More than a decade ago, I used 3G WCDMA for the first time. Yes, WCDMA, not TD-SCDMA.
I remember cycling away from a China Unicom store, pulling out my Windows Phone, and scrolling through Weibo. The page loaded almost instantly. I still remember that feeling. Not because there was anything important on Weibo, but because 2G had been so painfully slow. You waited for pictures. You waited for web pages. You waited for everything. Then one day, suddenly, you didn’t have to.
Back then, I had no idea that short-form video, livestream shopping, mobile payments, food delivery, and ride-hailing were coming. If someone had asked me what 3G’s killer app would be, I wouldn’t have had an answer.
But I knew one thing: there was no going back to 2G.
Lately, using AI has often given me a similar feeling.
From chat, to GPTs with Instructions + Tools, to Manus with its cloud sandbox, AI started browsing the web, working with files, and running on its own for a while. Claude Code and Codex plugged in the terminal, codebase, and Git. OpenClaw added channels, memory, and persistent execution. Now everyone is rushing toward personal agents.
The direction is pretty clear: we keep giving AI more things.
More tools, broader permissions, more persistent environments. A computer, a browser, a file system. Then email, calendars, contacts, accounts, and payments.
What task justifies an agent being around 24 hours a day? What can sustain a repeatable business loop over time? Nobody has quite figured that out yet. Everyone just feels it must be incredibly important.
I think there’s another very practical reason personal agents are hot again: chat and coding agents may not make enough money, at least not enough to justify the valuations and investment pouring into AI.
Chat has lots of users, but every conversation keeps burning inference costs. Coding agents are closer to productive work, and people are willing to pay for them. But they still may not be enough to support expectations of the next computing platform.
So the search moves outward. Into more workflows and more transactions, gathering more personal information and creating stronger user lock-in.
Personal agents are both an expansion of capability and an expression of business-model anxiety. They haven’t been proven to be the answer. But everyone needs a bigger answer.
Of course there’s a bubble in AI right now. A lot of valuations aren’t pricing today’s cash flow. They’re pricing the question: “What if AI really is the next Industrial Revolution?”
So everyone has FOMO. Investors are afraid of missing out. Founders are afraid of missing out. Big companies are afraid of missing out. The rest of us wage slaves are afraid we’ll wake up as yesterday’s workforce. Money and companies are pouring in, while only a handful of business models have actually been validated.
A bubble isn’t necessarily a bad thing, though. When nobody knows the answer, it acts as an accelerator. Capital subsidizes compute, models, and a whole lot of people trying all sorts of possibilities. Most of it will die. But some things that might have taken ten years to try could get tried in three.
That’s the contradiction in this industry right now: there are already plenty of very good AI products, but not nearly as many very good AI businesses.
Models are expensive. Agents are even more expensive. Vertical applications could be swallowed by the next generation of foundation models at any moment. Many of today’s so-called moats may just be scaffolding built around what current models can’t yet do.
Still, I look forward to seeing what pops up each week. You couldn’t have extrapolated TikTok and livestream shopping from “Weibo loads faster now.” How AI will change white-collar work, and when, I don’t know either. But once you’ve used it, going back really does get harder.
The bubble will probably burst, too. Just try not to land on your head.