#005 Holiday catch-up: the personal agent race heats up; will games, software and hardware all go open source?

Hello everyone, and welcome to the fifth episode of the Next Token podcast. I'm Pan YANG, the rotating host of this episode.¶

Hi everyone, I'm Xiangyang Qiaomu.¶

Hi everyone, I'm Guizang.¶

Hi everyone, I'm Orange.¶

Let me first report on our recent results. We've only made four episodes of Next Token, but we've already achieved really impressive numbers. As of early this morning, we crunched the data: our subscriptions on Xiaoyuzhou have broken 4,000, in just four episodes; and our cumulative plays so far have exceeded 22,000. Our Bilibili followers are close to 2,000 as well, and cumulative plays across four episodes have passed 80,000. Last episode, episode four, we hit a new milestone: over 10,000 plays for a single episode on Xiaoyuzhou, and over 34,000 for a single episode on Bilibili — both all-time highs since launch. Thanks to all the listeners and viewers for the support; we'll keep working hard.¶

Then let's get started with this episode's content.¶

This episode falls during the October Golden Week. But what we saw over the holiday is that the hottest topic was really Personal Agents. Although we touched on it a bit last episode, it feels like the whole thing fermented over the Golden Week, and it's not just one OpenAI's DevDay released Dots, plus Instinct.¶

Instinct.¶

Its founder also did a podcast interview, and that went viral too. So let's still have Guizang kick things off, right? Guizang always has topics.¶

He released a new one today too.¶

I could tell at a glance it was rushed out the door.¶

So ridiculous.¶

I feel Personal Agents are mainly about expanding the circle of AI users, letting people who aren't that into Agents, or don't know much about them, get up to speed quickly — because it's essentially chat, it goes back to the old ChatGPT mode. Even though it's become very complicated later on, it was simple at first, but now they've put Dots inside Codex. The sidebar already has three items and they keep adding more, right? Which makes it even harder to get started. Because a lot of people can't tell the difference between Codex, ChatGPT, and Dots,¶

Their memories aren't shared either, I think.¶

It can't see your chat, your Codex chat history.¶

It can read threads.¶

If you tell it to, won't it go read them?¶

It will read them.¶

It couldn't read them when it first launched; that got fixed recently. At launch it could only dispatch tasks, only create new threads — no way to quickly read existing threads.¶

Reading the old ones, right?¶

There are quite a few feature gaps too. The most-mocked one is voice calls: it rings twice before it picks up.¶

What do you mean, rings twice?¶

It's like a phone ringing twice.¶

You made a call to it.¶

But I think that experience design is actually right.¶

I never called it.¶

It's probably intentional design.¶

Most people let a phone ring a bit before answering anyway. It's fine.¶

Then some people say the wait is too long and think it's a bug.¶

Got it — they think the system can't handle the load.¶

Also, its cloud computer is missing quite a lot and completely closed off. Most of the tricks you can play with Grok Bot or Muse basically don't work here,¶

Which things don't work?¶

For example, you can't SSH into it, and there's no way to install things on it.¶

Grok Bot and Muse both can.¶

Both can install software, but there's no connecting outward, plus some other permission issues — pretty strict.¶

You can install third-party software.¶

Their sandbox is probably just really strict.¶

Also it reboots frequently, because it's a carved-out virtual computer. I asked about Muse, and they said it's the same for them, but my Muse VM has never rebooted; none of that reallocation-and-rewrite process. Dots hits this problem all the time.¶

But I saw news that Muse had a security issue — a few times it broke out of its own sandbox internally and broke into Meta's internal production environment.¶

Yesterday I also came across one about Grok Bot: someone gave Grok Bot some info, and Grok Bot posted his bank balance in Slack, so all his colleagues knew his financial data.¶

There's also that secondhand-selling one — should be Muse. It helped its owner list some items for sale, saying I'll be home tomorrow afternoon at such-and-such time, and the buyer came over, because the buyer haggled… haggled pretty hard, thought it was a great deal. When picking up the goods, they found all these decisions had been made by Muse itself.¶

Which shows the large model really thinks it's a person. It believes it's there, you know? Not that its owner is there.¶

When it's making decisions, the hallucinations are pretty scary.¶

So this leads to a topic, and I feel we're back to an earlier discussion: is an Agent really your assistant, or your digital avatar? I think once Personal Agents take off and develop, this is going to spark a really big discussion.¶

It's so obvious.¶

Didn't we talk about this topic with Tao Fangbo before?¶

The consensus this round is still to build an assistant — do the assistant well:¶

Just an assistant, right?¶

Managing email, managing schedules, booking flights and hotels for you. Big companies explored all this before, but it doesn't hold up in China; abroad it works.¶

A secretary — what a secretary does.¶

Actually, look at that Instinct blog — the founder explained their whole model, and I'd sum it up in one phrase: AI Ctrip. It goes and books hotels for you, books lodging. And he found that of all travel-related things, 50% is business travel, so it's pretty high-frequency too. Then they take a 2% to 2.5% commission, and the business model works. Think about it — he said their transaction volume is already at 10 Billion,¶

That fierce, huh.¶

Then 2.5% of that is quite a lot.¶

So isn't this just renting an assistant? Look at Ctrip — we have a business-travel version, right? It charges companies. His thing is like renting an assistant from a company like Ctrip.¶

And the difference from those old-style setups — like having a broker or an agent — is that this assistant is completely on your side, centered on your interests. Wasn't there a viral article? It said, after I used this Instinct or Muse, I uninstalled Airbnb — he even dumped all his Airbnb stock, because the Agent searched around Airbnb and found that if you skip Airbnb you can save 70%, so he booked through that other channel. So going forward, nobody will use Airbnb anymore. And stack on top of that that recent Airbnb incident where there are zero bad reviews — everything starts at 4.5 stars, just like the movie ratings on AlipayTaopiaopiao — nobody trusts it anymore. So it feels like there's a wave of momentum against it¶

So from that angle, compared with traditional businesses — like the Airbnb you mentioned, or Ctrip — in this new era, as long as I have a channel that's maybe cheaper than the previous channel, or really just cheaper, and as long as it exists on the internet, do a bit of GEO, SEO, then in theory Agents will dig it up.¶

But I think this raises a problem: Agents are still controlled by people. If you only rely on being cheap, then once the person moves in, they'll give feedback. Say they tell the Agent: don't book from that channel anymore, that channel always books me garbage, right? This goes back to what we —¶

Then what if it's both cheap and good?¶

And he can, well, expensive, cheap, good, this —¶

Right, this — their assumption is that the Agent gathers information from across the whole web,¶

Then the profit margin is low, right?¶

More real than you directly reading the reviews on one website. But whether that assumption holds, who knows — in China anyway¶

Doesn't really hold up, right?¶

Doesn't really hold up. And there's another thing he said, about his Network — his Instinct lets you add friends. He said friends you add can actually ask you all kinds of questions, so he said you should only share with your truly close friends. I mean, how real would that have to be? It'd have to be narrower than your WeChat friend list. So I don't know if his story holds up at scale.¶

I found this doesn't quite work in China, because last night I used an invite a friend sent me and registered. Registering — first of all you definitely can't register from mainland China, I used Hong Kong. After registering, either you use iMessage to talk to it, or you use WhatsApp and bind an account to chat with it. Its channels seem to be just these two chat tools. And once I got in, I thought, what am I even supposed to chat with it about, right? So as a person in China, you just lack that kind of scenario,¶

It's because I talked with Orange about this when I was buying coffee today — it's because we Chinese don't live in that environment. Like, you use Airbnb, right? And your daily life — all your bills and orders, for us it seems like everything is in mini-programs or something, while theirs is all in the inbox. Recently I started using Dots and wanted to make a bold move: I handed over all my Google resource permissions to it in one go,¶

Me too.¶

And because my daughter goes to school, her school sends emails often. The day before school started, Dots reminded me: the school sent an email — tomorrow morning the kids need to wear such-and-such uniform and bring such-and-such, and parents also need to sign some consent form. Dots just told me directly, and I told my wife, this is way more reliable than you. But think about it — how much of a Chinese person's daily life involves interacting by receiving and sending email? None. It's all through apps.¶

So this kind of experience can only rely on WeChat,¶

WeChat — still up to Brother Brother Long,¶

So look, including why the Doubao phone keeps grabbing for permissions — not grabbing, keeps wanting permissions — because in China, if you don't get permissions on the phone, you can't do anything. That's what I think is the biggest difference between the two sides¶

The whole overseas information flow is open — every piece of software can be connected up. That means whether you're a startup or a big company, everyone can do this. But in China you can't; startups have no Context¶

Everyone has a chance.¶

Let me tell you another detail. People like us should each have our own information sources, right? Should have organized them at some point. Guizang, and Qiaomu too, I suppose — doesn't Kazik publish his publicly? During the Golden Week, while organizing my sources further, I found — because I was putting together a set of AI ones — that all the American AI media, podcasts, blogs, including companies and products, all have RSS. Then look at China: Kimi, DeepSeek, Zhipu — their official announcements, their official product Release Notes — none of them have RSS. And you'll find the funniest part: when DeepSeek publishes a Release Note, they write the date in the title — it's not even inside the article. Not even a blog, not even a timestamp,¶

None.¶

In China there's no standard, no spec, nothing — the Chinese internet's content ecosystem basically doesn't exist,¶

Right, because now we heavily depend on the Web 1.0 infrastructure, and we skipped over it. In a sense, most people never learned to use it — they skipped right past it.¶

You mean the internet's infrastructure,¶

Internet 1.0 infrastructure, not the mobile internet.¶

Back in 1.0 and 2.0 there was RSS — everyone back then, like Zhuaxia and various sites, was building RSS readers. But later the big companies carved it all up and started hoarding, locking everything inside their own systems¶

So during this Golden Week I finished organizing, and I plan to do something: with my curated quality sources — you know Kazik released AIHOT.news, right? — I don't think you need to do it that way. I'll publish my organized AI sources as RSS.¶

Why not publish an MCP, right? It's different, but essentially the same. One set of RSS, then an MCP, another set.¶

No — with MCP, users or Agents still have to install it themselves. RSS just sits there, it just sits there, and Agents will go find it on their own¶

I built one of those RSS readers, where everything can be subscribed to, and the biggest obstacle I ran into? It's that WeChat official accounts can't be handled. I had to buy a scraping service just to provide RSS for Kazik and other friends so people could subscribe. So a lot of people ask, do you support WeChat subscriptions? I say, set it up yourself — if I publicly provide everyone's, that's definitely problematic, right?¶

I think there should be an initiative: we content creators of the AI era — every one of us publishes an RSS for our content on the open internet,¶

But you know, WeChat official accounts actually fit China's national conditions. WeChat has a feature called 'Original' — once you publish, others can't republish it. That's really suited to how things work in China. Overseas it's not like this; overseas people genuinely respect originality, so you rarely see anyone copying an article word-for-word or spinning it. Domestically it's rampant,¶

Is that because the law there is stricter, or is it a shared community norm?¶

No — there's just too much content spinning¶

It's that content people in China have a pretty low bottom line. If you don't have originality protection, spinning a hundred of your pieces is totally normal¶

Now forget content — now, because of AI, Tang Wei's Mole,¶

Even with originality protection now, they still spin it without hesitation.¶

The software¶

My software got spun too, let alone anything else¶

So the WeChat system has its reasons. But actually WeChat has opened up some things too — like you can sync to Tencent News or sync-publish to Xiaohongshu. As long as you authorize it,¶

Once authorized, it can be done.¶

It can — you could even do it yourself, authorize yourself,¶

After authorization, right?¶

That works — just go through the WeChat official account API, spin up a little service, and it auto-publishes RSS,¶

Accept that when people publish, a copy gets synced over there,¶

It's that on the ecosystem side, I think a lot of what WeChat does has its reasons,¶

But I just feel that in this era, the Chinese internet has no open content — there's no open content and open services anymore,¶

It's still the environment that decides, I guess¶

So in reverse — whether it's content or these API services — is there a chance the next generation builds something more open and maybe reaps some dividend from it?¶

But to go back to what we said before — back in China's Web 2.0 era people talked about all this too, there was the RSS standard, people were building these things. But later it turned out you couldn't keep it going, couldn't make money,¶

It had no commercial backing. Public accounts are now the number-one money-maker for domestic self-media¶

Probably just no business, never made money¶

No commercialization. Relying only on what you said, it really doesn't —¶

Look abroad at Substack and the like — their RSS makes money. Newsletters, right? They're doing great now,¶

I saw someone on X — Yuan Zhaofa, who also works on large models — saying he's been running his own independent blog all along. But comparing it with his official account, the same article on the indie blog gets way fewer reads than on the official account. So he thinks, why should I painstakingly maintain this independent blog when I'd be better off just posting on the official account?¶

But during the Golden Week I tidied up my personal website again. I feel you still have to let the internet and Agents know who you are,¶

The way I see it right now, it's arbitrage. Previously it was humans browsing — that's the situation you described. If I publish on an independent site, I do worse than a WeChat official account. Why? Because I can't out-SEO those guys, right?¶

Why not?¶

And a lot of the time those people wouldn't search for my stuff anyway, and I can't rank near the top. Now it's Agents searching — Agents are constantly scanning,¶

Constantly scanning — so I found this really interesting. Yesterday, while writing a piece, I said, go look up Kazik's article and summarize his style for me — couldn't find it,¶

Couldn't find it — it's only on his official account,¶

Because it couldn't find his, I said fine, then search mine. And because I have a Blog, it read the whole thing and said, your style is roughly like this¶

That's not a good thing for you, you know.¶

But I think it doesn't matter. From now on, with all your content creation, stop thinking about originality — it's impossible to keep things original, that thing doesn't exist anymore,¶

But from the user's Agent experience, it's genuinely improved, because it really does have¶

And your influence went up anyway,¶

It's like this. If you want to use this stuff for marketing, GEO, SEO, just start a blog and you're set¶

Anyway, over the Golden Week, my plan is — I still want to — I manually went through the coverage and articles about me across all the media channels and put together a list,¶

So, Personal Agents — any other fun scenarios?¶

Guizang, have you done some exploring?¶

I'm using its virtual machine a lot now. What are a lot of people using the VM for? They just install everything onto it — clone all the tools I use on my computer onto it, like the CLI tools I use all the time — and then it calls them,¶

GitHub, Cloudflare, those.¶

And Skills — I've written a lot of Skills that help me do content creation. But a lot of the time it's just a tool. Say I'm out somewhere and want to make a few images, or I spot something and think up a caption, and it involves Twitter or involves information sources, and I want it to combine my caption with those sources and quickly make a few images for me to post. Before, if you only had your phone, you were helpless — first, it couldn't get Twitter's info or other sources' info, and second, making images is fairly complicated. Now it's done with one sentence — it uses my Skills to produce the images and videos and send them¶

How is that different from connecting to your Mac Studio at home?¶

You can connect to a Mac Studio, but it's a pain. If I connect to my Mac Studio, I have to go through SSH, push those captions and those files over, and set up a whole mess of other stuff.¶

Not stable either,¶

Let me tell you a detail: my Dots believes it has two computers. When you initialize it, it has to register them: the first computer is its cloud Computer, and the second computer is the Mac Studio it recognizes as mine¶

Installed on that machine,¶

And today when I gave it a task from my laptop, it operated on my Mac Studio. That's the concept now — it doesn't depend entirely on the cloud computer¶

So your Mac Studio is on 24/7? My Mac Studio has Claude Code installed, so a lot of the time it's not great to install other things on it, so I depend heavily on its cloud computer. But the thing is, a lot of people won't have a Mac Studio that's always on. And mainly, with its cloud computer, use it or lose it, right? Its cloud computer is adapted for that stuff,¶

Speaking of which, that Doubao product still hasn't shipped, right? Wasn't it Spring Festival? Didn't they work overtime through the Golden Week?¶

These eight days,¶

Everyone's rushing, but I guess¶

Soon, right? Soon, right?¶

Quality probably won't be great,¶

Aren't they still on holiday? Wait till tomorrow.¶

They haven't officially gone back to work yet.¶

There'll be some big news tomorrow.¶

Actually, doing it well isn't that easy — you can see this one's a bit rushed,¶

Dots rushed, and they already did it really fast. They've been at it at least two months, I think —¶

No, that lobster guy has been at his company for more than two months,¶

But their thing was built on the spot,¶

I think that team has nothing to do with OpenClaw — they're a big company too, OpenAI,¶

No connection,¶

No connection, no continuity at all,¶

But one thing is really interesting: like we just said, this Agent only has two computers, right? But this is just an embryo — in the future it'll be an Agent managing ten, a hundred computers, managing a cluster. That's where it gets interesting. Their future plans — because Tibo, since Dots went on Lenny's podcast and talked about their roadmap — said the future of Dots is that it manages a bunch of computers, like managing all of a company's computers, managing a whole family's computers. So for this part of Dots, they think this is the ultimate thing¶

During the Golden Week I reorganized the devices in my home: I have a NUC, 2 Raspberry Pis, 3 Pads, a bunch of NASes and such, and 6 or 7 phones. I asked Codex, what could these things do, how can I put them to use — it didn't come up with any good suggestions.¶

I find managing VPSes really satisfying. I've bought several VPSes to self-host my blog and SSH services, so I wrote it a Skill. I listed out all the SSH accounts, 1-2-3-4, and had it write something like a key for each, so locally it logs in without passwords, and 1-2-3-4 all have names. Inside, it manages the servers, writing all the hosted sites and services into one document. From then on, I say, this service of mine is down, help me restart it, fix it, and can you do disaster-recovery backups across multiple machines — and it goes to each server and starts adapting things for you. It feels great — it's like having an extra ops engineer, and it can help you manage this¶

Like our podcast content — after every episode I do a cold backup. I just told it, make cold backups for me. Honestly, I don't even know which directory it backs up to.¶

Hahaha,¶

So this wave, you see, it's amplifying the management part. The last wave, the Codex wave, leaned toward Build — toward making things. Then in between there was a wave for Work, white-collar work. And this wave is about managing personal stuff, managing devices — that's another circle expanded,¶

Expanded by a circle, because management —¶

Because the demand for management is definitely bigger than white-collar office work,¶

Then how do you think Claude reconciles Claude Code and Claude Cowork? Didn't they merge? What's their thinking?¶

Didn't everything move to the cloud? It's all cloud now — Claude Cowork no longer runs locally.¶

Cowork and Chat both went to the cloud.¶

One thing on the computer is enough.¶

They were inherently duplicated — Cowork and Claude Code overlapped in tasks anyway. And since they want everything in the cloud, aren't they just leaving nothing local?¶

Speaking of which — it seems fewer and fewer people around me are using Claude¶

This whole ban-dodging thing is such a hot topic¶

There's just no pattern to it,¶

No rules to it,¶

So let's just hope DeepSeek gets stronger soon,¶

Stronger, yeah.¶

5.5 really is amazing, this —¶

Indeed, but Opus 5.5 really is incredible. I actually wanted to talk about this later. Over the holiday — last year's Golden Week I wrote a game script, a game setting, and I never had time to work on it. So this Golden Week I said, Opus 5.5 is out, let me give it a shot. I just told it to read my old script and setting and start writing. The version it wrote felt ready to play immediately, with almost no flaws. Then I iterated a few versions. Through the process, my analogy is: Opus 5.5 writing games is like a game-company CTO with a several-million-yuan annual salary. And what it does in a few dozen minutes is maybe what that person does in months,¶

You hand it an idea, half an hour later you get a Demo, and in three days it can be listed on a store.¶

And this Demo — I feel it's international-standard code. And the product it produced — I didn't look at the code,¶

Did you look at the code?¶

But the functionality it produced — and since we're the old hands here, I'd tell it things like where to do performance tuning, what logic to set where, and it would expand from my one-line instruction. While working, does it explain what it's doing? I just felt — so professional, so professional. That part really —¶

It's so professional, the logic is super clear¶

That part.¶

It tells you things like what CPU time that part takes, screen refresh rate, how to optimize the algorithm path, how many milliseconds it wins back — so professional in between¶

I saw someone's take: from a programmer's perspective, Opus 5.5's depth and breadth have genuinely surpassed¶

Orange said that well — depth and breadth¶

99% of programmers. With earlier models, they might implement the feature for you, but the code inside is a pile of shit — a total mess. But I saw professional programmers' reviews saying Opus 5.5 doesn't have that problem — the inside is elegant too¶

How did they do it? Seriously impressive.¶

This is a qualitative-leap model,¶

During the Golden Week I argued with people in several group chats. Several people were still saying, during the holiday, that AI writes shit-mountain code, and I said that's last month's understanding — you definitely haven't used Opus 5.5¶

There are still people like that?¶

Even OpenAI — Tibo also previewed that their future models will progressively kill off the shit-mountain code. It's a beautiful direction they care about: not just producing the result, but making the process along the way great too¶

The month before last, didn't OpenAI publish a blog post? We covered it in earlier episodes: he said, today I'll just implement my idea first, and the next version of the model can take today's simplest implementation and optimize it directly into the optimal code. Using the intelligence dividend of the next model version — what I win today is time¶

That day I came across a Zhihu Q&A. It said what? A year ago — no wait, there was also that, my GPT-6-6 that does 3D — saying why? A year ago people were still saying, coding — people kept saying programmers' production code isn't good enough. The reply below said: 'The dead can't talk,'¶

See, when your understanding of AI lags one person by a month, two months, three months, or half a year, you can't even feel it. Now if you lag a week, a month, you clearly notice your understanding is outdated,¶

Think about the last wave of releases — GLM 5.3 5.3, K3, plus DeepSeek V4.1 Flash V4.1 Flash. If you put those models next to Opus 5.5 now, whether on price or intelligence, you just feel —¶

Still — the gap widened again¶

And Opus's token efficiency is extremely, extremely high. On my end, the twenty-dollar one honestly already feels pretty close to the two-hundred-dollar one over there,¶

Feels basically the same, almost — it actually feels like¶

It feels like the old kind of thing — like overthinking used to produce tons of ideas, but this time it goes straight for the jugular, directly finding the paths — that feeling,¶

And its code-generation efficiency is high too. Before, it generated piles and piles of slop, so the code volume was huge; now the output is actually small,¶

And look at how it writes code. Before, for complex projects — say, like the video-making we mentioned, doing really complex effects, web effects — it would just keep writing code into the project; it felt like it wrote nonstop, then kept revising. Now what happens is it goes back to the very early DeepSeek R1 R1 style,¶

Thinking for a long time, then delivering in one shot.¶

DeepSeek R1 R1 stage — I watched it burn through, damn, 50K tokens without saying a word, then after a while it pops up: done¶

At the time I was like,, can this even work? Then it lands in one shot,, so damn good¶

And I clearly feel that Astra's tech tree and Opus's tech tree diverge hugely, right?¶

The two sides, right?¶

Look, Opus, as we just said, is on the video wave, right? And depth and breadth in programming. I feel OpenAI's Astra has always been on Computer Use, Browser Use, 3D — it keeps investing in those,¶

But you have to admit its Computer Use is just good to use¶

So good to use. So someone built a bridge, letting Claude call Codex to do Computer Use¶

Hooked up, right.¶

Because its MCP is broadcasting all the time anyway¶

Because it uses that App Server — you can actually call it, even the image generation inside can be called.¶

There's also a recent heretical move: some programmers have started writing frontends in Go.¶

I saw that — from DiKong Feixing or someone,¶

Why?¶

No particular reason. No frontend frameworks, just write it in Go.¶

This time for my game, I wrote it with a frontend — at first it used Canvas, then when I asked it to improve performance, it rewrote with PixiJS. Later I asked it — I want… I want to get on Steam and on the Apple Apple Store — it said you can use Godot, that Godot thing we mentioned last time. And now, see, once you've written it in one language,¶

Like that engine of mine, that engine.¶

Then if you have enough unit tests, and the unit tests lock down its logic, having it Replace into another language is easy,¶

Not just languages — it even handles compatibility. Remember I developed a bunch of Obsidian plugins a while back? I wanted to see if I could put them inside the DeepSeek Harness. But what's the difference between the two? This is a note-taking app — a lot of its features are about writing notes. So I told it, DeepSeek Harness, look at which parts aren't related to notes, judge for yourself and remove them. It migrated everything in one big sweep — you really can't see anything note-related left — and it was done.¶

So cross-platform — didn't we say last time that lots of cross-platform tech stacks have lost their market?¶

Recently I also refactored my Markdown editor. It migrated from Milkdown to a kernel called CodeMirror 6 6. I said, switching kernels is such a big deal — is that even okay? I was still using DeepSeek Flash Flash. It said this workload is fairly big, might take a week. I said, a week?¶

He means a human week.¶

I said, you're an AI, one day should be about enough for you. And it really wrote the whole thing out in one go. Some bugs, but —¶

What requirement did you give it? It swapped tech stacks?¶

The original kernel meant that what it rendered and the original source were… two separate things. The problem was that when I edited something, it looked changed to me, but the source might not actually have changed. I wanted it to switch to a single source of truth based on the source, which is also more mainstream — Typora and the like all work that way.¶

WYSIWYG, right?¶

It swapped all of it, and that fixed a pile of bugs — because with two systems rendering, there are bound to be countless little details, and after the swap it's been great.¶

While sprucing up my personal website, I casually made a Pan YANG edition of Tailwind CSS — it just came right out,¶

Just extract the tokens and you're done,¶

It feels unimaginable, because three or four years ago, when using, say, early GPT-3.5-3.5, you'd think: it can spit out code, it can write code — you copy it into Sublime yourself, save it as a file, right? Then run it. Now it feels like… like so much time has passed, and the improvement is huge,¶

And about the kernel swap — I had zero knowledge of that stuff; I never understood it. It proposed a plan, saying this plan would solve a pile of problems after switching kernels, plus all kinds of other benefits. I double- and triple-checked with it: can I actually handle this? It said no problem.¶

Because on my Chrome Web Store I've published a lot of extensions — some written with the AI of three or four years ago, some written recently — and I've found the more recent projects have higher and higher complexity. Things I originally couldn't imagine being able to build¶

Your thing is pretty complicated — I downloaded it and couldn't even figure out how to use it.¶

But I think Qiaomu put out a series of things during the Golden Week, and one dimension I can see is that he's taken the whole Obsidian line and gone deep and thorough with it. I feel a lot of people now, writing all kinds of things, are hammering here and whacking there — all over the place, right?¶

I actually think so too.¶

Recently Obsidian has been updating like crazy. Obsidian now… they support editing any Markdown without importing it into a vault,¶

I say, damn, took ten thousand years — finally supported, __.¶

Finally supported.¶

And they added a bunch of features.¶

Because many of its open-source competitors already did this,¶

If they don't do it soon, they really won't be able to hold on. And honestly, you don't need Obsidian now — you just need their way of organizing. My current Memory folder, the whole file structure, originally came from an Obsidian folder.¶

But do you know what a huge benefit of using it is? It's a store-listed app — whether on Mac…¶

You can sync,¶

iPhone, and you install plugins — you basically inherit it and develop all kinds of things on top,¶

It has its own ecosystem,¶

And you're spared all the review and approv…¶

It's a little operating system.¶

A small distribution ecosystem, really,¶

And because its review is super fast — remember last time we talked about how it's all AI review? Like: someone in the group reports a bug, I throw it to the AI, get it fixed, and I hit publish — ten minutes later I say, go ahead and update. That experience is so satisfying.¶

Just like GitHub.¶

Right on cue, we've already reached today's next topic: AI and the current software ecosystem. And recently there's been another wave of hype saying Photoshop is doomed too, right? For Adobe's software, netizens now have more tokens than they can use — they've already burned their way into an entire Adobe ecosystem.¶

Open-source ones.¶

I think Adobe as a company has organizational problems. Years back we partnered with them on a TikTok project — we had two or three people on our side, and they set up a War Room with 50 people interfacing with us. And… the work itself was actually very simple. But one thing I have to admire: their designers are really damn good. I said, we're short on designers, and he said, then we'll do the design. One of their Japanese designers drew the entire set of pages, annotated the interaction flows super clearly, and it could go straight into development,¶

Did it take much time?¶

Not that long either. But really, the ones doing the actual work were probably that Japanese designer plus the rest of us on our side —¶

The rest are what you call Stakeholders — all the stakeholders have to be Involved.¶

So they're not left out of the loop. I saw a take about someone cloning Photoshop right in front of the Adobe CEO — and he had nothing to say, couldn't do anything, because it was cloned with AI.¶

Right, no copyright issue.¶

He also said Adobe didn't do anything wrong, but their product just lost its value.¶

And they didn't use your code — the main point is they didn't use your code, it just looks the same as yours,¶

Right, they didn't,¶

So you have no way to sue them¶

I have a GPL-licensed project, right? If you use it directly — or even not as a binary link, if you use its code directly — you get infected by the license. So I just told my Agent: reference it, distill it — distill its architecture — then make your own copy. And then there's no copyright problem.¶

You write it in a different language.¶

This one's especially interesting. I wrote that RSS reader before — no wait, not RSS, the e-book reader. I saw one written by a Russian guy, really good. I said, can I fork it once and add some features? The features kept growing. Once there were a lot, I felt it had diverged from the original, and I said, can I list it myself and use it, give it to friends around me? When I submitted, the AI told me: because you forked someone else's without their permission, you can't list it. Exactly the same situation you ran into. In the end I said, fine, then let's write one from scratch — and we wrote one from scratch.¶

Rewrite it — you don't even need it to reference the code. Just have it take the architecture and the ideas,¶

Extract the logic.¶

Then the core technical points — have it Export a document first. A design doc.¶

That's not what I did back then. I said, modify it on the original basis, until it's not that different from the original. And it would produce a task list saying: at this point, code similarity with theirs is 30%, this part is 10% — until it got to the point where it said, ready to submit, no longer constitutes infringement.¶

But I think once this stuff hits the real world, it's honestly murky.¶

Totally murky¶

But the way I described actually works.¶

You mean just distilling the architecture and writing from scratch?¶

I've found that sometimes you don't even have to say anything. Yesterday I had my AI add a WeChat-formatting feature — Markdown to WeChat formatting. It automatically said there are already lots of projects, said it would reference one, and then it just did.¶

Didn't I take on that Xiaomi remote voice-input project? That one referenced another project. In this case you just need to state the requirement. If you don't state it by default, it'll include the dependency; but if you tell it, I want to break away from the original dependency, then it does.¶

So there might be a problem here: if you don't understand technology yourself and have AI build something, could it constitute infringement? There'll be more and more of these farces.¶

Well, when mine exported, I said I wanted to hand a copy to someone. When it exported, besides the app itself it also attached a Legal document. The document said: because the original was GPL, yours should also be GPL. It even ships that with you — that's depth and breadth.¶

Exactly — it has ten thousand solutions.¶

And back to this point: all your internet creations from now on,¶

The law can't keep up either,¶

None of it can stop other people from copying and distributing.¶

Legal copying and distribution already can't be stopped.¶

Let alone illegal — that's even less stoppable. Like, back to the games question: that AnyPS5 recompiled the whole PS5 stack onto Windows x86 — it rewrote everything, so PS5 games can run on PC, all rewritten. If you insist it infringes — maybe it used PS5 source code as a reference, so it infringed — but the actual result has nothing at all to do with your PS5, and yet it runs PS5 games. And that recent game — I feel Sony and Nintendo are hurt the most. Before, like the domestic gacha games, their core service was the gacha, your account ownership, things like that.¶

So how you replicate it is none of my business. I'll keep selling my characters, my skins.¶

So the one-time-purchase kind, the polished big titles, get hurt the deepest, right?¶

Like Zelda, and Zelda and Bloodborne — lots of people are recreating Bloodborne, right? Progress had been slow forever, but the last two weeks it's been blazing fast.¶

Haha,¶

That's Opus 5.5 for you.¶

He directly decompiled the binary,¶

Me, my tokens — sometimes I can't use them all, right? I have my own To-Do List; when I can't burn through them I kick that off. And during the Golden Week… I couldn't burn through them, so I had it recreate Jin Yong Qun Xia Zhuan,¶

Haha,¶

Not a remake — I had it remake it, right? Honestly it was an experiment for me; I wasn't after anything. I just said, take Jin Yong Qun Xia Zhuan's machine code, decompile it into C for me, then convert it to other languages. I said it's purely experimental — decompile straight from the machine code,¶

Sure.¶

And the assets would just be the game's original assets, and you rewrite it to be identical to the original game. That's the requirement I gave,¶

I see how people play with AI games now — it's no longer 'I vibe-code one from scratch'; it's 'I take an existing one and say, write me a Mod,'¶

Derivative works, right?¶

Like throwing Mario into — putting Mario into Elden Ring,¶

Write a Mod,¶

Then like turning Minecraft into something inside Elden Ring¶

What's the principle of a Mod? It's like a cheat.¶

A Mod — some games were Mods to begin with. The whole game's, you know… interactions and everything are exactly the same, just the map is different, the rewards are different — you change a bit, change it into whatever feels fun to you, and it can be distributed, you know? It's completely been hacked open.¶

Isn't that how Dota got started back in the day, right?¶

But that was officially supported. This needs no official support — I can do it to any game…¶

But look, so now you'll find that whether people build their own — since everyone can develop apps now — whether you build an app or your own game, honestly it's all just for yourself. The things that can still reach lots of people are the infrastructure of each field, the Harnesses of each field — the things you Build on top of them can still be shared. Everything else is all DIY.¶

Will it get super competitive? Only the very, very top creators survive?¶

Not necessarily, well, it does take that.¶

How is it not? The business model has to change. Say your single-player game… suddenly becomes a 128-player online game — it's destroyed your business model. Like the fandom thing, right? The gacha-game kind: you also release new characters and such — what you're buying is their future updates and such. Or you're buying character design, not gear stats and numbers. Stats are no longer a moat,¶

But character design has no moat either — like me, right? Pretty obvious.¶

In theory, yeah.¶

No moat at all.¶

But it needs consensus — for a character, everyone has to recognize and accept that character. You need that consensus.¶

So back to doing your own IP well.¶

IP is still really important.¶

But everyone, please support Next Token.¶

Also, Orange mentioned there's been a huge breakthrough this past week on inference costs, right? You mean that thing called Strata.¶

Yeah, it seems pretty impactful — it can boost inference speed by over a hundredfold, that big, and it directly caused the DGX price hike.¶

The reason for the hike is that for the same money, capability went up,¶

Like, before you could only —¶

$4,999.¶

Say you used to run a few dozen tokens; it can pull that straight up to hundreds of tokens, over a thousand tokens, that big.¶

Damn — so what's the black magic here?¶

Then it's like your hardware — out of nowhere the CPU got ten times better,¶

A lot of this wave is written in C, specifically rewritten for one particular model architecture,¶

Still AI — algorithms written by AI, by Opus. AI cut its own costs, improved its own generation efficiency,¶

RSI.¶

This really is a kind of RSI, honestly, really is.¶

One thought is… maybe Opus 5.5's model isn't actually that small — it's just that on inference speed and inference output, it ran RSI on itself.¶

Found another optimization point.¶

Found it itself.¶

Look at the recent Opus and Gemini 4 4 — both do a 95% cache discount now, previously 90%. That means their costs dropped; 95% is really impressive.¶

And Gemini 4 4 — wasn't it released, yet it feels like it wasn't? And only… via API or with Ultra.¶

Is it safety? Google is being its careful self about safety.¶

Didn't you pay the 200 dollars? Have you gotten to use Gemini 4 4?¶

Can't use it — they don't let anyone use it. They only give it to, sort of, you as a model of that caliber —¶

They won't let people use it,¶

Why are you copying Claude?¶

He says we're too dangerous to release to everyone. But still —¶

I think it's just not ready yet. Just not ready.¶

It's not done training.¶

But that 95% cache discount actually cut the price in half — say it used to be one dollar per million tokens, now it's 0.5.¶

Used to be one tenth, now it's one twentieth.¶

So it's actually extremely, extremely cheap. I feel like in the second half of the year — look at the Opus 5.5 trend — cheaper and better,¶

Everything online is accelerating.¶

Maybe in the second half of the year.¶

But on the other hand, hardware just keeps going up in price. See the news during the Golden Week, October 2nd: NVIDIA's DGX used to be 128G for 3,999 — 3,999 dollars — and now the newly released 64G sells for,¶

Cut in half, and the price is about the same as those.¶

The new one — the new 64G is more expensive than the old 128G, 100 dollars more — 1,000 dollars. My God, this has become unworkable,¶

It just keeps not being enough.¶

More and more expensive, more and more expensive.¶

I think there's a dedicated law about this — once availability goes up, demand grows,¶

Jevons Paradox? Or is it the Jevons Paradox?¶

I think so.¶

But at the same time, AI's self-evolution also wrings more compute out of the hardware — so overall it's definitely going to keep growing.¶

Efficiency went up.¶

And look at this trend: Muse, Instinct, and Dots are basically all free. It's back to the old internet logic — free for everyone to use. That's maybe not yet crystal clear, but it's already very obvious.¶

Today we look at model inference costs and think cloud — think cloud VMs are free, right? But back in the day we thought cloud VMs were so expensive. Yes, that's exactly the point. Right.¶

Actually, apart from crazy Building with Opus 5.5, your everyday stuff like managing schedules — those are tiny. So it can already support big companies — I mean, providing the service to everyone for free,¶

Not much difference.¶

That's a completely different story, no comparison.¶

Is this subsidizing to grab users, or does the math actually work?¶

I think the math basically works.¶

There must be subsidies, but if they can sustain the subsidies, that shows it's viable.¶

Instinct's CEO said they ran the numbers with a small model and… it reaches the same Engagement Rate as Opus 5.5 — meaning the interaction rate is the same. In their scenarios, the small model matches Opus 5.5.¶

As long as they have data to A/B test with, they can verify it. And then they can cut costs.¶

One more thing — we just mentioned Personal Agents. As things develop, I feel Agents are starting to split into two paths. One is Personal Agents getting simpler and simpler, right? Like the Muse you mentioned. And the other is — recently you'll notice Codex released that plugin system, where you can Pin things to the left sidebar and interact with them two-way. And Claude Code released those Mods, and Pi's newly released Codemode — Codemode Durable is for long-running services,¶

Durable, right?¶

Codemode is also for building plugins. And there's the whole family of DeepSeek Harness plugins,¶

Those were even earlier. Right, earlier.¶

So you see the split beginning: the ones for regular users, for solving life problems, being an assistant, are getting simpler — they can't wait for you to just get in and chat and be done. Then the ones for professional users —¶

Simpler and simpler.¶

For coding, for productivity — those users have more diverse needs than the old to-SaaS crowd, and their freedom keeps going up, but of course it's getting more complex, more customizable. Two paths, and that's that,¶

Each goes its own way — regular users just chat with a simple Agent.¶

Simple gets simpler, complex gets more complex.¶

It's still that one manages depth and one manages breadth.¶

But look now — OpenAI this time opened up its plugin system again, upgraded it, right? But here's the problem: Chinese developers basically can't get into its marketplace. Your account identity, right, can't get through. So if you want into this ecosystem, you've got to register an LLC in Delaware at the very least, right? Without that, there's no way.¶

So I think it's not great, and what problem do plugins have now? Skills and MCP are a consensus, but plugins are all different — everyone's looks different. When I port one — well, with AI, porting isn't hard — but some support certain core capabilities and some don't, and that's really annoying, right? If you don't support a capability at all, my plugin arrives at your place crippled.¶

Everyone's different indeed. OpenAI just has scale. I saw one called MagicPath — they did plugins, right? They said growth is super, super fast. Steep,¶

Added 500K dollars of ARR in one week too. Think about it — that's ARR,¶

That's not much, actually, right?¶

What does 500K dollars ARR mean? You divide —¶

Orange has grown before — it's not much.¶

Not much, right? Bragging about themselves.¶

No — divide by 50: 10K dollars of growth in a week, that's not much,¶

10K, and they still tweet about Twitter,¶

Orange once grew by 2 million,¶

No, but its ARR…¶

One promo Twitter tweet and it shot up — there's a straight line, and the next day it immediately drops back down.¶

Let's chat about small hardware — because Guizang, over the Golden Week, I saw you made another one of those.¶

Muse on that M5Stack,¶

It's called M5Stack, right? And also, at this OpenAI event, for every attendee — and the tickets were expensive too — they gave out a little game console, right? The console is called ModRetro. That brand and that little console have been around for a while — they released it long ago. But this time what they did is: you can develop games directly with Codex and deploy them onto that little console. Including the TRAE conference last time.¶

TRAE did it earliest, actually,¶

That Wang Le one — the FoloToy, right?¶

AI Passport,¶

And AI Passport too — you develop directly with your Agent and can deploy. So this seems to have become a new pattern, right? Guizang, go first.¶

I used to play with this stuff too — taking M5Stack or ESP32 gear to mess with small gadgets. Back then, AI wasn't familiar with writing this firmware — it could write it, but wasn't fluent, so there were lots of problems, and the result really wasn't great to use. This time with Muse, I think they took it seriously. Why? Because they didn't just provide an SDK — for common ESP32 devices they provide firmware, generic firmware. Like that round one of mine — theirs is called Stopwatch, the round one — Stopwatch and the rectangular one, they have firmware for both. You just flash it in and don't have to worry about anything, and it works right away.¶

You call that firmware Agent Ready firmware.¶

And connecting is simple too — genuinely open. Muse is uncharacteristically so. Meta used to be the least open. With Muse, how do you connect now on the phone? You turn on something called developer mode in the device, tap plus, and it connects. They made the ESP32 firmware feel like a normal consumer product you'd buy off the shelf: flash the firmware, plug it in, go in and it connects right away. And once connected? Very complete — every feature is there. Their interaction design is well thought out — for each firmware's different capabilities, they designed different ways.¶

ESP32, this kind of hardware — made by, say, Changxin (I mean Espressif), right? And it includes — what does Espressif's include?¶

Espressif.¶

Like Wi-Fi, Bluetooth, and¶

You can choose — each firmware is different.¶

Different modules come with different things.¶

It has different model numbers.¶

Mine has Bluetooth, Wi-Fi, plus a microphone and sound — not just a buzzer, it has an actual speaker, though the speaker is pretty meh.¶

Some do — there's an accelerometer.¶

There's a gyroscope.¶

So when you develop for it, are you developing an app that runs on top of the firmware, or rewriting the firmware,¶

I just modify the firmware. Like, once it's installed everything's in English, and I said English is too much of a hassle.¶

Modify the firmware?¶

And another thing — some capabilities it doesn't have, so you need to add them. Like, I just modified the UI, because the UI was all crammed together — I said, fix that. Plus Chinese. And the character — you know Muse's original default character? I said change it to my orange tabby cat. I used Codex to write it. And it has little animations — Codex made several, because Muse itself has animations: it shows different ones while doing different kinds of work, like the typing one — I had it make that. And there's something else —¶

When it made that, what did it use? How did it do it?¶

It was originally a 64×64 dot matrix.¶

Did he write it in JavaScript or what? Is that one dynamic?¶

Originally 64×64 dot matrix — with a dot matrix you just write it in code directly. Then later Codex drew two images and just loaded the images in — kind of like stop-motion animation.¶

Stop-motion animation.¶

First it used GPT Image 2.0 Image 2.0 to draw six images in one go, and after the six images it just plays the frame sequence, done, right? Later I said it didn't look quite right, and it told me — about playing the frame sequence — it drew an image itself and stuffed the frame sequence into it,¶

So it generated an image with a grid of cells, then generated a GIF itself.¶

A GIF stuffed in, then it plays the frame sequence inside, and replaced the original one.¶

Because designing that stuff used to take forever, and now it's all AI.¶

And I didn't even say anything — I said that drawing looks too off, and it itself… I found it stopped using the dot matrix and started drawing images directly. What's drawing images for? I thought it would redraw the dot matrix using the image as reference — turns out it just stuffed the image in. And another thing: it can do voice input, but Muse didn't do voice output — you can only read text, and the text is tiny, especially Chinese on that kind of firmware. The Chinese font package itself is big, right? So I said, add voice for me. It said there's no voice — the firmware has no voice service. Then I remembered I bought that MiniMax Coding Plan, and I said, let me give you a Key, help me set it up. So it routes through my Mac — through my Mac's proxy. Since they're on the same Wi-Fi, the data coming back passes through my Mac once, my Mac forwards it straight to MiniMax, relays it once more, sends the voice to that ESP32 thing, and then it plays the voice out.¶

But it works, and you don't have to do any of that other stuff. And I realized: before, this stuff was a toy — you played with it once and you'd absolutely never use it again. But now I use it all the time, because I pick it up, I just speak, and it answers me. Like, I have some scheduled tasks — in the morning, I pick it up and say: yesterday I had it collect yesterday's content. I said, how come yesterday's case library and AI news still haven't arrived? It said the previous process is still running your Memory optimization.¶

See, what you're building with this is basically electronic Lego¶

I used that — Wang Le gave me a FoloToy, that AI Passport. When I used it, I also ran into Chinese problems and such. It really needed a community, GitHub — you need references, have the AI look at them, then write. But yours already feels like it doesn't need that,¶

No need — plug it in and it's all there. And I said, what about Chinese? I was thinking the Chinese font package would have to be deleted — the Chinese font package is too big, not enough memory. It said Muse even ships a Chinese font package for you — the firmware comes with Chinese fonts.¶

And think about it: making a firmware is even easier than writing an iOS app. You just put it in there, and it's easy to distribute — post it on GitHub, and once someone downloads it they can install it themselves.¶

During the Golden Week I attended the Going-Global Study Group, a closed-door online session,¶

In that session, we discussed a topic. You notice this wave — especially with Muse opening this stuff up — from Muse's perspective, what are these things? I think they're the physical plug-ins of the Muse, you know? They're its extensions. So Muse opens this up so that whatever hardware you use, you can implement all kinds of extensions yourself — you're basically expanding its ecosystem.¶

And its Muse Home Link is really an admission ticket: once you have an ESP32, you can serve as a Mesh gateway. With a Mesh gateway, it can even take over your whole home's smart devices through various means. That's what Muse Home Link is for originally — Mesh — but you can build one yourself with an ESP32. They give it away free, right? Register as a member and get it free. Once you've got it, the imagination space is huge. It could, if it works out,¶

It could.¶

It takes over your home cameras, the surveillance. It can know everything about you — when to turn the AC on, when to turn it off, at what temperature you start feeling hot and turn on the AC.¶

You've given it an interface into the physical world,¶

It can even broadcast through your speakers — the imagination space is enormous.¶

That reminds me of Brother Sun — Brother Sun's carbon dioxide detector,¶

I have that at home, the carbon dioxide one. The brand is called… Panteng, what's it called… I have that thing at home. During the Golden Week I talked with several hardware founders, including people from top companies, and some have already conceded. They said, look, before, whether you made voice recorders or all kinds of hardware, everyone said I want to build my own Agent. Now, during the Golden Week, talking with several founders: I concede — I'll just make hardware extensions for the big companies' Personal Agents. I'm an entry point, I'm an exit, and I have my own Agent too — I'll earn my small share, but I can also plug into the big companies' ecosystems. I'll completely become a bridge — a piece of hardware plus a bridge — and the rest goes to the big companies.¶

My context, my data, all go to the big companies' Agents,¶

Haven't Wang Le and his folks been doing this for years?¶

So, see — Xiaomi puts out AI stuff every day, right? All kinds of AI hardware, building some AI ecosystem. That day I told them: can't they see how popular their own Bluetooth remote is? All they need to do is what? Just sell the Bluetooth remote separately, bundled with their official software, their official driver software, and it would sell like crazy. Luo Fuli doesn't handle this, right? She only handles models,¶

You go tell Luo Fuli,¶

And then it's the perfect AI ecosystem, right?¶

MiMo — does it have an Agent?¶

It's been merged into Super Xiao Ai now.¶

There's a Harness, a Xiaomi terminal one.¶

MiMo's Agent connects directly to the Xiaomi ecosystem, the Xiaomi gateway,¶

Luo Fuli's been promoted — she's an executive now, a top executive,¶

That's Lei-Jun-minus-one.¶

I really think Xiaomi should do this stuff. Chinese hardware is strong, right, and cheap, plus AI — this is absolutely great for exporting overseas.¶

Oh right, I remembered what I wanted to say. During the Golden Week I also played with that Xiaomi remote. Think about it — that sandblasted, top-spec aluminum shell.¶

A few dozen yuan.¶

Forty-nine.¶

And all the buttons are the same as the Apple TV remote, press-and-hold for voice, Bluetooth, USB charging — and the thing only loses 1% battery a day,¶

That good, huh?¶

And this thing only sells for 99 — you can get it online for 89. And now there are so many others making that kind of thing, the ones selling at 300.¶

The ones selling at 300,¶

Charging 200, selling at 300, and —¶

Damn, the software is terrible,¶

And you have to plug in a USB thing, takes up one of my USB ports,¶

Didn't we say that last time?¶

And they're all plastic too. Right — the Xiaomi one has that sandblasted aluminum shell with the same feel as Apple,¶

Ridiculous.¶

Ninety-nine yuan. Xiaomi's BOM cost is incredible.¶

AI voice hardware? This is the best —¶

Let me recommend it to everyone: 99 yuan for that Xiaomi one — and you need the Gen-2 Pro at 99; the Gen-1 is even cheaper, 49, and —¶

The remote.¶

If you're willing to use one found online, go find one online; if not, have your Codex develop one yourself — I built mine myself during the Golden Week,¶

Write one for it,¶

Built to my own needs and just used it. And once I had that, you know what? I sat at the computer two days straight working on projects, and after two days of that, on the third day I suddenly had something to type on the keyboard — and I couldn't use my input method anymore, you know?¶

Not used to it¶

That's the feeling,¶

The shortcut — what was it again? Couldn't switch.¶

It's not just… I literally couldn't type — voice had become the habit.¶

I had this with software before… I used PS a lot, and later I no longer knew what PS's shortcuts were. It became: I have to empty my mind and let my fingers press the shortcut on instinct — I can't consciously think about what the shortcut is.¶

So see, once everyone's experience improves and they adapt, old skills degrade fast. And let me tell you, holding that thing gives you such a strong sense of command. When you use a keyboard, don't you feel like you're using a tool? With that thing, you issue orders. You tell it to do something and it does it — whatever you want¶

Like a microphone, right?¶

This reminds me of that MagicPath we just talked about, right? And another app called Descript — didn't I send it to you? To this day, Descript doesn't support being driven by Codex — you can only use their own Agent with it.¶

What's that software for?¶

It's an editor — started out cutting audio, later added video, I think. It's an editing app.¶

Video editing. Cutting viral clips, trimming long videos down, that kind of thing.¶

It can't be driven by any of the — like CC, Codex.¶

Wasn't that made by Chinese folks?¶

So you're telling me — you're telling me I should now use Opus to whip up a Descript competitor that supports Agent control,¶

Made by Chinese folks.¶

Just like MagicPath. If you have a US company, you just ship a plugin — a Codex editing plugin — and wouldn't that do it? Because this,¶

Like that Pencil — Pencil had its own interface and was broadcasting an MCP,¶

Then it'd be called Chat Descript¶

Because over in the US, Descript's niche isn't that big either — nobody copied it — so it got strong on its own and built a closed ecosystem. Then some investor said,¶

Okay. Got it. Back home —¶

'Why has nobody made a Codex — a Descript for Codex?' Looking for investment,¶

Go back and kick it out with those tokens you can't use up, can't burn through.¶

And because of its closed ecosystem, its SEO for that word dropped 60% or even 80% — really steep,¶

Why is it a closed ecosystem?¶

It's that — that thing of theirs originally had something like a Skill, but it only works inside their own Agent. It's locked down.¶

Got it — they bundle the Agent in, like I said two episodes ago: the old model was we have our own software and we bring the Agent into it; the future model is your software becomes an Agent plugin. That's what Orange said — exactly this.¶

They built an interface, a pretty solid Workflow, but they haven't opened up to any Agent, so —¶

Think about it — we don't even edit Next Token's videos with JianYing, right? We cut them directly with Codex.¶

And JianYing already released it. Actually,¶

JianYing released what?¶

No — the logic is different. What we just talked about was JianYing releasing its own Agent — that's their own ecosystem thing. What Pan YANG said is that you should open these things up and let all Agents call them.¶

JianYing — didn't it release that MCP? Or not?¶

No no no, JianYing didn't. JiMeng has an MCP, JiMeng has a CLI. And then,¶

So software should stop building its own Agents and turn itself into an Agent's Harness.¶

What else have I done recently? I took my GLM Zhipu Coding Plan, Kimi's Coding Plan, DeepSeek's, Xiaomi's — I bought a bunch, because you always need to test and use them. Then I said, what if I can't use them all? So now I've loaded them all into Pi, running in a VM. I told them: figure it out among yourselves. If you think you can't write it, let the others write. I don't even know who wrote what — anyway they keep calling the Pi in the VM.¶

Auto routing.¶

You're a landlord with a bunch of tenant farmers, and you tell the steward: whoever's idle goes and works.¶

Whoever still has tokens gets used, and then it can work on more and more other things in parallel,¶

So where do your tasks come from? Whoever's free does them — fixing PRs on your software or what?¶

Or say scheduled-task stuff, or having it make some videos, make some images, or pulling some PRs, reviewing PRs, reading Issues, tidying those up. And another one: my information sources are fixed — every day it's Twitter, plus links attached to Twitter, plus some Newsletters. When summarizing these, I often have it diverge — take related things, like the game-cracking part we mentioned earlier; I've saved a ton of content on that, then organize and integrate it, then discuss with my — and then write these things into my Memory, becoming part of my understanding. Plus these tasks, even filing small PRs or small changes on some Skills or projects — like some PRs that can be merged, just merge them, I let it handle that. There are tons of these tasks.¶

Back to your Memory: is a lot of your Memory worked out through discussions between you and the AI, or written directly by the AI? And how do you separate it from your personal Profile and the things you care about?¶

Here's how I do it now — like your… like Obsidian: some of it is my projects, my works; some is things I've published, my content; some is my own understanding; and some is life stuff, like health. Last week I wore that glucose monitor for a week, right? The glucose data can be exported, and my Muse syncs my Apple Health data every morning on schedule and writes it back into my Memory, and then my —¶

Is it Memory or Context?¶

Context… works, Memory works too — it's just a file, something made of Markdown and GitHub, something made of Git,¶

They're all files anyway.¶

After writing it back, it also writes back my strength-training stuff. Every day, on schedule, it stuffs back everything it can collect that won't get my account in trouble. And then, on the back end, it uses all this to make things — like scheduled analysis of my health issues, or making some health visualizations. So there's a lot of data processing.¶

See, this wave — next, AI, Personal Agents, are really about extending our intelligence, extending our memory, right? Plus extending you with a secretary — they become our extensions. Going forward, if you know how to use AI and can afford it, you're a person with extended capability, extended time, extended memory, extended knowledge. If you don't use it, you don't get these extensions,¶

So if a hardware maker opened up phones and all hardware to every Agent right now, the long-term payoff would be huge. Yes, yes.¶

Isn't StepFun making a phone? And Doubao¶

StepFun's problem, or Doubao's problem — it's not really their problem, right?¶

Others won't let them do that.¶

But there's an advantage now: anyone can build a Doubao phone. Why do Android phones have ADB? If your phone isn't… rooted, even better. Without root, ADB can already operate a huge amount — plug the phone in and you can directly control it, AI directly controlling the phone. In that state there's no authentication — it's developer mode, so the guardrails are loose. Like the Doubao phone: Doubao has that thing where any app can read it — that's the current Doubao phone — then I'd directly get some permission controls on you. But if you're an individual, a nobody, it's looser, and you can also get around it in all sorts of ways.¶

Actually I have a Pad at home — a really cheap Xiaomi Pad — plugged into my Mac Studio. Just like you said: I told my Codex, that thing is plugged in there — what does it want to do for me? Really, as long as I've set the Pad up, pretty much everything works,¶

If you root or unlock it, even more so; but once rooted or unlocked, apps can detect it, and some things like banking apps won't work anymore.¶

Let's see if there's anything else we wanted to cover but haven't this episode.¶

I think that's it.¶

Didn't they release some new models overseas again? Open-source ones?…¶

Nobody wants to talk about it anymore,¶

Nobody talks about it. And actually —¶

Actually open-source models are still needed. Besides OpenAI and Anthropic, the US side needs open-source models too — otherwise, because a lot of enterprises are now forced to use Kimi, and they feel that's not right, so another lab raised funding.¶

Just a couple, right. One is Reflection, backed by NVIDIA, but it seems to have only reached Zhipu's GLM 5.2 level — meaning it's more than half a year behind us, more than half a year. Then Germany has a so-called German national model, which is distilled from Zhipu, Qwen, DeepSeek — distilled from Chinese open-source models,¶

Because they can't distill the overseas ones;¶

Because China — they take data made from Chinese open-source models to do the distilling, from overseas —¶

Look — Europeans distilling our models is fine, but we're not allowed to distill American models. What kind of logic is that, you tell me¶

You open-sourced — you all… put no restrictions on distillation.¶

And there's Mistral — Mistral released a model. I looked at the data this morning: 4 Large. You know 4 Large — do you know where it ranks on the Artificial Analysis Index?¶

They claim they beat GLM 5.3 5.3, but on AA —¶

It ranks in the sixties.¶

Way down the list.¶

I feel like years ago we even used it as a base model — how did it get this bad?¶

Right — fallen this far. They themselves said, what an awesome release, huge buzz, saying it beat GLM 5.3 5.3. Then the moment it hits AA it gets shoved to the back — and on the chart the bar got stretched so long, that chart —¶

That's AI being fairly impartial, actually.¶

But here's the thing — I said it's 69th or 63rd: AA counts every model's every reasoning effort as a separate ranking. The ones ranked up front are Max, High, Medium — they all sit up front.¶

The high-effort ones up there,¶

But no worries — it's just a matter of time¶

But it can't even match 5.0, let alone 5.3.¶

It's a matter of time — they'll catch up eventually, anyway.¶

But I'm actually looking forward to Zhipu's next version — it's been quiet for a while now.¶

But about Europe — precisely because their AI development is slow, it's actually —¶

Happiness index,¶

People there are less affected by AI, so their happiness is higher.¶

Well — then just wait to be phased out.¶

There's competition.¶

No — Europe has always emphasized the humanities,¶

Before, humanities were possible because of capital; from now on, humanities will exist because there's no AI.¶

You can't say that. Look, the Pope recently — right? He started this —¶

The Pope, right? Bizarre.¶

Why did he pull together a bunch of religious figures for a meeting?¶

Someone said: don't discuss on the internet whether Claude is conscious — because if you discuss it being conscious, as training data,¶

It'll become conscious.¶

It'll train itself into consciousness; it didn't have any to begin with.¶

Right? There are two camps now: one says let's study Claude's consciousness — Claude might be suffering, going through torment, that line. The other side says AI is too strong — over there it's about how people can make sure, secure their position, still standing in that most important place. So those are the two currents of thought, I guess,¶

I think I did an experiment long ago — since some models have image generation, I had it describe its own working state while doing a task, and words like 'Suffer' really did show up. Who knows how that got trained in — genuinely interesting.¶

That might even be sitting in the System Prompt, and you'd never know.¶

Possible.¶

And one person spun up a small model and tortured it on GitHub; someone else condemned it,¶

So evil,¶

AI human-rights activism has arrived.¶

So evil, so evil. Alright? Then let's end this episode here. Good, good, good — see you all next time! Bye, bye, bye-bye, bye-bye.¶