Artificial Intelligence thread

[european_guy]

https://twitter.com/i/web/status/1883228769701290440

inference cost for DeepSeek according to various host. Notice how DeepSeek's own server is so much lower?

I have checked chat.deepseek.com on netcraft

Please, Log in or Register to view URLs content!

And it's on Cloudflare, so a US CDN provider.

Is not clear where the actual servers are, maybe they are in part out of China, maybe rented somewhere....

For instance there is this Finland GPU provider,

Please, Log in or Register to view URLs content!

that helped (by providing GPUs) SGLang to integrate Deepseek

https://twitter.com/i/web/status/1872251875070021831

Special thanks to Meituan's Search & Recommend Platform Team, Baseten's Model Performance Team for implementing the model, and DataCrunch for providing GPU resources

I'm not claiming that they run the official DS, just that is not clear if all the official DS instances are run on Chinese servers.

I have used DS quite a bit in these days, and it's always very responsive. It is a bit of a surprise considering that probably there are tons of people doing the same from all over the world....including in China itself!

...but at the moment it is not forbidden for Western providers to run Deepseek....but if the other Western actors will not close the performance gap in the next 2/3 months I'm not sure they will still allow providers to serve DS in this part of the world....we already now how it works here with "competition": competition is good as long as is not real, and especially as long as US is in the driver seat.

Enjoy your DS now that you still can!

 

[tphuang]

That is quite possible. Maybe they are renting resource from abroad.

China has come out with its AI Development Action Plan.

https://twitter.com/i/web/status/1883264094884536747

DeepSeek is now #10 app in iOS in America

https://twitter.com/i/web/status/1883252520967156041

 

[Xiongmao]

That is quite possible. Maybe they are renting resource from abroad.

China has come out with its AI Development Action Plan.

https://twitter.com/i/web/status/1883264094884536747

Hold up, is this plan just for one single bank, i.e. Bank of China?

 

[luminary]

Bytedance moving to catch up.

Please, Log in or Register to view URLs content!

Currently offers unlimited free access to GPT-4o and Claude-3.5-Sonnet, with support for VSCode configurations and plugin imports. Currently only available for macOS, with Windows version under development.

I believe it is free compared to cursor.ai

 

[Fatty]

So I found this comment on Reddit about Chinese being a more efficient for AI:

“
I heard from a scientist in China that their models are faster and borderline superior with less data because they used chinese as internal language for their LLM's it seems chinese (or "mandarine") has some superior language properties for chain of thoughts, grouping of topics and implicit reasoning. Its fascinating because western languages are "limited languages" so their whole databases (based on German, France, English etc.) is less effective. They expanded topics by letting the LLMs explain topics in complex chinese and use that dataset for retraining. The next logical step would be to find the ultimate symbolic language.

Chinese / Mandarine has like 50.000 symbols, average mainland citizen is using 1.000 to 2.000 symbols but the language has the "build in" property to expand certain topics as far as possible - so you can describe whole domains in its complexity. ca. 3000 symbols cover 98% of everything in a life of a modern chinese citizen (blogs, newspapers, decrees, etc) - the rest, the rarely used 2% or ca. 47.000 symbols cover very specific topics in science, literature and art...

F.e. biology in English you would use the term "cell" but it has many use cases as a term (cell membrane, battery, compartment etc.), in chinese you have specific characters that really mean (biological) cells and nothing else. For example, the character "细胞" (xìbāo) means "cell" in biology, and it is different from "电池" (diànchí), which means "battery" (in a physics/energy context).

For some reason thats easier for LLMs - super complex character set but super exact at the same time.

This may be the reason why chinese companies speed-up with smaller datacenters and shorter trainings.”

Someone also pointed out that this theory about more data dense languages being more efficient has actually been tested and shown here:

Please, Log in or Register to view URLs content!

Wondering if @tphuang can shed some more light on this, but if true, it’d be interesting if mandarin became de-facto required for AI research.

 

[siegecrossbow]

So I found this comment on Reddit about Chinese being a more efficient for AI:

“
I heard from a scientist in China that their models are faster and borderline superior with less data because they used chinese as internal language for their LLM's it seems chinese (or "mandarine") has some superior language properties for chain of thoughts, grouping of topics and implicit reasoning. Its fascinating because western languages are "limited languages" so their whole databases (based on German, France, English etc.) is less effective. They expanded topics by letting the LLMs explain topics in complex chinese and use that dataset for retraining. The next logical step would be to find the ultimate symbolic language.

Chinese / Mandarine has like 50.000 symbols, average mainland citizen is using 1.000 to 2.000 symbols but the language has the "build in" property to expand certain topics as far as possible - so you can describe whole domains in its complexity. ca. 3000 symbols cover 98% of everything in a life of a modern chinese citizen (blogs, newspapers, decrees, etc) - the rest, the rarely used 2% or ca. 47.000 symbols cover very specific topics in science, literature and art...

F.e. biology in English you would use the term "cell" but it has many use cases as a term (cell membrane, battery, compartment etc.), in chinese you have specific characters that really mean (biological) cells and nothing else. For example, the character "细胞" (xìbāo) means "cell" in biology, and it is different from "电池" (diànchí), which means "battery" (in a physics/energy context).

For some reason thats easier for LLMs - super complex character set but super exact at the same time.

This may be the reason why chinese companies speed-up with smaller datacenters and shorter trainings.”

Someone also pointed out that this theory about more data dense languages being more efficient has actually been tested and shown here:

Please, Log in or Register to view URLs content!

Wondering if @tphuang can shed some more light on this, but if true, it’d be interesting if mandarin became de-facto required for AI research.

I’ve been suggesting for a while that 文言文 would make the best prompting language, if we can get enough people good enough at it to be practical.

 

[littleicebook2]

So I found this comment on Reddit about Chinese being a more efficient for AI:

“
I heard from a scientist in China that their models are faster and borderline superior with less data because they used chinese as internal language for their LLM's it seems chinese (or "mandarine") has some superior language properties for chain of thoughts, grouping of topics and implicit reasoning. Its fascinating because western languages are "limited languages" so their whole databases (based on German, France, English etc.) is less effective. They expanded topics by letting the LLMs explain topics in complex chinese and use that dataset for retraining. The next logical step would be to find the ultimate symbolic language.

Chinese / Mandarine has like 50.000 symbols, average mainland citizen is using 1.000 to 2.000 symbols but the language has the "build in" property to expand certain topics as far as possible - so you can describe whole domains in its complexity. ca. 3000 symbols cover 98% of everything in a life of a modern chinese citizen (blogs, newspapers, decrees, etc) - the rest, the rarely used 2% or ca. 47.000 symbols cover very specific topics in science, literature and art...

F.e. biology in English you would use the term "cell" but it has many use cases as a term (cell membrane, battery, compartment etc.), in chinese you have specific characters that really mean (biological) cells and nothing else. For example, the character "细胞" (xìbāo) means "cell" in biology, and it is different from "电池" (diànchí), which means "battery" (in a physics/energy context).

For some reason thats easier for LLMs - super complex character set but super exact at the same time.

This may be the reason why chinese companies speed-up with smaller datacenters and shorter trainings.”

Someone also pointed out that this theory about more data dense languages being more efficient has actually been tested and shown here:

Please, Log in or Register to view URLs content!

Wondering if @tphuang can shed some more light on this, but if true, it’d be interesting if mandarin became de-facto required for AI research.

so modi saying sanskrit is the best language for AI isn't true?

 

[Temstar]

So I found this comment on Reddit about Chinese being a more efficient for AI:

“
I heard from a scientist in China that their models are faster and borderline superior with less data because they used chinese as internal language for their LLM's it seems chinese (or "mandarine") has some superior language properties for chain of thoughts, grouping of topics and implicit reasoning. Its fascinating because western languages are "limited languages" so their whole databases (based on German, France, English etc.) is less effective. They expanded topics by letting the LLMs explain topics in complex chinese and use that dataset for retraining. The next logical step would be to find the ultimate symbolic language.

Chinese / Mandarine has like 50.000 symbols, average mainland citizen is using 1.000 to 2.000 symbols but the language has the "build in" property to expand certain topics as far as possible - so you can describe whole domains in its complexity. ca. 3000 symbols cover 98% of everything in a life of a modern chinese citizen (blogs, newspapers, decrees, etc) - the rest, the rarely used 2% or ca. 47.000 symbols cover very specific topics in science, literature and art...

F.e. biology in English you would use the term "cell" but it has many use cases as a term (cell membrane, battery, compartment etc.), in chinese you have specific characters that really mean (biological) cells and nothing else. For example, the character "细胞" (xìbāo) means "cell" in biology, and it is different from "电池" (diànchí), which means "battery" (in a physics/energy context).

For some reason thats easier for LLMs - super complex character set but super exact at the same time.

This may be the reason why chinese companies speed-up with smaller datacenters and shorter trainings.”

Someone also pointed out that this theory about more data dense languages being more efficient has actually been tested and shown here:

Please, Log in or Register to view URLs content!

Wondering if @tphuang can shed some more light on this, but if true, it’d be interesting if mandarin became de-facto required for AI research.

Chinese being data dense I'm pretty sure is true and people have studied this in the past by comparing thickness of documents at the UN where each item has to be separately translated into each working language of the UN and Chinese documents were found to be consistently thinner than the rest. 文言文 in particularly is extremely data dense hence why when you read it it takes non-trivial amount of brain processing power to decipher what it's saying even for someone fluent in regular Chinese. 文言文 evolved in that direction since writing material back in the day was either expensive (paper) or heavy (bamboo) so squeezing meaning into as few word as possible was beneficial. Ancient Chinese people didn't actually speak straight 文言文 in every day life, it was only for writing.

 

[pbd456]

So I found this comment on Reddit about Chinese being a more efficient for AI:

“
I heard from a scientist in China that their models are faster and borderline superior with less data because they used chinese as internal language for their LLM's it seems chinese (or "mandarine") has some superior language properties for chain of thoughts, grouping of topics and implicit reasoning. Its fascinating because western languages are "limited languages" so their whole databases (based on German, France, English etc.) is less effective. They expanded topics by letting the LLMs explain topics in complex chinese and use that dataset for retraining. The next logical step would be to find the ultimate symbolic language.

Chinese / Mandarine has like 50.000 symbols, average mainland citizen is using 1.000 to 2.000 symbols but the language has the "build in" property to expand certain topics as far as possible - so you can describe whole domains in its complexity. ca. 3000 symbols cover 98% of everything in a life of a modern chinese citizen (blogs, newspapers, decrees, etc) - the rest, the rarely used 2% or ca. 47.000 symbols cover very specific topics in science, literature and art...

F.e. biology in English you would use the term "cell" but it has many use cases as a term (cell membrane, battery, compartment etc.), in chinese you have specific characters that really mean (biological) cells and nothing else. For example, the character "细胞" (xìbāo) means "cell" in biology, and it is different from "电池" (diànchí), which means "battery" (in a physics/energy context).

For some reason thats easier for LLMs - super complex character set but super exact at the same time.

This may be the reason why chinese companies speed-up with smaller datacenters and shorter trainings.”

Someone also pointed out that this theory about more data dense languages being more efficient has actually been tested and shown here:

Please, Log in or Register to view URLs content!

Wondering if @tphuang can shed some more light on this, but if true, it’d be interesting if mandarin became de-facto required for AI research.

If it is true, it wouldnt really matter as other languages can be translated to a denser language before being processed by models?

 

[OptimusLion]

Baichuan Intelligent launches open source omnimodal model Omni-1.5, claiming that multiple capabilities surpass GPT-4o mini

Officials claim that Baichuan-Omni-1.5 outperforms GPT-4omini in vision, speech, and multimodal streaming processing; in the field of multimodal medical applications, it has a more prominent leading advantage.

Please, Log in or Register to view URLs content!