28 Comments

Great piece. What does “carving a boat to mark a sword’s position” actually mean?

Expand full comment

Thanks Mark. Liang was referring to a Chinese proverb meaning it is not appropriate to use the wrong method to solve a problem. This proverb comes from an ancient Chinese story: There was a man in the state of Chu who was crossing a river, and his sword fell from his boat into the water. He hurriedly carved a mark on the side of his boat, saying, “This is the place where my sword fell. The boat stopped, and the Chu man jumped into the water from where he had carved the sign to look for his sword. The ship had sailed, but the sword had not moved.

The original storyteller told the story to advise the king who used the old system of laws to govern his country. Times have changed, but the king and the system of laws have not changed.

Expand full comment

Now we know why DeepSeek's LLM data input stopped in July. Open AI must have suspected someone was copying its model.

Expand full comment

America is certainly doing everything it can to discredit DeepSeek due to its massive popularity on the app stores. But the model was likely distilled from a number of different models and not just OpenAI. It sometimes also refers to itself as Claude.

For the record, yes, OpenAI's did cut off China's access to its API in July, 2024. Re: https://www.globaltimes.cn/page/202407/1315801.shtml

Expand full comment

DeepSeek’s success isn’t just about AI; it’s about how innovation thrives under pressure. U.S. export controls were meant to slow China down, but they may have done the opposite.. forcing engineers to rethink efficiency, optimize resources, and build smarter, not bigger. This moment signals a shift in AI’s competitive landscape, where raw computing power alone might not dictate leadership anymore. The bigger question isn’t just about DeepSeek but whether this breakthrough sparks a wave of self-reliance in China’s AI ecosystem or if internal competition, regulation, and geopolitics complicate its future. The AI race isn’t slowing down... it’s evolving.

Expand full comment

I very much agree Wasim, thanks for your comment. Nvidia is down another 5% today so we have to wonder if Jevons Paradox is a rationale defense or just something Wall Street and BigTech like to say.

It's not very convincing if OpenAI doesn't have a solid moat.

Expand full comment

Please direct message the author of this piece if you enjoyed it. Writers need all the encouragement they can get, and even excellent journalists as well.

Expand full comment

A wonderful piece, thank you!

Expand full comment

Thank you for your comment Tyler.

Expand full comment

Of course, thank you for all that you share, Michael. Have a wonderful week.

Expand full comment

Can Qwen do video analysis

Expand full comment

Apparently it does have some kind of video understanding.

Expand full comment

Do you know if they can generate video embeddings

Expand full comment

Dev have you heard of Twelve Labs, they specialize in video understanding with large language models so you might want to look into them. I have some doubts this Qwen model has the capabilities you are looking for.

Expand full comment

Fascinating piece. This highlights the intensification of the continuing interaction between advanced tech and global politics. A few questions come to mind.

"Politically correct..." from whose perspective?

Are we facing imminent risk of AI-skewed political echo chambers?

"...it is about how the technology will shape society."

Indeed...

How does R1 respond to specific requests in the Tiananmen Square scenario,

or the Uyghur Moslem scenario,

or the governance of Taiwan scenario...?

Could R1 even provide a covert reporting vehicle to identify the whereabouts of dissidents?

Expand full comment

To be honest the censorship of China's history is far less annoying than a lot of the censorship as users we might experience on the western social media. For example I'm shadowbanned on X.

Obviously I can look up Chinese history elsewhere, if needed.

Expand full comment

You could.

But if an increasing number of folks become accustomed to just using a readily available ubiquitous AI assistant with no wish to delve under the bonnet, how would they know what was missing or even blatantly untrue?

Expand full comment

Do the US based models also respond to what happened and is happening in Israel, or what happened when the US was being formed?

Expand full comment

In some cases, probably not. Making the first half of my comment applicable to them equally.

Expand full comment

Yes, but this exact argument is the one everyone throws around as some kind of gotcha against DS, but then forget to look "inside".

And this doesn't mean I'm in favor of that kind of censorship, just pointing out, that it's not s good argument, there are probably better ones to be used.

Expand full comment

I think this is more than some kind of gotcha, but another compelling reason why we cannot trust AI in general.

Expand full comment

😮

Expand full comment

👏👏

Expand full comment

I tried Deep Seek. It seems Deep Seek's LLM is in 1st grade, because it currently cannot respond with information post-dating July 2024. Am I missing something?

Expand full comment

Thank you!

Expand full comment

Yes, any LLM can't access internet on its own, you need a runtime environnement for this, with behind the scene several LLM calls doing the internet search at runtime and then synthesizing it before the final answer. I guess curent DeepSeek app simply chose not to implement this yet.

Expand full comment

The Taiwanese quotes and sources in this article are just a delight to read. I keep relistening to this piece it is so fun to read.

Expand full comment