Artificial Intelligence thread

E

Eventine

Junior Member
Registered Member
Compared to Anthropic's Claude Sonnet 3.5, that's a ~7x price advantage, but it should be remembered that Sonnet 3.5 is not a reasoning model, and as such it produces fewer output tokens (since the thought chain is not included). Sonnet is mostly used by coders (and is probably the best model for coding) via intermediate services like Cursor, and so there is still an argument to be made that Sonnet is competitive with R1, which probably explains why it remains at the top of the Open Router list.

A Deep Seek v4 that beats Sonnet 3.5 in coding (but is still cheaper to run) should be the next goal. Anthropic is the least open of all the Big AI companies in the US and also actively campaigns for chips embargoes on China. Crushing them will not only be commercially and technologically useful, but would also be a huge publicity win.
 
tphuang

tphuang

General
Staff member
Super Moderator
VIP Professional
Registered Member
https://twitter.com/i/web/status/1886163621849202807

This guy has some good investigation and post here. I wouldn't accept everything he says, but the parts about pricing on H100 and H800 and how they've dropped in the past few months indicate the glut of hopper chips that are available in China.

This is from October, the H100 rental prices were plummeting across China.

Please, Log in or Register to view URLs content!

demand for this is simply not as high as people think.
 
Last edited:
E

Eventine

Junior Member
Registered Member
GulfLander said:
Pku graduate

https://twitter.com/i/web/status/1886216911777919257
....
"SoftBank and OpenAI to set up AI joint venture in Japan
Masayoshi Son's group to pay $3bn annually to use new Cristal services"
Please, Log in or Register to view URLs content!
Click to expand...
The final piece of the puzzle towards AGI is theorized to be embodied agents using RL in the “real world” to self improve and self evolve (e.g. modifying its own network architecture and learning algorithm to get better at its set tasks, and then eventually learning to define its own tasks to get better at the ultimate goal of being able to answer any question and achieve any goal).

Deep Research (as well as R1-Zero) are getting closer; but they’re just the public releases. Behind closed doors, I wonder how much closer we are. It’s starting to feel like it could happen in the next 20 years.
 
FairAndUnbiased

FairAndUnbiased

Brigadier
Registered Member
european_guy said:

DeepSeek Becomes Top AI App, India Leads the Way in User Adoption​


Please, Log in or Register to view URLs content!




Well... :)

It seems these guys provide DeepSeek here:

Please, Log in or Register to view URLs content!

Apparently they have their own cloud infrastructure on Intel Gaudi CPus:

Please, Log in or Register to view URLs content!

But they also partner with AWS, so it is not clear if the servers for DeepSeek are actually in India or they simply act as front interface for AWS.
Click to expand...
The best shot for India to become independent in AI is for them to go 100% all in on the Deepseek-Huawei ecosystem and learn everything they can about it. They must fully humble themselves. They must drop all of their ego and compete with Pakistan on how to best improve their value to China.

They will soon be faced with a choice of geolocked and paywalled systems, and open source, free, transparent systems. If India wants to survive, this is their only hope.
 
O

OptimusLion

Junior Member
Registered Member
The first in China: Baidu successfully lit up the Kunlun Core 3rd Generation 10,000-card cluster, and will also light up 30,000-card clusters

Baidu Smart Cloud announced today that it has successfully lit up the Kunlun Core 3rd Generation 10,000-card cluster, which is also the first self-developed 10,000-card cluster officially lit up in China. Baidu Smart Cloud will further light up 30,000-card clusters


IMG_20250205_084350_982.jpg

Please, Log in or Register to view URLs content!
 
You must log in or register to reply here.
Top