What Large Models Need Is a Lei Jun

"Tents Can Fly"

At this stage of the foundation model arms race, what we desperately need is the kind of product definition capability Xiaomi shows with the Pengcheng and its built-in tent.

Adding a tent to an SUV is neither technically sophisticated nor a large share of the cost, yet it carries enormous weight in users' minds. A few nights ago I was up late seriously browsing Xiaomi's Pengcheng website, even configuring my own build — only because I can't afford it. I usually ride my Ninebot, which is very cool.

But I was genuinely tempted. The whole time, I never once looked at technical specs like how many seconds it takes from 0 to 100.

After years of brutal competition, EVs have moved past the stage of pure spec wars — at the same price point, the technical metrics don't differ much anymore. Users are numb to it.

The same goes for foundation models, which are also past the pure parameter-war stage. Fable 5.1's improvements were far from awe-inspiring. And after these past few months of leaps by Chinese models, people's sensitivity to raw parameter counts and benchmark scores is fading too.

Take coding leaderboards first. Scores there have practically become talismans for Chinese model launches — teams will stop at nothing to squeeze out a few more points, incidentally feeding a whole crop of data companies that specialize in targeted distillation for them.

These data companies, with their main business of distillation outsourcing for model labs and a side hustle of reselling test questions, have lately been busy preaching RSI — which is itself a kind of RSI, isn't it?

As for frontend websites, the domestic scene has been completely dominated by Moonshot AI's Kimi. When a newly released model claims its frontend is amazing, nobody cares anymore — everyone has particle effects now. And that one is a ripoff of Lxson's website; they should consider suing.

Then there are the multi-dozen-hour long-horizon tasks that everyone keeps showing off. Beyond letting consumer users exclaim "holy cow, AI," nobody actually uses them.

So here's where we are: every newly released big-parameter model claims to be world number three, every Flash-class model claims to be a killer threshold — it gives off the feeling of the third world endlessly hacking at each other.

Of course, this assessment only applies to the few players stretching for the ceiling. The second-, third-, and fourth-tier players who want a seat at the table should first honestly finish implementing the capabilities others have already defined.

What I think foundation models need is some hardcore stuff like a built-in tent. While every EV maker fought over cabin space, Lei Jun told everyone to look to the sky.

Just fly ✈️

But a built-in tent isn't easy to pull off — it's a systems engineering effort. What won me over was a whole set of features: the car has floor heating, so it's more comfortable than running the AC in winter; the tent has AC piped in too; and there's dedicated storage space for water.

You need someone to clearly define user needs, someone who can tie the engineers down — and have them happily, voluntarily tied down — teaching them that serving user needs well is the road to AGI, with everyone working together to nail down the supply chain.

GPT-6 Astra is actually doing the same job as Lei Jun. Its 3D capability is honestly mediocre — its generated human figures are a mess, nowhere near professional AI 3D models. But combined with its computer-use ability, it's very powerful; it manipulates tool software well. That's why the case of GPT-6 using Blender to build 3D models spread so widely.

One more thing here: GPT-6's computer-use ability really is strong. It can fully handle editing short videos from solid to rough cuts. Right now I'm using Codex to operate CapCut for editing — the final cuts basically only have some audio splice issues. A human just needs to review the cut and give feedback, and telling it to look at the waveform itself solves the rest. I've basically said goodbye to manually operating CapCut.

It's foreseeable that all the Chinese model makers will follow up on 3D and computer-use capabilities, trying to grab big results alongside OpenAI. It just takes time — what I mean is, distillation is a process. This isn't urgent, but it has to be fast.

Still, I keep feeling this is an area where independent innovation is possible.

GPT-6's emphasis on 3D is just picking low-hanging fruit — reaping the dividend from a space other models haven't optimized much. Chinese models can absolutely do the same thing. This is a product-and-technology-driven question, not a purely technical one.

All in all, compared to Luofu Li, foundation model startups need a Lei Jun more. I suggest that next time Xiaomi releases a new model, Lei Jun should host the launch event himself.

I want to hear Lei Jun talk again.

(This article's cover image was generated by ChatGPT; the writing is purely human)

⬇️

Subscribe to our Substack: funeralai.substack.com