This, it's been my experience since the early 2000s that western forums are just.... dumber.
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
Why do the Huawei Ascend chips have lower interconnect speed between multiple GPUs. What is the bottleneck there? Interconnects do not need lower than 7nm i presume.
Seems India is quite happy with DeepSeek open free model. Since it now enables India to get into the AI game for little cost which was something that was almost insurmontable for them before this.
At least they can also get in the game now.
The Chinese language itself definitely plays a role here, given how transformers and attention (basic building blocks of LLM) work. But the liveliness of available Chinese content, i.e., how creatively the Chinese people use the language, might be an even bigger factor.
So I've been thinking, we know Deepseek is the side project of HiFly fund and the founder had to pitch the idea to his partners. But what was his pitch?
Is it 1: I want to spend our resources (idle or not), pay for the extra power bill, redirect our existing people and hire new talent, so we can develop the worlds best model and give it to everyone for free, because we are a hedge fund and charity is our thing.
Or is it 2: same as one, but add "Americans are in an AI bubble and this will allow our fund to make more money than ever by shorting it"
Is it 1: I want to spend our resources (idle or not), pay for the extra power bill, redirect our existing people and hire new talent, so we can develop the worlds best model and give it to everyone for free, because we are a hedge fund and charity is our thing.
Or is it 2: same as one, but add "Americans are in an AI bubble and this will allow our fund to make more money than ever by shorting it"
I am somewhat familiar with this space. Short answer is that hedge funds are very willing to pay for research to find edge in the markets.
Some of the best weather models, for example, are at certain hedge funds. They pay meteorologists and physicists top dollar to help them generate edge in the commodities markets.
Wall street has also been using NLP models on earnings data for decades, so it’s not like LLMs are that big of a leap from normal hedge fund research. Furthermore, I have read that high flyer already has been using AI/ML techniques in their strategies already, so it probably wasn’t that big of a sell, especially since idle GPUs are considered a loss due to depreciation.
The whole “make money by releasing model then short the markets” seems like a big reach to me.
It's not creating the model to create competitive advantage that's interesting
It's releasing the model so everyone, including other hedge funds can share in your advantage, which nullify your advantage, that's interesting.
Nations and organizations that don’t adopt will be behind. This is why you see even US corporations adopting it despite the PRC stigma.