends list was full, and I've always wanted a feature where an LLM suggests who I should delete. That's another massive-data problem, and this Jev is actually a great fit for massive, high-frequency scenarios. Because yesterday someone mentioned some scenarios, and I think they said it saves tokens in the middle — but in a lot of scenarios, the examples people show nowadays are like: here's a spot that's 1% of the workload, I optimize it by another 90%, and you only got a 0.9% overall gain. If you don't apply it to something massive and high-frequency, the gains just aren't high enough. So I just had it run through all 10,000 of my friends, gave it some rules, and had it tell me who it suggests deleting — the return is also a probability. And another one: I had it scan all my daily chat conversations — I probably chat with dozens, a hundred people a day — and find the ones where, after chatting with me, there's like a TODO or a highlight, and return and filter those out for me. This one is still running; part of the results have come out. From what I've seen so far, the speed and the results are decent. Later I'll do a comparison with another small model, and we can put out some numbers.

**Xiangyang Qiaomu：**

Let me say something.

**Orange：**

These 10,000 — 10,000, how long does it take to run, roughly?

**Pan YANG：**

Mine hasn't finished yet. 10,000 — that's because, because you can call it concurrently,

**Orange：**

Right.

**Pan YANG：**

It's actually fast. You don't have to call it serially. Right, each individual call — ignoring network conditions — is a few hundred milliseconds, roughly, that's the level. You just turn on concurrency. Anyway, I don't have any ordering dependencies.

**Xiangyang Qiaomu：**

Right, after I saw it I applied right away — didn't expect it to come through that fast, it was approved the very next day. When I tried it back then I didn't think of a particularly good use case; I just got it working. Today I saw some cases online that I think are pretty interesting. For example, someone made an LLM version of Ctrl+F?

## 05 / Semantic search and database queries

**Xiangyang Qiaomu：**

We all know Ctrl+F is exact matching — I search for a word and it checks the page or the content for that keyword, exact match only. He turned it into, essentially, an LLM's fuzzy matching: you say "happy", and even if it's not an exact match, it can still find content with a similar meaning. And someone else turned GPT into a database query function — like, "I want to know who didn't come to the office today" — and it can query the records in that database-like setup and return the answer directly. Before, you might have had to query quite a few tables in a database. I think the application scenarios are pretty interesting, but from what I've seen so far,

**Pan YANG：**

This is basically a function with generalization ability!

**Xiangyang Qiaomu：**

Exactly!

**Guizang：**

A general-purpose classifier.

**Xiangyang Qiaomu：**

And I feel like right now, the things that clearly work — people are still exploring, but intuitively, Computer Use is definitely something this can speed up. Because it just reads the DOM tree to decide where to click. As for the rest, I feel people do have some doubts. For example — the architecture isn't public, and there's no good benchmark for its intelligence level — could it judge things wrong? When you have it make this many judgments, how accurate are the results actually? People do have those concerns. That's what I've observed.

## 06 / "No hallucinations" doesn't mean correct judgments

**Pan YANG：**

Right, there's a point in their marketing here — they say it has no hallucinations — and there's a bit of a bait-and-switch in there. What they mean is: when we normally use small models, for example if you have one output a strongly-typed result, output JSON, it might output a wrong format, and the program itself, when you build the harness, needs protection — like retries and such. What they're saying now is: I guarantee the returned result definitely won't need a retry. Whether it's actually correct in the middle — hard to say. There are a lot of people building benchmarks to compare now, right. And one more thing I saw: lots of people in the community saying, "that's nothing, we can reproduce it." From what I've seen in the community so far, nothing reproduced has turned out better — not in quality, not in speed. The ones that claimed to be reproductions eventually all got called out. But on the other hand, I think open-source replicas will come fast.

**Xiangyang Qiaomu：**

There are already 6 projects today. Let me see — someone said there are 6 open-source projects, all replicating this.

**Pan YANG：**

But a lot of them don't seem to work!

**Xiangyang Qiaomu：**

For example.

**Orange：**

The model — this model really does depend on raw intelligence. It seems to have a simple benchmark, at roughly Sol's level, if I remember right.

**Pan YANG：**

It can reach Sol's level.

**Orange：**

Between Terra and Sol, I think. It's pretty smart.

**Pan YANG：**

That said, now that OpenAI and Anthropic have proven the pattern works, they could also build one of these little models inside their own systems to do this!

**Orange：**

Right, and it should be easy to train!

**Pan YANG：**

Right.

**Guizang：**

They said they'd open-source it — at the time they said they'd open-source it later.

**Pan YANG：**

Will they open-source it?

**Guizang：**

They said yes. I remember the founder mentioned it in passing. They also wrote up their own limitations. They said, first, multi-hop reasoning and complex coreference are weak — the parts our Thinking models are good at, this is all pretty bad at. Like, if you need it to reach a result that takes two hops, two steps of thinking, working through a few layers of relations to draw a conclusion — it can't really do that.

**Xiangyang Qiaomu：**

System 1, right. Intuition — just give the answer directly.

**Guizang：**

Right,

**Pan YANG：**

Right,

**Guizang：**

And then.

**Pan YANG：**

So it pairs really well with other LLMs,

**Guizang：**

Right, they're complementary. And the old problem: unreliable at numeric computation, counting, and date handling — you still write a program for that,

**Xiangyang Qiaomu：**

Still have to call tools!

**Guizang：**

Right, call tools. And it tends to understand things literally — it can't understand metaphor. The "happy" you mentioned, it can understand that. But once there's a complex metaphor, it can't understand it. Slightly more complex, and it probably can't.

**Pan YANG：**

Have you noticed that when people think about these issues, we're kind of back to — remember 2023, when everyone had LLMs do arithmetic problems and count things? Nobody asks SOTA models to do that today. But today this new little model comes out, and people immediately go "this doesn't work, that doesn't work" — when your use case may not even fit its positioning. You're forcing a use case that doesn't match what it's built for,

**Guizang：**

That's what they wrote themselves,

**Pan YANG：**

Right.

**Guizang：**

It's the parts they themselves found problematic in their own testing. There's also something about context — with irrelevant context, accuracy drops. If you pre-filter and clean the data before you have it make judgments, accuracy may be a bit higher. I still think it's the problem from before: it does have accuracy issues, and it does have a hallucination problem — it's just that their "hallucination" isn't the kind we mean,

**Pan YANG：**

Right.

**Guizang：**

And alignment, or prompt-injection risk is fairly big — presumably they didn't do much work on that, didn't do much processing there.

**Pan YANG：**

Right, I have a question. This model itself is pretty simple in function. If they keep going with it, what's the moat? What else can you even build on top of something like this? I honestly haven't figured that out.

**Guizang：**

I think multimodal is a really important thing,

**Pan YANG：**

Right. Agreed!

**Guizang：**

For data retrieval and — a lot of what we do now goes through retrieval and recognition. The diversity and accuracy of its inputs, the accuracy of its outputs — I think all of that matters. Especially if you can solve fast matching and processing for audio, video, and images — which are extremely token-hungry and time-consuming — the prospects are big.

**Pan YANG：**

You mean bringing in multimodality, next, down the road,

**Guizang：**

Right.

**Pan YANG：**

Right. Come to think of it, there's something that's been stuck in my head all week — borrowing from some of Liang Wenfeng's earlier remarks, where he said something like, we won't do multimodal, we think language models are powerful enough. Everything that's come out since has basically slapped that down.

**Guizang：**

You can't skip it. Skip it and no one uses you — and the money still has to be made.

**Pan YANG：**

Right.

**Xiangyang Qiaomu：**

I also saw a demo of controlling all kinds of software on a computer by voice. Before you even finish speaking, it opens it for you — like "open Notepad", you're still talking and it's already open. The response speed is very fast. I never figured out how it converts that into driving software and driving the interface.

**Pan YANG：**

In there — there are two models working together, with a realtime one,

**Guizang：**

Right.

**Pan YANG：**

Something essentially like Computer Use — it's just speeding up Computer Use. Realtime computer operation has been around for a while.

**Guizang：**

All those tasks are probably paired with a program, or with an LLM, or with other models.

**Orange：**

The thing about Computer Use is it's essentially a finite set: buttons — of all those buttons, which one to click. It just becomes a conversion like that. The model supports 255 options, so for open-ended sets it doesn't do well. They have a fan site collecting all kinds of uses called Jevable. There's a demo on it — everyone knows there's an animation type called Live2D, where you have parameters that define a character's motions and expressions. Live2D — this model can take one sentence — like "you smile" — and immediately generate a bunch of parameters, making the Live2D character move.

**Pan YANG：**

So fast. The website,

**Orange：**

Right,

**Pan YANG：**

Found it — Jevable.com.com!

**Orange：**

Then I also had this model judge the 10 most valuable apps on that site. It made the call,

**Pan YANG：**

Love it.

**Orange：**

It said the most valuable first one is helping it set thresholds. The model has a threshold thing, right — 1-to-10 scale or 1-to-3. You feed the data to the model, have it run a threshold by itself first, and then its judgments get really accurate.

**Pan YANG：**

Isn't that like playing Heroes of Might and Magic? What's that genre called — the grid-walking ones, with just a few options,

**Guizang：**

Turn-based tactics.

**Pan YANG：**

Right,

**Orange：**

Tetris — his second pick was Tetris, since the falling blocks are real-time probabilities.

**Pan YANG：**

That's such a good fit.

**Orange：**

It played 66 lines and never died. Anyway, it's just getting started. I understand this model surely has plenty of problems, including that officially it hasn't gone through RLHF, so it won't have those biases. But there are some problems — like if you ask it certain slightly sensitive questions touching on different things — it might have a hard time, might not know, so it'll run into those issues too.

**Guizang：**

It always has to give you an answer.

**Pan YANG：**

Okay.

**Orange：**

Right, so.

**Pan YANG：**

The site Orange recommends is called Jevable.com.com, right? Jevable.com.com.

**Orange：**

Right — they swapped Lovable for Jevable.

**Pan YANG：**

Everyone can go try it out.

## 07 / Using the model to filter covert ads

**Guizang：**

Another case: someone built a tool for blocking covert ads while you scroll Twitter or Weibo. Overt ads — can't you just block anything with an AD label at the start? This thing can block ads in real time as you scroll: each time you feed it one tweet, the moment it loads, it blocks it for you. There's also one that blocks soft ads based on subtitles. That especially plagues us in China — there's no requirement to mark ads on screen, so a lot of the time you get ambushed and have to skip manually. It gives you an option directly,

**Orange：**

You give it an article and it can also pick out the AI-slop words in it. It feels like a really fast little tool an LLM can use. I think the positioning is great: just treat it as a System One Model.

**Pan YANG：**

Right, like Qiaomu said — it's a function with generalization ability,

**Orange：**

Exactly. A smart if-else.

**Pan YANG：**

Because before, look — you either had a model, or you had a program. This swaps certain parts of a program for a model. Yesterday someone called it a "joint model" — I thought that metaphor was pretty interesting.

**Guizang：**

Pretty interesting indeed.

**Xiangyang Qiaomu：**

Why is everyone chasing this model so feverishly? So much of it is people desperately hunting for scenarios to use it in. I don't know what its breakout point is, or what idea it hit for people. Besides being fast — and I think speed really matters — what else?

**Guizang：**

I think it's one thing,

**Pan YANG：**

Cheap.

**Guizang：**

Once you get swept up in the wave, a lot of people naturally follow the herd and go try it. Another thing: the concept is genuinely novel. AI hasn't had something in a long while that everyone can understand, is new, and you can pick up quickly. And another: it's fast, and when it's fast, it's easy to build demos. That gives it a viral advantage — it's easy to make those demos that amaze everyone, and the spread of demos further amplifies the herd-mentality problem.

**Xiangyang Qiaomu：**

Agreed!

**Guizang：**

Those two numbers sitting right there, and it—

**Pan YANG：**

100 times faster, 100 times cheaper — that's 10,000 times!

**Orange：**

He added another concept: since he was one of the GPT authors before, he says this model might have the possibility of AGI. In his promo video—

**Guizang：**

Every time I watch his video—

**Orange：**

He asks, why build models at all.

**Xiangyang Qiaomu：**

Reaching AGI with just these three types?

**Orange：**

You've got a novel now — like the 42 in The Hitchhiker's Guide to the Galaxy. 42 is the ultimate answer. Actually, that could work.

**Guizang：**

42\.

**Pan YANG：**

Look at what's written on their site: 193.6x faster, and 444.6x cheaper.

**Orange：**

As long as you can convert it into a question,

**Pan YANG：**

Right.

**Orange：**

Like, should rates go up or should they go down — give it a 0-1-2, and it can pick out a number.

## 08 / Plugging the new model into your existing agents

**Pan YANG：**

Here's what I did: I went straight to my GPT-6\-6 and said, TypeSafe AI just released a new model — first do your homework on it, go read and understand it yourself, since it's not the latest knowledge. Then I said, write it up as a doc and put it in a project. I just started working inside that project. I started discussing scenarios with it — what models to make, prototypes — just let GPT-6\-6 run with it.

**Guizang：**

The main thing is that it depends heavily on other things; its standalone use cases are pretty limited. And once you assemble it with today's agent models, integrating this model into an agent is actually fairly complicated, because its request types and patterns are completely at odds with today's LLMs. If you want to add it to the existing AI stack, it's still a bit of a hassle. So a lot of the time people just build new websites and little tools — there's no quick way to plug it into the AI they already have.

**Pan YANG：**

Let's see how long it takes for a real open-source replica to come out — see which episode of ours it shows up in.

**Orange：**

It can be replicated even without open-sourcing. Honestly the principle is just next token — it's simple, I think. And the author's core belief is that this thing doesn't need to go through language-based alignment — it hasn't been influenced by humans,

**Guizang：**

More objective,

**Orange：**

A relatively objective probability model.

**Guizang：**

Another interesting case: someone drew a grid for it — maybe 24×24 or 30×30 — and had it predict in parallel what color every pixel should be, achieving something like Stable Diffusion or diffusion-model-style image generation. It can actually draw some pretty decent stuff. I think that represents another kind of possibility.

**Pan YANG：**

It shows the essence of intelligence is the same,

**Orange：**

Right, you give it—

**Pan YANG：**

It's all probability.

**Orange：**

Give it the 26 letters and have it choose, and it can speak. The key is still the level of intelligence.

**Guizang：**

I'm stammering a bit. And what was the other one?

**Pan YANG：**

Orange just mentioned a lot of things — I bet a lot of people haven't actually read its official site carefully. A lot of what Orange said is on the site; there's a huge amount of information there.

**Orange：**

Yeah, I watched the official video twice; he explains it pretty clearly, because training a GPT-6\-6 actually requires a—

**Guizang：**

There's also a doc.

**Orange：**

For example, a 5T-plus-parameter model needs however many cards, whereas this model is probably tiny.

**Guizang：**

True.

**Xiangyang Qiaomu：**

Isn't the author a core contributor to InstructGPT and GPT-4\-4? Feels like there's something there.

**Orange：**

Right.

**Guizang：**

The models coming out these past couple of days all have fairly small parameters. I looked — Devin trained a 0.5B one themselves. Devin, that is.

**Orange：**

Right,

**Guizang：**

They trained a 0.5B one, modified from Qwen 2.5-0.5B, and then CUA is—

**Orange：**

There it is.

**Xiangyang Qiaomu：**

Qwen — modified from that?

**Pan YANG：**

I don't believe any of this — no matter what they use, I don't believe in taking a small model and directly LoRA-ing it or whatever these days.

**Guizang：**

Right, I think it must already be different at the pretraining stage; there's no way it purely relies on an LLM finishing the job through training,

**Orange：**

They talked about that too,

**Guizang：**

They said it's mainly in the data — the core of all their work is probably in the data!

**Orange：**

With an LLM, you need to constrain it to output in a format. And once you add that constraint, the LLM's intelligence drops.

**Guizang：**

Anything else anyone wants to say?

**Pan YANG：**

We've been going quite a while already!

## 09 / Multimodal: why audio and video should be understood together

**Guizang：**

Okay, let's look at the rest of this week's news. Still going by models, hardware, and other stuff. On models: Qwen released a Qwen 3.8-Omni-Flash, an omni-modal model supporting native audio-video understanding, reasoning, and tools. Capabilities are close to the 3.8 Flash over here, 1 million context, and video output cost down 89% versus the previous generation. Feels like it referenced — or whatever — something similar to what we talked about last week, that cost-reduction approach the 3.8 Flash here used for video understanding?

**Pan YANG：**

What are these Omni models even for?

**Guizang：**

I think the core is, for example, automated ads — automated image creation for video ads — and things like arranging short-drama videos depend heavily on this kind of model. Especially the persimmon stuff!

**Pan YANG：**

Versus pure video models — because every single one of its functions, we can find a separate model for,

**Guizang：**

Right.

**Pan YANG：**

So what's the point of making it an Omni model?

**Guizang：**

Because you can't find separate models for it,

**Pan YANG：**

It's just that I—

**Guizang：**

There's no way to understand it holistically. You know, a short drama or a feature film — audio, video, every frame and cut — is a whole. If you understand them separately, converting everything into text annotations and feeding it to a smarter model, you lose a lot of information. At the annotation level — that is, the understanding level — it can be really bad. It's like using it is the same as not using it; you get a very mediocre result.

**Pan YANG：**

Someone told me it could be much better for annotation.

**Guizang：**

Annotation isn't just annotation either. If you build image-and-video agents, you really need this model.

**Orange：**

Like replicating hit videos — very well suited.

**Pan YANG：**

But its audio-video isn't a SOTA audio-video model.

**Guizang：**

It can't generate video,

**Orange：**

Can't generate, right,

**Guizang：**

It can only understand.

**Orange：**

It's understanding.

**Guizang：**

It can only understand.

**Pan YANG：**

Still — understanding and breaking down hit videos, isn't that just annotation?

**Orange：**

A cheap understanding model doesn't have that many uses,

**Guizang：**

You could think of it that way.

**Orange：**

Some scenarios do save quite a bit of money,

**Guizang：**

Absolutely necessary.

**Xiangyang Qiaomu：**

Very profitable scenarios, especially ads,

**Orange：**

Right.

**Xiangyang Qiaomu：**

Back when we were doing ads at TikTok, the most painful thing for creators was how to adapt hit videos, and what the whole team wanted to do was help everyone replicate hit videos quickly. This has real business value.

## 10 / Quantization and local small models

**Guizang：**

So Gemini remains what everyone really needs to this day — in this area there's almost no beating it. Second item: a company fine-tuned Qwen 27B and did a ternary quantization — 9x smaller than full-precision Qwen 27B, only 5.6 GB. They say it retains 98.2% of overall performance, and it's even stronger than before on multimodal and long-horizon task calling, and it can run locally on a single 5090. That's pretty important, because the 5090's problem is its VRAM is too low — a 27B simply couldn't fit before. With multiple cards, forget it. A single 5090 — and for this kind of inference, 5090 performance, I think, is sometimes better than Apple's MLX optimizations. Quite needed for local tasks — even a 32 GB Mac Studio could actually run it.

**Pan YANG：**

For small models, NVIDIA's cards are actually pretty efficient,

**Guizang：**

Higher,

**Pan YANG：**

Right, mainly because its memory is small, one,

**Guizang：**

Thank you.

**Pan YANG：**

Bandwidth a bit lower, right, so its compute is strong,

**Guizang：**

Right — after all, the CUDA advantage is there. Second item: MiMo's livestream, pretty interesting, spread pretty widely,

**Pan YANG：**

Luo Fuli is a post-00s? Only the post-90s— no, only post-00s would do this kind of thing,

**Guizang：**

Post-00s, I think. It takes young people,

**Pan YANG：**

Right!

**Guizang：**

They're livestreaming MiMo V2.6's reinforcement learning — the training process as a dashboard with all the metrics and dollar amounts: how much money each model has consumed so far, what the cost is, and they keep adding metrics and data. That's pretty valuable for training teams.

**Pan YANG：**

There's not much informational content in it — watching other people burn money feels good and it spreads well, but you can't really glean much from it.

**Orange：**

There is something to it. At least it lets you see complete progress, not just benchmarks. If you only see the final benchmark, the middle might be watered down, but this kind of continuous process is relatively more believable.

**Guizang：**

And they have a scrolling board on there with text telling you when they're going to terminate training, when they'll retrain, what went wrong with the last training run, and what they'll change for the next one. I think that's quite meaningful, quite interesting. And livestreaming does burn money at a very high rate,

**Pan YANG：**

Is that a first, too?

**Guizang：**

People never really had a sense of how much model training actually costs. When you watch it tick up $20-30 every single second, you realize it really does cost — and that's just compute, not counting staff and all the rest.

**Pan YANG：**

Put another way: it's not a SOTA model, so marketing has to carry it.

**Guizang：**

Next up is — they had this before, I can't recall the name — they released something called Arrow? Is that what it's called?

**Pan YANG：**

Arrow?

## 11 / SVG, icons, and design workflows

**Guizang：**

Arrow 2 2 — a model that draws SVG, extremely strong. Find its demo: a pelican riding a bicycle, 10,000 times better than what all current models draw, tons of detail. You watch it outline the feathers stroke by stroke, and the bicycle wheels are wild too. I think lots of models have never gotten this part right. Drawing icons — drawing the SVG icons we normally have in interfaces — even GPT-6\-6 or other models don't draw them well. Drawing icons is painful for designers too: you work in a tiny grid in Figma, tens of pixels by tens of pixels, tracing Bézier curves with the pen tool, or cutting and boolean-operating basic shapes. A skilled hand is fast at it, but very few people are good at icons. It's hard — 20 icons might take a whole day.

**Pan YANG：**

See, Astra is strong at drawing 3D — this is basically the 2D version,

**Guizang：**

Right.

**Pan YANG：**

I think this kind of thing goes fast — SOTA models just train the capability straight in. And also,

**Guizang：**

It can be done.

**Pan YANG：**

And — offering it in the form of a standalone model — few people would spin up a separate model just to do this. I think the best form would be offering it as an MCP or something like that.

**Guizang：**

I think it's fine — it can be packaged as an MCP. Like just now,

**Pan YANG：**

Right.

**Guizang：**

Jev also has those prompts where you provide how your model gets called,

**Pan YANG：**

Right.

**Guizang：**

So they — for the design field, lately a lot of AI design tools or Figma\-replacing design software have come out, and their demand is pretty big. Since the parameter count isn't large, the training cost probably isn't high either — there might be a bit of commercial value.

**Pan YANG：**

See, remember I told you — your PPT Skill is missing images, it's all text. With this thing paired up, it would be divine,

**Guizang：**

It does have images — it calls an icon library and such, but drawing richer things is more troublesome,

**Pan YANG：**

Right,

**Guizang：**

With it, definitely.

**Pan YANG：**

If you pair it with an MCP, wouldn't that be awesome,

**Guizang：**

Right, that'd be great!

**Xiangyang Qiaomu：**

And it's vector,

**Pan YANG：**

Right.

**Xiangyang Qiaomu：**

I think vector really matters, especially for print publishing.

**Guizang：**

Right — otherwise how do you scale it up?

**Xiangyang Qiaomu：**

Like when we wrote books back in the day,

**Guizang：**

No worries.

**Xiangyang Qiaomu：**

You need illustrations, and image generation is definitely unreliable for that. There was a site called Napkin — Napkin as in the tissue — that turned selected text into vector graphics, but with very limited layouts. If this kind of model gets widely used, I think there's a lot of imagination for the print publishing industry.

**Guizang：**

And big-screen displays.

**Pan YANG：**

Could the OS come with it?

**Orange：**

I feel like Astra probably also used data made with Astra. Astra itself has some ability — you give it a photo and it can draw a line-art version with code, quite close. I feel Astra is accelerating; next it will make a lot of heavy models extremely fast, because making data becomes easy.

**Guizang：**

I think—

## 12 / Using GPT-6 Astra to make data and train small models

**Pan YANG：**

This is what I saw this week: Astra has given us a new arbitrage space — use Astra to create new things. For example, use Astra to fine-tune small models for specific domains. Today's Astra, together with Fable, is already capable of this. Before, you needed a professional algorithm-and-training person to tune a small model yourself. Now with Astra's help, you give it the requirement, it sets up all the data and the whole infrastructure needed for training, and then trains it for you from 0 to 1, done!

**Guizang：**

Didn't a bunch of similar GPT models come out today? I bet a lot of them were just run through Astra for a day and out they came!

**Orange：**

DeepSeek's 4.1 Flash — is it also built on Astra\-lineage operators?

**Pan YANG：**

Right — and Zhipu would say it was written with their own 5.5? And the DeepSeek guy's article says it was written by GPT?

**Orange：**

Right, because they said on Zhihu that operators can already be written by Astra.

**Pan YANG：**

Right — this is lifting yourself up by stepping on your own right foot!

**Xiangyang Qiaomu：**

Then what's the standard to judge that by? Is it only achievable with Astra and Fable 5? Do domestic models have it — can the current ones iterate on themselves too?

**Guizang：**

Didn't Tang Jie say the other day they're working in that direction?

**Pan YANG：**

Right!

**Xiangyang Qiaomu：**

Is there a node past which \[the others\] can never catch up,

**Pan YANG：**

That's also an inflection point.

**Xiangyang Qiaomu：**

Because we've talked about this before — if there's a huge, incredibly strong overseas model researching and iterating and improving itself, everyone else might just never catch up.

**Guizang：**

I think there may be a singularity. We used to talk about that moment — once it arrives, say, when self-sufficiency in the whole training process hits 80%, progress might explode overnight and no one can ever catch up again.

**Pan YANG：**

I think the base point may already be here,

**Guizang：**

But—

**Pan YANG：**

We just don't know it yet,

**Guizang：**

I've already bet on the right one.

**Orange：**

About there.

**Pan YANG：**

Note a date: by December of this year, ChatGPT will have been out for 5 years. There was a saying — 5 years after ChatGPT's launch comes AGI. I believe it.

**Guizang：**

What people chase now is ASI. And then— you were saying—

**Orange：**

Look, DeepSeek insiders have already started worrying about losing their jobs — it's already arrived. I think his article blew up this week too, everyone's seen it. It's a bit like in the middle of a battle, a soldier who already has nothing left to do.

**Pan YANG：**

Right, so people's understanding will shift quickly: these technologies, this engineering — none of it is a moat. What people used to talk about — what you have in hand today, what you're building, whether as an individual or a company, being the moat — that got discussed relatively little, in small circles. I think very soon it becomes a question every person and every company has to face.

**Guizang：**

Which brings us to his hottest part — I think the hottest is the last section, about,

**Orange：**

Hitler?

**Guizang：**

About a communist cyberpunk-style 2077 oligopoly?

**Pan YANG：**

Right, ask it. Just ask. Probability.

**Guizang：**

You could feed it this article, feed it that final question and—

**Pan YANG：**

Right, I think — both outcomes are possible.

**Guizang：**

Nobody knows! And the funny thing is — it's decided by humans!

**Orange：**

In the end it might not be decided by technology — that stuff might have nothing to do with technology,

**Pan YANG：**

Also a dice roll, then?

**Orange：**

In the end it comes down to how people decide.

**Guizang：**

On the very problematic — last week, weren't people talking all week about how Sam, Dario and Musk were discussing slowing down — it even looked like they'd reached a consensus.

**Orange：**

Right.

**Pan YANG：**

Haha.

**Orange：**

Everyone agreed to slow down!

**Guizang：**

Right — and then that evening Trump posted a tweet. Trumpy posted a tweet,

**Orange：**

Right.

**Guizang：**

The next day Sam rushed to clarify: that's not what we meant.

**Pan YANG：**

They were just about to ship a new model.

**Guizang：**

Right,

**Orange：**

No wait, Musk is such a hypocrite,

**Guizang：**

Shipping right away.

**Orange：**

On one hand saying "we agree," on the other saying our 4.8 will crush everything,

**Guizang：**

Shipping, benchmarked against 2.5—

**Orange：**

There's no way you people are stopping.

**Guizang：**

Then this Monday — this Monday at some All-In event or something? On stage, phoning Jensen Huang: ship it for me right now.

**Orange：**

Old Huang. The double act went viral across the whole internet — that post has already passed a thousand, it blew up.

**Guizang：**

Next: StepFun released three voice models, including the realtime-interactive StepAudio 3 Realtime 3 Realtime, StepAudio 3 ASR Max 3 ASR Max, and full audio content generation.

## 13 / Realtime voice and Computer Use

**Pan YANG：**

Feels like a ton of realtime models shipped this week — everyone is crowding into this lane. I think the time has come. On the other side, Computer Use is about there too — more and more scenarios of operating the computer through voice interaction. I've already tried the ASR models in this space myself, because we make a podcast and need transcripts — I set up the environment locally. After testing, I think it's better than what had been SOTA on my end, Qwen ASR — mainly stronger at recognizing proper nouns, and better at handling spoken language too. I think there are still a lot of low-hanging fruit in speech models. So many talented people went off to work on LLMs, nobody cares about this area — invest a little and you still get big gains.

**Guizang：**

Next is Google — Google also released voice models. It follows Gemini's updates: when Gemini updated to 3.8 Flash, its voice models updated to 3.8 Flash too. One Live and one Thinking — it's two, right? One for real-time conversation, the other for deeper, more complex tasks. It talks while it reasons — it can keep talking with you while reasoning, and once the reasoning result comes out, it naturally carries the conversation back to that result.

**Pan YANG：**

That said, I think for quite a long stretch, very few people will actually bring in realtime interaction models. A lot of the examples are on mobile, and phones — including the Apple — are constrained by the OS; it's hard to do what you want, so it usually happens inside apps.

**Guizang：**

Inconvenient.

**Pan YANG：**

I think the next big breakout point might be realtime models plus Computer Use capability, further completely liberating us. Before we started, didn't I post a video in the group — operating the entire computer purely by voice, completely hands-free. I think large-scale application here might arrive any moment now. Speaking of which — there's more; I'll bring it up when we get to hardware: the Doubao phone. Later.

**Guizang：**

That's about it for models. And another one: the robotics company that previously worked with OpenAI released a Helix 2.5 2.5.

**Orange：**

Figure?

## 14 / Embodied AI and zero-shot tasks

**Guizang：**

Figure released a Helix 2.5 2.5 robot model, mainly improving robots' zero-shot task success rates — things like tidying rooms, folding towels, making beds — raising the zero-shot success rate of those tasks from 9% to 56%. Feels like progress in embodied intelligence is still very slow, especially on the brain side.

**Orange：**

Right, its video demo is decent. When I watched the video it got a little scary at the start — it's picking socks up off the floor. Mainly in China its robots are all remote-controlled, right — it slowly picks things up, and it's quite interesting.

**Guizang：**

So, back to that Jev — I actually think it's very important in this area. The problem with controlling robots is that once it has a brain, it can't react quickly to its limbs and its voice,

**Orange：**

Right. Right, right.

**Guizang：**

With this thing, it can adjust joint parameters in real time. Like the steering angle and force of hydraulic or electric-assist systems,

**Orange：**

Right!

**Guizang：**

Output power — adjusting in real time to what level?

**Orange：**

Right.

**Pan YANG：**

Didn't we also say this week that Astra is a big boost for embodied too? Look at the two things this week — Astra plus Jev — that's basically maiming all the embodied folks.

**Guizang：**

Right, but I think it's a good thing,

**Orange：**

Right.

**Guizang：**

A good thing — because the embodied models used to have no visible hope; progress was extremely slow.

**Pan YANG：**

It further shows that once you pick the right base model, you just crush everything else, right.

**Guizang：**

Very important, right! And another thing — lately I,

**Orange：**

So—

**Guizang：**

You were saying?

**Orange：**

For embodied, this feels pretty important, because everyone has been waiting for this intelligence. Honestly didn't expect this: they've always been choosing among many models, many companies, and in the end it's OpenAI that comes and solves it for them. They're still selling hardware — doesn't matter — unless OpenAI also builds its own robots.

**Guizang：**

Weren't they partnered with them before,

**Pan YANG：**

I think I saw a news item,

**Guizang：**

Didn't they stop working together later?

**Orange：**

Makes sense.

**Pan YANG：**

I seem to have seen news that OpenAI is also doing embodied. Think about it from another angle: for a base-model company to improve model capability further, it has to devour data domain by domain. Who used to bother doing 3D? Who bothered doing Computer Use? OpenAI's capability rose on the back of these domains, pulling its capability up again and again. And sometimes you can't even tell whether it's the model's intelligence rising or its capability rising — the two have become hard to separate. Anyway, it just keeps getting stronger,

**Guizang：**

Right, and I also have that—

**Pan YANG：**

So next, OpenAI and Anthropic will definitely keep hunting for data, for domains, devouring them one by one — there's nothing they won't eat.

**Guizang：**

I have one more case. Just now, Fang wrote a very rudimentary assisted-driving algorithm. He modeled it for a car — the car makes quick judgments based on current road conditions, like traffic lights. I think that's also very important; it opens up a new path for assisted driving. Assisted driving needs very fast judgments and at the same time very high intelligence — before, those two were seen as hard to combine. And this also requires multimodal support. Once multimodal is supported, I think there could be enormous progress.

**Xiangyang Qiaomu：**

These are all Jev, right?

**Orange：**

I think Jev takes Transformer, tokens and those techniques and moves them to another dimension — pretty creative.

## 15 / After Vibe Coding fills up your drive

**Guizang：**

Qiaomu raised an issue earlier — we finished discussing models, and this fits right in: as model capability keeps growing, everyone will do more and more Vibe Coding. Disk space runs out. Lots of people have this problem — storage was already expensive, and we already had those small-but-beautiful apps taking up space, so things were already stretched thin, and then Vibe Coding came along. AI madly produces content for you and madly opens Worktrees, and for many people there really isn't enough. My previous computer was 512 GB — I thought, how could I ever use that up? Once Vibe Coding took off I was cleaning daily, and finally switched to 2 TB. This 2 TB machine is three months old, and just now I looked — 1 TB is already gone.

**Pan YANG：**

My 1 TB is already at the point where I have to start deleting things.

**Xiangyang Qiaomu：**

I did a complete reinstall yesterday. I couldn't take it anymore.

**Orange：**

I'm okay — I recently did my own cleanup. My old computer was 512 GB, nearly full; after cleaning, it's now at 200 GB. I found three things—

**Pan YANG：**

Still not enough!

**Orange：**

Your terminal keeps every package — installers and all kinds of packages all get kept. One is Chrome, including the browser — when an agent uses Chrome there's a huge cache, 60 GB or more. Feishu — any app with a built-in browser, have the agent run it once — takes even more space than WeChat. There's a cache, and that's easy to delete — just clear the cache.

**Guizang：**

A lot of the time it's never used. Wouldn't even go look things up? I took a look—

**Pan YANG：**

I've started shopping for drives lately,

**Guizang：**

My main— go ahead.

**Pan YANG：**

External storage, NAS and all that — I've started stocking up.

**Guizang：**

I got some too. For us content people, it's often video. I checked — raw video recording specs are huge; record for 10 or 20 minutes and 20, 30 GB is gone. And these are your masters — you can't delete them, but not deleting them doesn't work either; you can only deal with them later.

**Pan YANG：**

Then the other day I remembered an antique — my family later bought an external DVD burner, we still have it. I dug it out thinking, could some cold-storage stuff be burned onto discs? Then I looked it up — a DVD is only 6.9 GB.

**Guizang：**

There are big ones now. I remember watching MediaStorm — they also store data on optical discs.

**Pan YANG：**

A DVD is only 6.9 GB, and a hard drive today is several TB. We record one podcast and a DVD isn't even enough.

**Guizang：**

Filled right up, yeah.

**Xiangyang Qiaomu：**

Say you use a NAS — how does that solve it?

**Pan YANG：**

Today—

**Xiangyang Qiaomu：**

Like a NAS — but doesn't your network need to be fast too, otherwise there'd be latency.

**Guizang：**

No need — the NAS is on your LAN.

**Pan YANG：**

There's LAN too?

**Orange：**

Intranet. I was looking in the wrong place.

**Guizang：**

If that doesn't work, just plug in a Thunderbolt 5 straight to the NAS.

**Xiangyang Qiaomu：**

I saw someone say that when buying a Mac Mini there's a 10-gigabit Ethernet option, and they didn't pick it at the time and deeply regret it. So going forward, it feels like 10-gigabit networking should be standard, and the bigger the drive the better — even 1 TB isn't enough. I saw lots of replies saying 1 TB isn't enough either. When my 512 GB was nearly full I cleaned daily, got sick of it, and gave Astra a command: I want to recover 100 GB or more. It went directory by directory searching for me — in the end it did manage to free that much, but it nearly deleted every app I regularly use. I figured there was no point, so I just did a complete reinstall.

**Pan YANG：**

My biggest one is WeChat,

**Orange：**

Right, speaking of which—

**Pan YANG：**

And what else — Claude has a VM that's super big, and if you develop for iOS, the iOS images are super big too.

**Guizang：**

I just saw it — four systems add up to another 50 GB. One iPad, one iOS, and iOS and iPad each have the highest, lowest, and middle versions installed for testing — that's forty or fifty GB together.

**Orange：**

So—

**Xiangyang Qiaomu：**

Remember that.

**Guizang：**

Xcode itself is big too.

**Orange：**

So a bunch of people are waiting to use foldables — tapping the top of a foldable — starting at 256G, custom builds from 20,000 up.

**Guizang：**

1TB.

**Pan YANG：**

By that logic, in this era — is cheap online cold storage useful? I don't know if it's a business.

**Guizang：**

Isn't ByteDance doing hard drives? Doing cloud services means doing drives — doing cloud, doing netdisks, enterprise netdisks, enterprise netdisks for agents to use,

**Pan YANG：**

Right.

**Guizang：**

Volcano Engine is doing it.

**Pan YANG：**

That's one meaning. Like, I don't buy a NAS at home, or I have a NAS, but the NAS is local hot storage — I don't need RAID. And online there's cheap cold storage. What's cold storage? What I want isn't instantly downloadable — I can only see what files are stored, the file names. Like, if I want to retrieve that volume, I send a request, and within 24 hours it lets me access it — that's cold storage. It could actually be made into a service. I just don't know if it's too niche.

**Guizang：**

Indeed.

**Orange：**

24 hours — I don't think people today can wait,

**Guizang：**

I think the core—

**Orange：**

People today want it in 24 seconds.

**Pan YANG：**

That's hot storage then, not cold—

**Orange：**

No — users don't care,

**Pan YANG：**

Because it's not—

**Orange：**

Then you have no user demand at all.

**Pan YANG：**

Because the price differs by 10x — say a netdisk at 68 yuan a month, now you make it 6.8, same capacity?

**Guizang：**

Isn't the biggest chunk of a netdisk's costs traffic? This lowers traffic costs too, since there are fewer accesses, so traffic is naturally lower. It might download—

**Pan YANG：**

Right, this—

**Guizang：**

It might download something enormous in one go, which could be expensive.

**Xiangyang Qiaomu：**

Isn't it said that storage stocks keep rising — turns out it's because of data center construction. I feel consumer drives are slowly not enough either.

**Guizang：**

Right.

**Pan YANG：**

Hard drives are another category — they're not the same as memory. Not the same as SSD either!

**Xiangyang Qiaomu：**

You mean storage — it doesn't refer to SSD-type storage?

**Pan YANG：**

See: memory, then SSD, then hard drives? I saw news today that CXMT (ChangXin) is also getting into NAND flash,

**Xiangyang Qiaomu：**

Right, quite—

**Guizang：**

None of it is enough. Not good. Not good at it.

**Pan YANG：**

Because I think this could be a big revenue line later. Think about it: we've been in the digital era for twenty or thirty years — how many digital works and artifacts from the first 10, 20 years need archiving? Maybe every few years, the capacity we need to archive grows 10x. The digital content the few of us produce in one year might equal the total of the past several years. There will definitely be storage demand going forward. Buying a NAS at home is expensive too, drives can fail, and you need backups. To back up you have to buy two drives — the price effectively doubles. So, like I said, if there were a place that would keep a cheap cloud backup for me — pay a few dozen yuan a year for several TB — I'd be willing. Right now my Apple subscription is the 2 TB plan — 2 TB for the whole family at 68 yuan a month, and it's nearly full. On the next tier I might have to pay 200-plus a month for a few extra TB.

**Guizang：**

And just now there was another news item — they folded Apple TV and such memberships into that iCloud thing too,

**Pan YANG：**

Right.

**Guizang：**

And iCloud has AI tokens in it.

**Pan YANG：**

Not available in China, right?

**Guizang：**

Not in China — not in China or the EU.

**Pan YANG：**

Right.

## 16 / Making promo videos and soundtracks with code

**Guizang：**

There's one more thing we haven't covered. This week I had GPT-6\-6 make a promo video for my software — it made the music too. I thought it had found it on some open-source music library site. Then it made two, and I noticed the music was different each time, and both synced with the video content. I asked how the music was made, and it said with Python. I can't hear that it was made with Python. Before, AI-made electronic music was very monotonous — a few notes repeating forever, or a few drums thrown together; this one follows the rhythm of the whole video. I always assumed it was open-source music downloaded from a music library site — turns out it wrote it in Python.

**Orange：**

I saw someone specifically test this capability, especially Astra.

**Guizang：**

Impressive.

**Orange：**

Astra's capability is way up over Sol,

**Guizang：**

Right.

**Orange：**

When it writes music, it first writes a score, then converts it into music with MIDI.

**Guizang：**

Right. Right.

**Orange：**

It sounds great — indistinguishable from something written by a person.

**Pan YANG：**

Look — this is very much like the other thing we talked about last week. Last week we said Astra doing 3D means it can directly draw? It's the same principle,

**Orange：**

Right.

**Pan YANG：**

It's like it uses Blender to build 3D models to generate images and video, and Python plus MIDI to generate music — same principle. It just kills the native music, image, and video generation models outright.

**Guizang：**

I don't think it kills them — not yet, anyway,

**Orange：**

Right.

**Guizang：**

They still each have their strengths. The problem is its capability is getting scarier and scarier. You can see that every time Claude ships a feature, the videos are made by Claude itself — and those videos keep getting richer. I'm sure the people shipping features didn't put much effort into them, but the videos keep getting better and better. Astra is the same — the other day I had it try based on my codebase, and it did a great job. The music mentioned just now was made by it too; the components were off the shelf.

**Pan YANG：**

The absolute rise in intelligence means it can do everything. Those models or pieces of infrastructure we used to build specifically in some domain for a specific task — SOTA models, as their general intelligence rises, can eat all of them. Right.

**Orange：**

Right — this is something it learned on its own. It's definitely not from application training data, because the value of that is too low; nobody would prepare that data for it,

**Guizang：**

It must have generalized it itself.

**Orange：**

It just broke through that wall on its own.

**Pan YANG：**

Speaking of which — it can operate MIDI, and the basics it can definitely operate. So for example, the Journey to the West theme — what's it called, "Yungong Xunyin" — could you have it make its own version of Yungong Xunyin?

**Xiangyang Qiaomu：**

Right.

**Guizang：**

Right — electronic music, it can remix,

**Pan YANG：**

Right, right,

**Guizang：**

It'll mix it for you,

**Pan YANG：**

Right, right, totally,

**Xiangyang Qiaomu：**

It's all pure digital.

**Pan YANG：**

Rearrange and rework all the great tunes in history.

**Orange：**

Right — it can play you music while painting visuals, achieving a real-time video effect entirely from the code level. But it's still fairly slow,

**Pan YANG：**

Wait — next month, when we can't burn through all our tokens, let's save some and make a demo.

**Guizang：**

Let's try. Basically the whole video—

**Xiangyang Qiaomu：**

I saw a Vibe Coding friend of mine, Xiao'er — he has the model generate dynamic sound effects based on a character's vector-space position, and the results were spectacular. You can actually use this approach to make all kinds of visual effects — the mapping between music and characters,

**Guizang：**

Lately I keep getting fed a category of video: SVG and motion effects set to music as the visuals. You've probably seen lots of those creative websites — like grass growing along with the music. The grass is actually SVG-ified, a Bézier-curve animation, also generated from a single sentence. Like a sailboat — entirely through one one-shot SVG, drawn and grown line by line, with waves drawn around it using those lines. The sailboat moves, drifting with the wind — it's all stuff like that.

**Pan YANG：**

Really, the new era has its new-era artists — using the technology of the new era for all kinds of artistic expression.

**Guizang：**

Expression, right.

**Pan YANG：**

Right, expression.

**Guizang：**

Let's continue. On hardware, the first thing — why do I count this release as hardware? NVIDIA released a Rust version of CUDA? Before,

**Pan YANG：**

Right.

**Guizang：**

Before, he could only — he couldn't use it, he could only keep using the original. Now it's been rewritten directly — a rewritten CUDA — and since it's already very efficient, a lot of people have started writing with it.

**Pan YANG：**

Let me tell you, from this angle: NVIDIA is genuinely scared now. SOTA models will soon dismantle NVIDIA's CUDA moat completely.

**Guizang：**

Replicating a CUDA isn't that hard anymore,

**Pan YANG：**

Remember in our very first episode we talked about AMD? They've now started using Fable to write their whole driver suite, and just this week DeepSeek and Zhipu both started using models for inference acceleration. NVIDIA can't hold this moat much longer. Jensen Huang is clear-headed — he knows this perfectly well.

**Guizang：**

That's why he won't let those three giants keep suppressing the development of other open models — because they can do it and others can't, and the gap keeps widening. Next: Volcano Engine released a car AI assistant, the Doubao cockpit assistant. The main feature is phone-car interconnection. You can remotely control the car and check its status from the Doubao app, and push trips to your phone. Car controls, navigation, assisted driving — the car's whole set of functions can be executed through Doubao. For now it's first partnering with Roewe's Jiayue 07, and SAIC Audi is joining too—

**Pan YANG：**

These must all be the software-weak car makers, the weaker of the bunch, right,

**Guizang：**

But honestly, both the AI capability and the phone-car interconnection capability are not strong right now. The smart-cockpit part, I think, is far behind on cars — definitely far behind computers, and what it can do in the car is far, far behind phones.

**Xiangyang Qiaomu：**

You know what? That Brad Qiang — he's working at SAIC on the Doubao project.

**Pan YANG：**

I know.

**Orange：**

Look, the Doubao phone actually sells really well. You didn't know?

**Pan YANG：**

Didn't know — aren't there tons of complaints,

**Orange：**

No.

**Pan YANG：**

"Can't use this, can't use that"?

**Orange：**

It's selling like crazy!

**Guizang：**

Well, can't help that,

**Xiangyang Qiaomu：**

But the thing is—

**Guizang：**

Right,

**Xiangyang Qiaomu：**

Right.

**Guizang：**

It's not Doubao's problem?

**Orange：**

I think it's only you tech-circle people who think this thing is clunky here and broken there. The masses just see "AI", the marketing kicks in, and it sells out.

**Guizang：**

There are plenty of scenarios?

**Orange：**

Cockpits too. Look at the carmakers that used to care most about this — Xiaomi, Li Auto — their own in-house work isn't that great either. Xiaomi might be a bit better; the others are terribly weak.

**Pan YANG：**

Look — didn't the first version of the Doubao phone get roasted with "this can't do it,"

**Orange：**

It was pretty big.

**Pan YANG：**

"that can't do it either" — didn't the apps just block it right away? This version, ByteDance has actually been very restrained — for apps that haven't granted authorization, it proactively doesn't operate them either,

**Orange：**

That doesn't matter. Think about it — just like when Taobao blocked WeChat, and then Pinduoduo came along and now stands toe-to-toe with Taobao. Taobao blocked the Doubao phone, but Douyin now has shopping, group buys, food delivery — it has everything. Closed loop.

**Guizang：**

You mean me—

**Pan YANG：**

You mean, with these not open, ByteDance's Doubao phone can freely use AI to access its own businesses for smart assistance and automation — so ByteDance kills everyone off again.

**Orange：**

I think given the phone's sales— feels like—

**Pan YANG：**

What sales? Do you know its numbers?

**Orange：**

Go check the news!

**Guizang：**

What's the main thing? One is that it's not expensive — the phone is overall pretty middle-of-the-road, all the specs are adequate. Its price, I saw, is 6,500?

**Orange：**

Look — sales broke 100 million yuan in 1 second. And pre-orders were 370,000 units. Very hot.

**Guizang：**

That's probably Nubia's and ZTE's best performance in years.

**Orange：**

ZTE — a brand nobody buys — and it goes this hot just because AI got added. The masses are truly feverish!

**Guizang：**

The price, I think, is right — there's no AI premium; it's just a normal phone price,

**Orange：**

Right. Right.

**Xiangyang Qiaomu：**

Xiaomi — doesn't it also have some AI capabilities? And Guizang demoed it for me before — the Doubao phone is more system-wide, more low-level, so it's different — a completely different experience.

**Guizang：**

I don't think it's so much about being more low-level or being able to do things others can't — it's more about concept delivery and how it's demonstrated. The other day I suggested you watch Chen Baoyi's video — he did a Xiaomi HyperOS 4 demo showing how the system connects the person, the car, the home, and all things. Watching that thing, you realize you don't know how to use a phone. The comments all say "Lei should have him present it — enough of the other guy; talking about all those specs is useless." I think it's a communication problem.

**Orange：**

Think about it — if Xiaomi tried to interconnect all these services, it would connect even less. Xiaomi doesn't have Douyin's group buys either,

**Guizang：**

Its own services, right!

**Orange：**

Food delivery and such. In the end it comes down to competing on services — letting you enjoy all services through simple interactions, while having built out the full layout of those services itself. The ones who could fight it are maybe Tencent and Alibaba, but both of those are slow.

**Guizang：**

It's mainly that they own the content — consuming content matters a lot.

**Orange：**

Right, exactly.

**Guizang：**

Okay, and the next one — the Doubao phone assistant consumer edition. I think later—

**Orange：**

Charging.

**Pan YANG：**

Didn't we already cover this? Right,

**Xiangyang Qiaomu：**

Done discussing?

**Guizang：**

Right, I—

**Pan YANG：**

Didn't we already cover this?

**Guizang：**

I think later it becomes something — a different thing. He shipped the phone, and he split the phone assistant out. When AI isn't good enough — not just ZTE or the others — once those players see this, they'll go do some cooperation too. He's turned it into a service.

**Pan YANG：**

Isn't that the Huawei model? The AITO model?

**Orange：**

Right,

**Guizang：**

Otherwise you get throttled,

**Orange：**

Skip it.

**Pan YANG：**

What's wrong with that.

**Guizang：**

And then we're done—

**Orange：**

Don't mention that name?

**Xiangyang Qiaomu：**

Don't. Right.

## 17 / Voice input products and realtime feedback

**Guizang：**

Done — and we won't mention keywords, otherwise the comments blow up and it's over. Next: other things. First, Tencent also released an AI input method — I think this thing will clash with WeChat Input Method.

**Orange：**

Tencent — I'm not on WeChat— no wait, I've always been on WeChat—

**Guizang：**

Not WeChat,

**Pan YANG：**

Code?

**Guizang：**

It's like Typeless,

**Pan YANG：**

Right.

**Guizang：**

It's not an input method — it's a standalone app, right, that Chatterfly.

**Pan YANG：**

Right.

**Xiangyang Qiaomu：**

Isn't it Chatterfly?

**Pan YANG：**

Chatterfly — it competes with Typeless. And NetEase Youdao's what's-it-called, Bage is another one.

**Guizang：**

It's just that he released one himself.

**Pan YANG：**

Bage, I installed it and then uninstalled it. I thought about it: it's not that much simpler than Typeless, and it's not as good as WeChat Input Method's smart voice input. I use WeChat Input Method now, which isn't built-in, and I don't think I need to keep three. Keep one input method and one Typeless\-type thing — one of each category and done. I tried it, and there was no wow effect. And one thing I really can't stand: when you're dictating, it doesn't show what you're saying in real time — say 2 minutes, 3 minutes, you can't see anything, and then it all comes out at once. That's a bad experience for me.

**Guizang：**

Typeless is like that too?

**Orange：**

Typeless — isn't it like that?

**Pan YANG：**

Anyway I've barely used it lately.

**Orange：**

Typeless even raised its prices,

**Pan YANG：**

What I mainly use now—

**Orange：**

Cut the free tokens.

**Pan YANG：**

As long as WeChat Input Method's voice input gets a proper-noun dictionary, I don't need anything else!

**Orange：**

And WeChat made a press-and-hold input.

**Guizang：**

I depend on Typeless heavily.

**Orange：**

Everyone says it's really good.

**Guizang：**

Right,

**Pan YANG：**

It should have that too.

**Guizang：**

Great interaction, no problem. Is it just not in there?

**Pan YANG：**

Tencent's new Chatterfly lets you add a Skill. After you finish a stretch of speech, it can process it with a specific Skill — I'd call that an innovation, I guess.

**Xiangyang Qiaomu：**

Shandianshuo did that long ago.

**Guizang：**

Had it ages ago.

**Pan YANG：**

But in a way — what Shandianshuo—

**Xiangyang Qiaomu：**

Although it did it very early,

**Pan YANG：**

But in a way,

**Xiangyang Qiaomu：**

it overcomplicated it,

**Pan YANG：**

In a way, they overcomplicated the product,

**Orange：**

Actually—

**Pan YANG：**

Right.

**Orange：**

It's actually out of step with the times. The trend now is making things simple — all those prompt templates and such are being simplified.

**Pan YANG：**

Right.

**Guizang：**

The core issue is that its Skills are essentially still prompts. You can't operate the computer, can't execute tools. As a voice input method, it can't execute tools either — it's still just prompts. Polishing things with prompts — how do people put it? If you know, you know! Also, Claude Code finally supports AGENTS.md — I'd been maintaining so many separate copies,

**Xiangyang Qiaomu：**

About time. So stubborn.

**Guizang：**

Besides AGENTS.md, keeping a separate Claude.md — and I'd have to update both in sync.

**Pan YANG：**

It still hasn't supported Agent Skills,

**Guizang：**

Agent Skills still isn't supported — next step, hopefully they get that handled, otherwise it's still painful.

**Pan YANG：**

Still not supported? Right. Let me tell you what this reflects, sideways: it reflects that he no longer thinks he's the dominant one — he's been beaten into a sorry state by OpenAI.

**Orange：**

Right — didn't its OpenRouter call volume just lose first place to OpenAI for the first time?

**Guizang：**

OpenAI has subsidies,

**Pan YANG：**

OpenAI's cards should outnumber theirs.

**Orange：**

It's mad subsidies, mad discounts.

**Guizang：**

A lot more,

**Orange：**

Right, but this time it—

**Guizang：**

A lot more.

**Orange：**

The reason it could overtake actually still comes down to GPT-6 Astra\-6 Astra. And Astra doesn't get a discount at OpenAI, which is pretty impressive.

**Guizang：**

And the Codex experience itself — it has all kinds of problems, but compared to other desktop stuff it's honestly 10,000 times stronger, I think.

**Pan YANG：**

Right, and Computer Use is just too strong,

**Guizang：**

Right, the built-in browser!

**Pan YANG：**

Haven't I been using Zhipu's ZCode and GLM these past couple of days?

**Xiangyang Qiaomu：**

Right.

**Pan YANG：**

Finished a task these two days. I knew ZCode had Computer Use, but after using it, I found it's far behind Codex.

**Orange：**

Codex is now the de facto king,

**Guizang：**

On Computer Use it's out in front by a mile.

**Pan YANG：**

It has genuinely become my operation entry point. These days I don't open websites or apps myself for a lot of things.

**Guizang：**

It's like the operating system,

**Pan YANG：**

A lot of it goes straight—

**Guizang：**

You now—

**Pan YANG：**

Right,

**Guizang：**

Spread it all out, just use it — you can use your computer entirely through it, no problem at all.

**Pan YANG：**

Issue commands through it.

**Xiangyang Qiaomu：**

That's why I used to use OpenCode\-like routing relays to hook all the other models into Codex. People really can't be bothered to switch between multiple models and multiple harnesses — get used to one and stick with it.

**Pan YANG：**

What you said ties back to what I said last week: models and their homegrown agents are becoming more and more interdependent. If the model is bad, the agent can't push it; if the agent is bad, the model can't be pushed either. Right.

**Guizang：**

Next is another Codex piece of news — Codex's built-in browser now supports Chrome extensions. Now it's really over,

**Xiangyang Qiaomu：**

I tried it today — I found it so damn good,

**Pan YANG：**

Brutal.

**Guizang：**

Right!

**Xiangyang Qiaomu：**

It can load cookies, history, extensions — everything — it's basically a full browser now. Last time I was saying, should I look for a third-party browser,

**Guizang：**

Right.

**Xiangyang Qiaomu：**

Like ego (lite) or Tabbit, something like that—

**Orange：**

Wiped out.

**Xiangyang Qiaomu：**

Now I'm thinking it's probably unnecessary anymore — it's almost—

**Pan YANG：**

Right.

**Guizang：**

Before, logging into websites could be a hassle — now you just install 1Password for it. My 1Password is all authorization-based — no plaintext passwords ever appear; it authorizes with the key directly, it clicks once and you're logged in, very secure. My credit cards are stored in 1Password too — for payments and such, just authorize it and it can fill them in for you.

**Pan YANG：**

I used Safari for a long time, because I have choice paralysis — Chrome has too many extensions, Safari has basically none, so I just stopped using Chrome entirely. A big reason for using Safari was that it's tied to Keychain — I keep user passwords and such in Apple's Keychain. Now it's been split into a Keychain and a Passwords app. Chrome used to be unable to touch the Apple system's Keychain; now Chrome can directly read all the passwords and accounts in the Apple macOS keychain. I think beyond the 1Password features, for the basic experience, Chrome can already handle it on its own.

**Guizang：**

Right, it has that kind of thing itself.

**Pan YANG：**

One more thing it shows: the Codex team was quite committed to swallowing the browser — this is a previous decision landing. And there's another question I keep mulling: does anyone else feel that with Codex's built-in browser mode, the screen size no longer fits? I think we need a wider ultrawide.

**Guizang：**

Wider, right.

**Pan YANG：**

Keep a 16:9 browser, and add space on the left or right for chat and a session list. I've been wanting to buy a 5K ultrawide lately.

**Guizang：**

It fits well — splitting the screen is easier too,

**Pan YANG：**

Right.

**Guizang：**

Splitting screens on 16:9 right now is a real struggle — no way of splitting feels right, especially painful when browsing the web,

**Pan YANG：**

This way you get the original 16:9 browser experience while using the agent.

**Orange：**

Also — WeChat's recent update moved the browser to the right side.

**Xiangyang Qiaomu：**

So gross.

**Orange：**

Got flamed to death, right?

**Pan YANG：**

WeChat's update is brilliant, I think — it's exactly the same as Codex, you notice—

**Orange：**

But the experience is terrible,

**Xiangyang Qiaomu：**

But the logic is wrong.

**Pan YANG：**

Session list on the left, chat in the middle,

**Orange：**

The experience is terrible!

**Pan YANG：**

And the browser on the right. Right.

**Orange：**

Right, but it has no AI, so—

**Guizang：**

Also bad. Hard to use.

**Xiangyang Qiaomu：**

I was ready in advance!

**Orange：**

Right.

**Pan YANG：**

Whatever. Don't rush—

**Orange：**

Right,

**Pan YANG：**

Don't rush — everyone is converging on the same thing,

**Xiangyang Qiaomu：**

But like, you tap Moments and it pops up over there, and then on the left there's another notification for you to tap. You haven't gotten used to that behavior, so it feels like nothing changed — very weird.

**Guizang：**

Right, can't figure it out.

**Xiangyang Qiaomu：**

Have you used it for a while?

**Orange：**

They didn't do it well.

**Guizang：**

Right,

**Pan YANG：**

Well, I've used it for a while now,

**Orange：**

It didn't do tabs.

**Pan YANG：**

I think it's great.

**Xiangyang Qiaomu：**

Maybe your screen is wide. I feel it takes up space, and it keeps un-docking — sometimes it's docked together, sometimes you accidentally tear it apart.

**Orange：**

Mainly WeChat itself already takes up 3/4.

**Guizang：**

When its browser isn't showing, you have no idea where it will pop up. One moment it's the side browser, the next moment an independent browser window appears.

**Xiangyang Qiaomu：**

Right, there's no predictability!

**Guizang：**

Right.

**Pan YANG：**

I know what problem you're all hitting — there's a setting: open with the external browser, or open with the internal browser?

**Guizang：**

Isn't that—

**Pan YANG：**

At first I always used the external browser; later, after running into all your problems, I switched to opening links in WeChat's internal browser by default — and then there's none of that split experience you're describing.

**Xiangyang Qiaomu：**

That's not the problem.

**Guizang：**

It's a separate WeChat internal-browser window — you have a sidebar browser window open on the right.

**Pan YANG：**

Right.

**Guizang：**

Say you've opened one — like a WeChat official account article,

**Pan YANG：**

Right.

**Guizang：**

Then you open another WeChat official account — what pops up isn't the default browser, it's a separate window of WeChat's built-in browser.

**Pan YANG：**

I know. It just covers it, right.

## 18 / Codex multi-account and workspaces

**Guizang：**

It doesn't cover it — right now the two logics coexist, and you don't know which one will show up when. I hit it several times this afternoon while organizing information. Moving on: Codex's built-in connectors now support logging in with multiple accounts. Say you've connected Notion or Twitter MCP — the Notion or GitHub connectors now support switching and can bind multiple accounts. Say your company uses Notion and you personally use Notion too — you wanted both your company and personal Notion logged in so it can read that content. Now it can read both; it can read content from multiple accounts in one go.

**Pan YANG：**

Isn't that just a Chrome feature?

**Guizang：**

These connectors are on the MCP side, nothing to do with the browser. The connector itself supports multi-account login.

**Xiangyang Qiaomu：**

Its account system feels pretty well designed — see, it can switch back and forth between multiple OpenAI accounts? And its other connections can also combine multiples — feels like a many-to-many relationship?

**Guizang：**

These aren't things you couldn't think of — it's that they move so fast. While other agents are still working on the basics, they're already going deep into these very fine-grained things. And then—

**Orange：**

It iterates together with the model, the model's environment — so this place is basically already in a self-accelerating state.

**Pan YANG：**

As for eating the browser — nothing left to discuss. I think Anthropic has just been way too slow on this.

**Guizang：**

Their engineering is just too weak.

**Orange：**

Isn't Anthropic now pivoting to life sciences?

**Guizang：**

A villain trending toward a full-on supervillain image.

**Xiangyang Qiaomu：**

I'm curious — the AI browsers of earlier years all seem to be gone now. Like, I use Chrome and occasionally ask Gemini something, at a very low frequency. The original selling point of the second browser seemed to be that you could have an AI assistant — that whole form seems to be fading away.

**Pan YANG：**

Right, you're right.

**Xiangyang Qiaomu：**

Don't know—

**Pan YANG：**

It's two routes. Look: Chrome adds Gemini inside itself — it's a browser adding an agent. Now Codex says: I'm an agent adding a browser. Let's see who wins — obviously the Codex route wins.

**Guizang：**

I think what Codex lacks now is tab and project management — managing browser tabs and projects together, in a linked state. That's needed. Its tabs are like Dia's — didn't Dia's former designer and interaction designer move to OpenAI? They worked on Atlas before, and now presumably on the browser — they'll definitely have thought of this.

**Xiangyang Qiaomu：**

Right!

**Guizang：**

Later they'll definitely do projects — linked management of the left-side chat projects and browser tabs. And then, honestly, I might not use a browser at all, not even Chrome. Because all your context is truly unified.

**Orange：**

In the end we'd use the browser to watch videos,

**Xiangyang Qiaomu：**

Man, I feel like—

**Orange：**

No other use for it.

**Guizang：**

A lot of the time I use it to scroll Twitter, and to read articles.

**Orange：**

Right, just consumption — beyond consumption,

**Guizang：**

Like Gemini — I rarely use the Mac app; if I can use the web, I don't use the app.

**Xiangyang Qiaomu：**

Codex wants to eat everything — eat a whole pile of software; all it needs to do is make Codex good.

**Pan YANG：**

Codex is now the strongest agent inside an operating system. The only thing I think is ahead of it is ahead on a different dimension—

**Guizang：**

An interface?

**Pan YANG：**

It's what I've kept insisting on these episodes: Grok Bot — and didn't Meta's one also rise up?

**Guizang：**

OK. Next news item: Grok Bot.

**Pan YANG：**

That one Muse, and what's it called — Instinct?

**Orange：**

Instinct

**Pan YANG：**

I don't know what that one's about — it's pretty hot too.

**Orange：**

Right now Silicon Valley is hyping this concept — personal agent. Tonight?

**Pan YANG：**

I think it will definitely get there. I firmly believe it!

## 19 / Always-on agents and multi-platform content distribution

**Guizang：**

Next up is Grok Bot!

**Xiangyang Qiaomu：**

Need a use case. A few days ago in Hangzhou I met Tw93 and asked him how he uses Grok Bot. He said he has Grok Bot automatically do multi-platform marketing promotion for him — promoting his Mole cleaning app. Because he has a dedicated computer running it around the clock, scheduled to visit niche forums in Europe and the US, then handing it to Grok to judge each forum's vibe. Some forums love indie devs telling their stories, some prefer knowledge-heavy, practical sharing. It posts based on the vibe, and it also judges when people are slacking off to time the posts. He only needs to provide the content.

**Pan YANG：**

Doesn't that wipe out a bunch of AI marketing companies?

**Xiangyang Qiaomu：**

Yes.

**Guizang：**

Right.

**Pan YANG：**

Right — using it is using Browser Use, Computer Use and such? Isn't this delivering outcomes?

**Orange：**

Right.

**Guizang：**

That template sharing of his—

**Orange：**

Right — made into a service, I think a lot of people would pay for it,

**Guizang：**

And Grok Bot's templates can be shared too — you can share your whole setup directly to its marketplace for others to use.

**Xiangyang Qiaomu：**

So I effectively gain independent software-promotion capability — way more powerful than a Skill!

**Orange：**

And Grok's English is pretty strong too.

**Pan YANG：**

Start delivering outcomes!

**Orange：**

What it writes has a nicely human flavor.

**Guizang：**

And then—

**Orange：**

Just in time.

**Pan YANG：**

Next I'm handing our website's growth task to it. See — Qiaomu isn't waiting for you anymore; just have Xiangyang Qiaomu do it,

**Xiangyang Qiaomu：**

Haha,

**Guizang：**

It might not adapt well domestically in China. Can't log into anything.

**Xiangyang Qiaomu：**

And everything is—

**Orange：**

Right.

**Xiangyang Qiaomu：**

Xiaohongshu blocks you at the first cut.

**Pan YANG：**

No need for domestic — there's no Internet in China.

**Orange：**

You can translate it into English and compete abroad.

**Guizang：**

No internet meaning—

## 20 / Coding and Office: two work modes

**Pan YANG：**

While we were talking, ZCode just got a new version update — it now supports Coding Mode and Office Mode. Coding Mode and Office Mode — a code-writing tool is starting to become, essentially, CodeBuddy turning into WorkBuddy.

**Guizang：**

About what I mentioned earlier — I think we could try using Codex's built-in browser as the main browser for a while.

**Pan YANG：**

Sure, I'll try it together with you — I'm interested.

**Guizang：**

Next: Grok Bot's supported voice — voice conversation capability. I think for low-cost agents, low-usage-cost agents, that's pretty important. Next is Claude Code — as mentioned, Claude Code launched Projects in beta. In one conversation it automatically handles the work — there's no more concept of chat conversations. Only the concept of projects: once you start a project, all your chats live together, and you don't switch chats yourself — it splits threads on its own, runs them in the cloud, and passes context between the different conversations. You can step away from the computer and it keeps running in the cloud. Feels like they're heading toward Grok Bot.

**Pan YANG：**

Feels way too complicated!

**Guizang：**

But — there are just too many services; you don't even know which one to go to.

**Pan YANG：**

No idea how to use it — the concepts are too complicated.

**Guizang：**

Right. Right.

**Pan YANG：**

Right, you—

**Guizang：**

He always—

**Pan YANG：**

Either natively in the cloud,

**Guizang：**

Spin off a—

**Pan YANG：**

Or natively local,

**Guizang：**

Or just spin off a separate app, a dedicated Claude client. I think there's no fixing it anymore.

**Pan YANG：**

Hasn't been updated in ages.

**Guizang：**

No way around it.

**Pan YANG：**

Feels like it hasn't had any big changes in forever.

**Guizang：**

And this time it's taking Cowork—

**Orange：**

Feels like they've already given up on treatment.

**Guizang：**

Right,

**Pan YANG：**

Why though?

**Guizang：**

And another: they merged Code, Cowork, and Claude Chat into one — copying Codex. Codex copied them before, right — like Codex — the ChatGPT part: ChatGPT and ChatGPT Work in one place.

**Pan YANG：**

Right.

**Guizang：**

Cloud processing.

**Xiangyang Qiaomu：**

No need to split,

**Orange：**

Right,

**Xiangyang Qiaomu：**

I think there's no need to split.

**Orange：**

So it's a big trend — simpler and simpler. Very few people truly understand what Work, Code and those things mean.

**Guizang：**

No need to split by task — don't carve the interface into different entrances by task type and make people switch.

**Pan YANG：**

Conversations, sessions—

**Orange：**

Right. Right, conversations!

**Pan YANG：**

Like we just said — the WeChat session list on the left used to be people and groups; now it's all agents.

**Guizang：**

Now it's chats and projects.

**Pan YANG：**

Right.

## 21 / Cloud Macs and agent execution environments

**Guizang：**

And another: Devin now supports macOS virtual machines. They bought a bunch of Mac Minis themselves and offer you the whole-machine service. In their virtual machine you can, like Grok Bot,

**Xiangyang Qiaomu：**

That's pretty interesting,

**Guizang：**

The VMs used to be Linux; their machines are Macs now, so a lot of AI tasks run more smoothly. You can also open remote desktop like Grok Bot and operate the software inside.

**Xiangyang Qiaomu：**

And they set you up with a Mac Mini.

**Pan YANG：**

This business has been around for a while — in the US lots of companies do exactly this: renting a Mac environment.

**Guizang：**

Right — they've now integrated it all together.

**Pan YANG：**

Right.

## 22 / Enterprise data and log boundaries

**Guizang：**

Then there's NVIDIA and other top US companies — concerned because Anthropic introduced a rolling 30-day usage-log feature in June. It saves your chats and information — whether you're a B2B company or a consumer user, there's no opting out; it retains your conversations on a rolling 30 days. Combined with the report a few days ago — they're all banning employees from using Fable 5, or banning employees from using Fable models.

**Pan YANG：**

Anthropic is effectively announcing to the whole world: I can access and store any company's content at will — and I'll say so out loud, too.

**Guizang：**

Right,

**Pan YANG：**

Who would dare use it?

**Guizang：**

As long as I decide you're unsafe, good luck to you — I'll just, you know. Last one: Apple released the new Siri AI AI — barely anyone discussed it.

**Pan YANG：**

Haven't even gotten to use it!

**Orange：**

Seems like nothing much.

**Guizang：**

A bit underwhelming.

**Orange：**

Barely made a splash.

**Pan YANG：**

Haven't used it yet — anyway, yesterday I upgraded all my systems to 27,

**Xiangyang Qiaomu：**

And it looks a lot better too,

**Pan YANG：**

Back then it still—

**Xiangyang Qiaomu：**

The way I see it — when Liquid Glass first came out it felt like a disaster, but now that the design language has grown on us,

**Guizang：**

Speaking of which—

**Xiangyang Qiaomu：**

with the patches added on, it feels okay.

**Pan YANG：**

You can adjust the transparency, the blur level.

**Guizang：**

The higher the blur, the more readable it gets. At maximum it's the original frosted glass.

**Xiangyang Qiaomu：**

Right, I turned it up a bit.

**Pan YANG：**

I'm on the default middle setting!

**Guizang：**

And that's all the news. There's some more stuff — about Liu Shengyu's article mentioned earlier, it spread a lot, including on Twitter; I saw lots of people sharing it.

**Pan YANG：**

Think about it: a DeepSeek prodigy, and on top of that taking shots at Anthropic — that's already a hot topic on its own.

**Guizang：**

Stacking traffic hacks?

**Pan YANG：**

Three traffic hacks stacked together. When have you ever seen a DeepSeek employee come out and write an article? Never seen it before — the moment one appeared, everyone rushed to share.

**Guizang：**

Anything like that can go viral.

**Pan YANG：**

Right — psychologically it closes the distance. You retweet an article, and psychologically you close the distance between yourself and DeepSeek.

**Orange：**

I think it's more than that. He himself evokes resonance — an algorithm engineer that brilliant,

**Guizang：**

The very top of the top,

**Orange：**

even he feels like he's about to be replaced — the resonance is extremely strong.

**Guizang：**

And when someone like him comes out and says this, people accept it — or rather, people truly agree!

**Xiangyang Qiaomu：**

DeepSeek — isn't its engineering strength really strong?

**Guizang：**

And then—

**Xiangyang Qiaomu：**

Because I heard from a friend who often chats with people inside DeepSeek — their verdict on Claude Code was like, where's the engineering in this? The harness is terribly written. That's how they see it.

**Pan YANG：**

Because of the DeepSeek team's background — they came from quant, so their engineering is stronger than other model-algorithm teams. Look at DeepSeek's whole journey: they consistently invested a bit more in engineering relative to other model companies, and that bought them huge gains over their peers.

**Guizang：**

Right — like your 4.1?

**Pan YANG：**

Right.

**Guizang：**

Inference speed, cost, and price.

**Pan YANG：**

What we ran into this week was leaning on the top models—

**Guizang：**

Truly outrageous.

**Pan YANG：**

to smooth this thing over again,

**Guizang：**

Right, exactly. And then finally there's a text—

**Pan YANG：**

It's still the same question: what is our value going forward? I don't think this is a pessimistic question — it's that we have to think and find our new position and value.

**Guizang：**

Right — what are you going to do?

**Pan YANG：**

There will always be something. Like we just said about art, about the music bit — there will be new-era artists who can wield the technology.

**Guizang：**

There are already lots of them — people doing expression, doing abstraction, including content and commercial content. I think this group is already pretty clear.

**Xiangyang Qiaomu：**

Judging from how the short-drama industry developed, I'm very optimistic — it shows that as long as people have good tools of expression, they can be incredibly imaginative and create things completely unlike anything before.

**Pan YANG：**

Right — see, before, to make this kind of thing you needed imagination, then you needed to write scripts, then you needed to make video, you needed to edit footage. Now AI handles everything behind it for you — as long as you have imagination and are willing to get hands-on, you can make a film today. Right — aren't people talking about the Americans obsessing over that American-style drama made by Chinese creators? Those two people have never been to the US, and they made an American drama that's hot in the US.

## 23 / AI dramas and content production

**Guizang：**

Indeed — lately every week there's a new hit American-style series, AI dramas, including new AI IPs, with very good numbers. It's not short-lived: click in and you'll find the whole dataset is very stable. Some may not be super hot, but people keep watching. Since it's stable, there's some ability to monetize, or other avenues.

**Pan YANG：**

One more thing — this week there was the Pinglu Canal news story, right? I noticed a phenomenon: this batch of Pinglu Canal short and mid-length videos is all really high quality — every single one is a gem, no matter whose shop produced it. There are all kinds of forms: standard blockbuster-style ones, like Planet Research Institute style, and ones made by all the big accounts. On one side there's editing capability, material-gathering capability, storytelling capability; and there are AI versions too — like turning the Pinglu Canal into a Nezha-style animation, AI short dramas — the results are all excellent. Through this event, one of my takeaways is that our whole society's capacity and scale for creating content is already completely different from six months ago.

**Guizang：**

Right — back to the problem: the time for consuming content is limited,

**Pan YANG：**

Yes.

**Guizang：**

Hey, and the even more brutal part—

## 24 / AI in R&D and human oversight

**Pan YANG：**

We'll have to grind on growth and on delivering results — because creation has already been max-leveled for everyone.

**Guizang：**

Last one: a frontier AI lab released an AI development-velocity metric — anyway, there are three metrics. The first is the degree of AI's involvement in its own R&D — the ASI automation trend index we keep mentioning. It classifies the company's internal R&D by human job categories and assesses automation levels. As of August 2026, it says Claude has not yet achieved autonomous operation in any stage, but plays a leading role in 26% of R&D work, and sits above the AI-collaboration level for over 90%. The second measures the ability to supervise AI agents — that safety stuff of theirs. The third measures how compute is allocated, especially the share of compute going to AI research. They say they put 6% of compute — about 12% of AI-R&D-driving compute — into safety work. Feels like apart from the first one,

**Pan YANG：**

Right — it feels like nobody believes what they say anymore.

**Guizang：**

The second and the third go back to the old narrative — they've started fear-mongering again.

**Pan YANG：**

Let's not get into it.

**Guizang：**

The first one is still pretty interesting. Okay.

**Pan YANG：**

How about we stop here for today?

**Guizang：**

Stop here — it's 10 o'clock today.

**Pan YANG：**

This is our first episode recorded like this.

**Xiangyang Qiaomu：**

Go record it online?

**Guizang：**

Recording online.

**Pan YANG：**

Online recording. Okay.

**Xiangyang Qiaomu：**

Right.

**Pan YANG：**

So — goodbye everyone!

**Guizang：**

Aren't you going to wait the 60 seconds?

**Orange：**

Hello.

**Pan YANG：**

No wait — first say goodbye to everyone. Okay, goodbye.

**Guizang：**

Bye everyone.

**Pan YANG：**

See you next episode. Bye!

**Orange：**

Bye.

**Xiangyang Qiaomu：**

Bye.

**Guizang：**

See you next time!

## Related resources

### Brands

- [Next Token | Beyond Tokens](https://nexttoken.tv/wiki/brands/next-token)
- [TypeSafe AI](https://nexttoken.tv/wiki/brands/typesafe-ai)
- [OpenAI](https://nexttoken.tv/wiki/brands/openai)
- [Anthropic](https://nexttoken.tv/wiki/brands/anthropic)
- [Apple](https://nexttoken.tv/wiki/brands/apple)
- [NVIDIA](https://nexttoken.tv/wiki/brands/nvidia)
- [DeepSeek](https://nexttoken.tv/wiki/brands/deepseek)
- [Zhipu](https://nexttoken.tv/wiki/brands/zhipu)
- [StepFun](https://nexttoken.tv/wiki/brands/stepfun)
- [Google](https://nexttoken.tv/wiki/brands/google)
- [Figure](https://nexttoken.tv/wiki/brands/figure)
- [ByteDance](https://nexttoken.tv/wiki/brands/bytedance)
- [Volcano Engine](https://nexttoken.tv/wiki/brands/volcengine)
- [CXMT](https://nexttoken.tv/wiki/brands/cxmt)
- [AMD](https://nexttoken.tv/wiki/brands/amd)
- [SAIC](https://nexttoken.tv/wiki/brands/saic-motor)
- [Xiaomi](https://nexttoken.tv/wiki/brands/xiaomi)
- [Li Auto](https://nexttoken.tv/wiki/brands/li-auto)
- [Nubia](https://nexttoken.tv/wiki/brands/nubia)
- [ZTE](https://nexttoken.tv/wiki/brands/zte)
- [Tencent](https://nexttoken.tv/wiki/brands/tencent)
- [Alibaba](https://nexttoken.tv/wiki/brands/alibaba)
- [Huawei](https://nexttoken.tv/wiki/brands/huawei)
- [NetEase Youdao](https://nexttoken.tv/wiki/brands/netease-youdao)
- [Meta](https://nexttoken.tv/wiki/brands/meta)

### Products

- [Jev](https://nexttoken.tv/wiki/products/jev)
- [Cola](https://nexttoken.tv/wiki/products/colaos)
- [WeChat](https://nexttoken.tv/wiki/products/wechat)
- [GPT-6](https://nexttoken.tv/wiki/products/gpt)
- [Jevable](https://nexttoken.tv/wiki/products/jevable)
- [Live2D](https://nexttoken.tv/wiki/products/live2d)
- [Heroes of Might and Magic](https://nexttoken.tv/wiki/products/heroes-of-might-and-magic)
- [Tetris](https://nexttoken.tv/wiki/products/tetris)
- [Lovable](https://nexttoken.tv/wiki/products/lovable)
- [Twitter](https://nexttoken.tv/wiki/products/x)
- [Weibo](https://nexttoken.tv/wiki/products/weibo)
- [The Hitchhiker's Guide to the Galaxy](https://nexttoken.tv/wiki/products/hitchhikers-guide-to-the-galaxy)
- [Stable Diffusion](https://nexttoken.tv/wiki/products/stable-diffusion)
- [Devin](https://nexttoken.tv/wiki/products/devin)
- [Qwen](https://nexttoken.tv/wiki/products/qwen)
- [TikTok](https://nexttoken.tv/wiki/products/tiktok)
- [Gemini](https://nexttoken.tv/wiki/products/gemini)
- [Mac Studio](https://nexttoken.tv/wiki/products/mac-studio)
- [MiMo](https://nexttoken.tv/wiki/products/mimo)
- [Arrow](https://nexttoken.tv/wiki/products/arrow)
- [Figma](https://nexttoken.tv/wiki/products/figma)
- [Napkin](https://nexttoken.tv/wiki/products/napkin)
- [Fable](https://nexttoken.tv/wiki/products/claude)
- [ChatGPT](https://nexttoken.tv/wiki/products/chatgpt)
- [All-In](https://nexttoken.tv/wiki/products/all-in-summit)
- [StepAudio 3 Realtime](https://nexttoken.tv/wiki/products/step-audio)
- [Doubao](https://nexttoken.tv/wiki/products/doubao)
- [Helix 2.5](https://nexttoken.tv/wiki/products/helix)
- [Chrome](https://nexttoken.tv/wiki/products/chrome)
- [Feishu](https://nexttoken.tv/wiki/products/feishu)
- [Mac Mini](https://nexttoken.tv/wiki/products/mac-mini)
- [iOS](https://nexttoken.tv/wiki/products/ios)
- [iPad](https://nexttoken.tv/wiki/products/ipad)
- [Xcode](https://nexttoken.tv/wiki/products/xcode)
- [Apple TV](https://nexttoken.tv/wiki/products/apple-tv)
- [iCloud](https://nexttoken.tv/wiki/products/icloud)
- [Blender](https://nexttoken.tv/wiki/products/blender)
- [Taobao](https://nexttoken.tv/wiki/products/taobao)
- [Pinduoduo](https://nexttoken.tv/wiki/products/pinduoduo)
- [Douyin](https://nexttoken.tv/wiki/products/douyin)
- [Xiaomi HyperOS](https://nexttoken.tv/wiki/products/hyperos)
- [WeChat Input](https://nexttoken.tv/wiki/products/wechat-input-method)
- [Typeless](https://nexttoken.tv/wiki/products/typeless)
- [Chatterfly](https://nexttoken.tv/wiki/products/chatterfly)
- [Bage](https://nexttoken.tv/wiki/products/bug-say)
- [Shandianshuo](https://nexttoken.tv/wiki/products/shandianshuo)
- [Claude Code](https://nexttoken.tv/wiki/products/claude-code)
- [OpenRouter](https://nexttoken.tv/wiki/products/openrouter)
- [Codex](https://nexttoken.tv/wiki/products/codex)
- [ZCode](https://nexttoken.tv/wiki/products/zcode)
- [GLM](https://nexttoken.tv/wiki/products/glm)
- [OpenCode](https://nexttoken.tv/wiki/products/opencode)
- [ego (lite)](https://nexttoken.tv/wiki/products/ego-lite)
- [Tabbit](https://nexttoken.tv/wiki/products/tabbit)
- [1Password](https://nexttoken.tv/wiki/products/1password)
- [Safari](https://nexttoken.tv/wiki/products/safari)
- [Notion](https://nexttoken.tv/wiki/products/notion)
- [GitHub](https://nexttoken.tv/wiki/products/github)
- [Dia](https://nexttoken.tv/wiki/products/dia)
- [Atlas](https://nexttoken.tv/wiki/products/atlas)
- [Grok Bot](https://nexttoken.tv/wiki/products/grok-bot)
- [Muse](https://nexttoken.tv/wiki/products/muse)
- [Instinct](https://nexttoken.tv/wiki/products/instinct)
- [Mole](https://nexttoken.tv/wiki/products/mole)
- [Grok](https://nexttoken.tv/wiki/products/grok)
- [Xiaohongshu](https://nexttoken.tv/wiki/products/xiaohongshu)
- [CodeBuddy](https://nexttoken.tv/wiki/products/codebuddy)
- [WorkBuddy](https://nexttoken.tv/wiki/products/workbuddy)
- [Cowork](https://nexttoken.tv/wiki/products/claude-cowork)
- [ChatGPT Work](https://nexttoken.tv/wiki/products/chatgpt-work)
- [macOS](https://nexttoken.tv/wiki/products/macos)
- [Linux](https://nexttoken.tv/wiki/products/linux)
- [Siri AI](https://nexttoken.tv/wiki/products/siri)

### People

- [Liang Wenfeng](https://nexttoken.tv/wiki/people/liang-wenfeng)
- [Luo Fuli](https://nexttoken.tv/wiki/people/luo-fuli)
- [Tang Jie](https://nexttoken.tv/wiki/people/tang-jie)
- [Sam](https://nexttoken.tv/wiki/people/sam-altman)
- [Dario](https://nexttoken.tv/wiki/people/dario-amodei)
- [Musk](https://nexttoken.tv/wiki/people/elon-musk)
- [Trump](https://nexttoken.tv/wiki/people/donald-trump)
- [Jensen Huang](https://nexttoken.tv/wiki/people/jensen-huang)
- [Brad Qiang](https://nexttoken.tv/wiki/people/brad-qiang)
- [Chen Baoyi](https://nexttoken.tv/wiki/people/chen-baoyi)
- [Tw93](https://nexttoken.tv/wiki/people/tw93)
- [Liu Shengyu](https://nexttoken.tv/wiki/people/liu-shengyu)