Welcome to Ten Blocks. I'm Jordan McGillis, economics editor of City Journal.

On January 20, 2025, the Chinese artificial intelligence firm, Deepseek, released its R1 model.

The model is competitive with top American models, but Deepseek has reportedly achieved this feat

at a tiny fraction of the cost that American firms have been pouring into their training.

The next day, President Donald Trump held a press conference at the White House

with the heads of Open AI, Oracle, and Japan SoftBank

to announce a $500 billion plan to build a system of AI data centers in America called Stargate. To discuss the latest

in AI geopolitics, I've invited Judd Rosenblatt on today's show. Judd is the founder and CEO of

A.E. Studio and a leading advocate for aggressive, thoughtful, American AI development.

Judd, thanks for coming on.

Thanks for having me, Jordan.

First question for you.

What does Deep Seeks release tell us about the state of the AI arms race?

Well, it's impressive that they were able to make so much algorithmic improvements with limited compute.

And one thing that's fairly interesting about the deep seek work is that it strongly reinforces this idea of a negative alignment tax,

where improving alignment techniques, investing in trying to make AI more likely to be more capable by virtue of its

alignment, not only mitigates risks, but also enhances capabilities because it uses reinforcement

learning to induce chain of thought reasoning. And that winds up increasing. So basically,

optimizes for transparent reasoning and structures and also

increases model performance in all these different complex tasks, math ones, especially. And so,

so basically, instead of just using reinforcement learning for preference alignment, which is what

Open AI's RLHF does for politeness and stuff, it uses these reward signals in reinforcement learning

...