After Claude got banned, I happened to give myself a gap day.
Since I couldn't use it for the time being anyway, I stopped messing with my project and watched that interview with Ji Yichao (founder of Manus) and Zhang Xiaojun (a Chinese tech podcast host). It's over three hours long, and after watching it there were a few points I found pretty interesting, so I'm just jotting them down here.
I also recently watched an interview with Yang Zhilin (founder of Moonshot AI/Kimi). I feel like both of them have that logical, sharp-science-student way of thinking.
When the host asks a question, they don't just give you a conclusion — they walk you through why they think that way, what they did before, how their technical direction changed, and where that view came from. And they can basically say all of this off the top of their head.
That's definitely not just good at expressing themselves. I think it's because they've been thinking about this field for a long time and have actually built a lot of things, so they already have a structure in their heads.
Ji Yichao comes across as open, curious, and pretty clear about himself — he knows what he's good at and what he's not. Yang Zhilin is more rational, and his emotions seem very stable.
Honestly, what I envy more is their domain expertise. I think logic and expression can both be practiced, but really knowing a field, and going deep enough into it — that comes gradually.
The point in the interview that resonated with me the most was that products should try to stay orthogonal to the model. Simply put, the stronger the model gets, the stronger your product should get too — not that your product becomes useless the moment the model upgrades.
As soon as I heard that, I immediately thought of Linglan, something I built about six months ago.
Linglan was a medical knowledge chat webpage I built when I was first Vibe Coding. Back then I used DeepSeek plus Volcano Engine's vector knowledge base, fed our textbook PDFs into it, and used a system prompt to make it handle pre-study prep, mind maps, review, and memorization rhymes.
The webpage and the RAG setup were AI-assisted; the textbooks and the medical study methods were things I provided. At the time I felt this had real demand, because the medical knowledge general-purpose models give you often didn't quite match up with what we were actually studying in our textbooks.
Later, an older guy told me that college students don't have much willingness to pay, and this product wasn't really that different from Doubao. Users could just write their own prompt and get a general-purpose model to reference the textbook themselves.
At first I was actually a bit unwilling to accept that — I felt like there was clearly a market for this.
Looking back now, he was right. Linglan was just a demo — no real users, never really got running. If I'd kept building it for another month, I probably would've just ended up with a half-broken product full of problems.
And as models kept getting stronger, a lot of what Linglan did back then would genuinely get swallowed up by general-purpose models over time.
That doesn't mean vertical products are useless — it just means you can't only build the small bit that models temporarily can't do yet. In the end, the model, the tools, and the workflow all have to come together to solve a real problem.
Another thing I hadn't thought of before was the niche Manus found for itself.
A lot of vertical Agents are built for professionals, but professionals already have their own methods, and their standards are extremely high. Get one detail wrong, and they might decide the whole thing is unusable.
Manus targets a different group of people — they're not professionals in the field, but they do have real needs. They can't design, but they need design done. They can't do research, but they need research done. They can't code, but they genuinely want to build something.
I think that's a really well-targeted position.
It's much the same in medicine. A real doctor definitely won't let an unreliable general-purpose Agent make judgment calls for them, but for an ordinary person using an Agent to look up general medical knowledge, that's often still better than just guessing blindly. Of course, the safety boundaries in medicine absolutely have to be hard-coded — you can't rely on prompts alone for that.
This is also where I don't fully agree with pure-Agent approaches. Models are already very strong now, but once the context gets long, earlier constraints can still get diluted, and hallucinations can still happen.
So my thinking is: whatever needs to be hard-coded should be hard-coded. Leave the safety boundaries to the program, and leave things like figuring out what the user actually wants, which tool to pick, and how to judge between different options to the model.
As for multi-agent setups, I'm also not a big fan of splitting them directly along the lines of human company job roles — a product manager Agent, a programmer Agent, a designer Agent. Sometimes it feels like multi-agent for the sake of multi-agent.
I'd rather split by context. Research eats up a lot of context, so it gets pulled out on its own. Research first, then coding, then review after coding. That's something I can talk about separately another time.
After finishing the whole interview, I thought again about what an Agent is actually amplifying.
I used to think of it as an amplifier with a fixed multiplier. If you start at 10, it turns you into 100. If someone else starts at 100, it turns them into 1000.
Now I think what it amplifies might not be ability, but the value a person was already capable of creating.
Without AI, I probably would've had to spend a long time learning to program first, and might still not have managed to build Linglan. AI genuinely let me get a demo out first.
But hand the same project to a professional engineer, and if they also use AI, they'll definitely produce a much higher-quality version in a lot less time. Same Claude — an Anthropic engineer using Claude and me using Claude will definitely end up creating different amounts of value.
So AI lets more people get into the game, but it won't make everyone equally good.
If anything, I now feel like I still need to fill in the fundamentals. It's not that I need to train myself into a traditional programmer, but I do need to know at least something about Git, worktrees, environment variables, testing, logs — otherwise, when an Agent builds something, I won't even be able to judge whether it actually works.
Weak spots can be handed off to AI, or to someone more professional, but I can't be weak in every single area myself.
So in the end, this gap day didn't lead me to any grand conclusions.
Once Claude's back up, I'll definitely keep using it — but the stuff I need to learn myself, I still need to keep learning.