[zamalek]

You misunderstood my comment. My hypothesis is that it did _neither:_ it accelerated along an axis, and American SOTA/frontier is now laggard (efficiency).

Deepseek and Kimi are writing paper after paper with substantial architecture improvements for efficiency, because they can't just throw more hardware at the problem.

And China is now doing something on the hardware axis; which it may have never explored were it not for the sanctions.

[throwup238]

> And China is now doing something on the hardware axis; which it may have never explored were it not for the sanctions.

Gaining parity on the semiconductor fab front has been official government policy as part of their Five Year Plan for at least the last decade, straight from the Politburo. They were always going to go down this path, and with AI playing front and center on their upcoming plan, there’s even more pressure.

There was never a possibility of them not exploring it.

[throwa356262]

The difference is that this time the both the government and the (semi) private sector invested billions in this area.

They have always been able to do this, but this time they did have the option to pass.

[throwup238]

There’s an ambiguity here: what is “this area”? If we’re talking about the semiconductor industry, they’ve been investing billions since the early nineties. The really big push came with the China Integrated Circuit Industry Investment Fund in 2014 at which point they were plowing billions into the industry across the public and private front. That one fund alone has $100 billion in assets under management at this point, and there are many other funds involved.

They’ve also invested in AI separately (before LLMs) in that time period but I’m less familiar with that sector.

[aspenmartin]

Oh no Chinas chip manufacturing efforts have a long and rocky history, they would be pursuing a sovereign stack no matter what.

The tradeoff is worth it. They’re even publishing papers which blows me away — their efficiency gains quickly become incorporated into frontier models because they are open sourcing them. They would be aggressively pursuing the same chip pipeline strategy as they are today.

US is lagging in efficiency work because the ROI is better elsewhere for us. We have the same tier of talent, once the script flips so can the research.

[linzhangrun]

LongCat 2.0, 1.6T-scale, released yesterday, was trained on Huawei's GPUs.

DeepSeek is doing similar work too.