Artificial Intelligence thread

[tphuang]

And good news for all the users, DeepSeek has finally gotten it worked out with inference to Zhejiang data center.

https://twitter.com/i/web/status/1894429573124624677

now, it's actually up, latency is down and it's just available.

 

[Wrought]

But remember that comparative advantage doesn't work if:
1. China has an absolute advantage in both manufacturing and service industries
2. Plus the spare resources to fully apply AI to both manufacturing and service industries

Comparative advantage only applies if both sides are limited or face resource constraints

???

That's exactly the opposite of Econ 101. Comparative advantage applies even when one country has an absolute advantage across the board.

Why countries trade​

Ricardo observed that trade was driven by comparative rather than absolute costs (of producing a good). One country may be more productive than others in all goods, in the sense that it can produce any good using fewer inputs (such as capital and labor) than other countries require to produce the same good. Ricardo’s insight was that such a country would still benefit from trading according to its comparative advantage—exporting products for which its absolute advantage was greatest and importing those for which its absolute advantage was comparatively less.

Please, Log in or Register to view URLs content!

 

[AndrewS]

???

That's exactly the opposite of Econ 101. Comparative advantage applies even when one country has an absolute advantage across the board.

That assumes there is a constraint that stops the country from producing and/or exporting

---

On the production side, China has spare capacity and the ability to ramp up in every industry.

---

On the export side, in the old days, if you exported a lot:

1. You ended up with a strong currency which naturally dampened exports. But now, governments can keep the exchange rate weak by printing their own fiat currency and using it to buy foreign currency/assets.

2. Also, the importing country would previously struggle to pay for those imports with hard currency. But now? You have fiat currency (USD, etc) which are printed and are generally accepted everywhere.

 

[Wrought]

That assumes there is a constraint that stops the country from producing and/or exporting

---

On the production side, China has spare capacity and the ability to ramp up in every industry.

---

On the export side, in the old days, if you exported a lot:

1. You ended up with a strong currency which naturally dampened exports. But now, governments can keep the exchange rate weak by printing their own fiat currency and using it to buy foreign currency/assets.

2. Also, the importing country would previously struggle to pay for those imports with hard currency. But now? You have fiat currency (USD, etc) which are printed and are generally accepted everywhere.

No it doesn't, it just assumes that:

1) Resources are finite.

2) Absolute advantage doesn't have a uniform distribution (and opportunity cost is therefore uneven).

Both of which are, yknow, always true in reality. This is an extremely basic concept in economics—given finite resources and relative efficiencies, making more of X means making less of Y. Which is why you trade for one of them, so you can have more of both.

 

[OptimusLion]

DeepSeek Open Source Week Day 3: DeepGEMM.

DeepGEMM is a library designed for concise and efficient FP8 general matrix multiplication (GEMM) with fine-grained scaling, as proposed in DeepSeek-V3. It supports both normal and Mixed of Experts (MoE) grouped GEMMs. The library is written in CUDA and does not require compilation during installation, compiling all kernels at runtime using a lightweight Just-in-Time (JIT) module.

Currently, DeepGEMM only supports NVIDIA Hopper tensor cores. To address the problem of inaccurate accumulation of FP8 tensor cores, it adopts a two-level accumulation (lifting) method of CUDA cores. Although it borrows some concepts from CUTLASS and CuTe, it avoids heavy reliance on their templates or algebra. Instead, the library is designed to be concise, containing only one core kernel function with about 300 lines of code. This makes it a clear and accessible resource for learning Hopper FP8 matrix multiplication and optimization techniques.

Up to 1350+ FP8 TFLOPS on Hopper GPU
No heavy dependencies, simple as a tutorial
Fully JIT compiled
~300 lines of core logic - but still outperforms expert-tuned kernels on most matrix sizes
Supports dense layout and two MoE layouts

Please, Log in or Register to view URLs content!