kkzz99

You may not know, but China has also adopted pretty elaborate privacy laws called Personal Information Protection Law(PIPL) which is pretty close to GDPR.

The blatant white-washing of corrupt oligarchs by the usual suspects is truly off the charts.

How does that work when the card only has 230MB of SRAM?

How does the Groq PCIE Card work exactly? Does it use system ram to stream the model data to the card? How many T/s could one expect with e.g. 36000Mhz DDR4 Ram?

I don't think this is the argument you think it is...

American weapons, American ISR, American planning, American funding... US politicians talk about a "good investment" and "to the last Ukrainian". I don't understand how you could NOT call it a US-proxy war.

Its because they don't want you to actually use it and see how far behind they are compared to other companies. These announcement are meant to placate investors. "See, we are doing a lot of SotA AI too".

You have to specify the amount of layers to put on the GPU with ollama. Ollama defaults to far less layers compared to what is actually possible.

You use the vector embedding only for retrieval of the actual text, which then gets injected into the prompt.

Huh, why would using uuidv4 be a problem? Collisions?

RTX 3090 has 24GB of memory, a quantized llama70b takes around 60GB of memory. You can offload a few layers on the gpu, but most of them will run on the CPU with terrible speeds.

Which is very ironic, when you compare 1st tier chinese cities with 1st tier US cities, you could probably say the same thing.

Correct, the US is not a reliable partner and should be viewed very warily.

Market is more dominated by Adidas.

Mistrals biggest risk factor is the EU. The overzealous regulation could be stifling innovation, especially open source.

Doesn't matter what hardware they release if the software doesn't support it.

Do these AI accelerators even support common implementations for LLM inference or stable diffusion models?

Last time I checked, AMD's support was terrible.

No matter how many times I clicked the back button it wouldn't let me return to hn. Its a terrible implementation.

China filed more IP than rest of world combined in 2022. There is no reason to keep it secret with IP-Laws and Patents. https://www.chinadaily.com.cn/a/202312/12/WS6577b803a31040ac...

For Chinese EV insights I recommend https://twitter.com/tphuang

It was primarily the USSR that liberated us from fascism, while the USSR stopped its occupation of east Germany, the US never stopped. And wasn't it Christina Nuland that said "Fuck the EU" while instigating the Ukraine Coup? Wonder why we should trust the US when it has been one of the worst "Allies" we ever had...

In my opinion, the US is as much a foreign adversary as China. The US isn't subsidizing defense, its occupation and hegemony.

I mean, sounds like you haven't kept up with Chinese EV innovation...

Doesn't really look like a cooldown cycle for me in China. The competition is extreme there. Weekly new releases with another cool state-of-the-art feature. Pure innovation going on there, which makes me, as a German, kinda worried...

This is no different then me being a German owning an Android or Iphone, which is going to ping back home to some servers in the US.

Usually, for efficiency, you use quantized models. Quantized models reduce the number of bits available for each parameter, saving space and reduce RAM usage.

How are yearly interest payments of 1tn$USD sustainable?

Will that not change with the new EU rules?

We are talking about very big models which training requires an enormous amount of hardware. Not sure how scalable HuggingFace transformer for training such models is.

Looks like hardware vs. software abstraction. Considereing the perspective of an LLM startup: Would you rather write 20k LOC of complex code that would make you be able to more easily switch hardware platforms - or - write 600 LOC of less complex code and be pinned to a single provider?