2020: I started toying with GPT-3. Getting it to craft a poem or churn out a catchy paragraph felt incredible. I thought this was real AI (how naive).
Then ChatGPT burst onto the scene in November 2022, together with all the hype. I thought, this is real AI (and I'm sure I'm not the only one who felt amazed).
Ever since, I've thrown various LLMs into my daily workflow. Feeding them legacy SQL and ETL code and asking for neat, refactored solutions for modern tools became the new normal. The productivity boost is tough to quantify precisely, but the results are undeniably real. Yes, they can be frustratingly wrong sometimes, but they've turned weeks of work into days in at least dozens of cases, or at least got me over the psychological hump to get moving. Real-ish, for sure.
Enter Claude Code and its competitors like Gemini CLI: writing code, deploying directly to Cloudflare, auto-fixing bugs, committing neatly to GitHub, all from a few simple prompts. Is this real AI?
But nothing prepared me for this week. I'm one of 300 thousand people attending WAIC, the World Artificial Intelligence Conference in Shanghai, and what I've seen here is AI stepping out from behind screens into the physical world. AI embodiment.
Robots of every imaginable size and function: playing musical instruments, giving massages, packing boxes, helping elderly people.
Out on Shanghai's streets (which, by the way, are impressively spotless), I caught sight of robotaxis, while drone delivery is progressing between trials and commercial deployment. Someone said around 60% of vehicles in Shanghai are EVs; judging by their plates, that's no exaggeration.
Even my hotel, which is comfortably average, greeted me with a robot casually rolling into the lift alongside humans as though it's the most normal thing in the world. Turns out room-service robots are practically standard here.
At Tencent HQ I saw, among many things, a console that lets someone drive a real truck remotely, obviously for a mining site. I also heard Ian Tong talk about enterprise AI versus consumer AI and its lack of governance. Spot on, and exactly what's been bothering me about consumer AI.
And in the toilets… well, let's not go there. I can't help noting the control panel though: the lid opens and closes automatically. Somehow I felt a bit eerie and certainly prefer my non-smart toilet back in Melbourne.
Overall, what I see in China: deploy AI at scale → fail at scale → learn at scale → progress rapidly. Elsewhere the case may be: talk at scale → limited deployment → limited learning → slow progress?
Massive thanks to Xue Huang and MLAI, who made this eye-opening trip possible.
First posted on LinkedIn. This is the canonical copy, lightly tidied.