DeepSeek just proved Lina Khan right | Hacker News Reader

DeepSeek just proved Lina Khan right

dropsitenews.com

8 pointskelthuzad7 comments

The US has forced China to build a Better Mousetrap! And do so with GPUs of lesser processing power than the west has access to! The west's policy to limit GPU technology access to China has necessitated that innovation that is utilizing lower power GPU hardware than was assumed was needed! And so what a mother of invention that's become to the tune of some trillion+ loss in market value among the US AI Monopolies! The US has become so good at stifling out fair competition and the very driver of that innovation that creates the real wealth!

So why did the Biden administration block China's access to advanced AI chips? Why did Biden not reverse Trump's Huawei ban?

DeepSeek did not prove Lina Khan right. These things are honestly not even that related.

I think you're missing the core argument here. Khan's warning wasn't about foreign competition - it was about how monopolistic behavior leads to technological stagnation. The Deepseek situation demonstrates this perfectly. Let's look at the actual numbers:

- DeepSeek built their model with 2,048 H800 GPUs and $5.5M in computational costs

- Meta targeting 600,000 H100 GPUs

- OpenAI Used ~25,000 A100 GPUs for GPT-4

- Altman claims to need $7T for his "dream AI"

This massive efficiency gap suggests something deeper than just trade policy at work. American tech giants have become accustomed to solving problems by throwing massive resources at them rather than through efficient engineering. The chip restrictions and Huawei ban are separate policy decisions based on national security concerns. But they don't address what Khan warned about - namely that allowing monopolistic practices would make these companies "fat and happy inside their castles," focusing on "performing alchemy with spreadsheets" rather than actual innovation.

DeepSeek's achievement is particularly interesting because they succeeded despite having access to less advanced chips (H800s instead of H100s/A100s). They had to optimize for efficiency, which ironically led to better results.

This isn't about whether export controls are justified - it's about how domestic monopolistic practices made our tech companies vulnerable to more innovative competitors, exactly as Khan predicted.

Deepseek trained their models on GPT4o and Claude 3.5 and Llamma outputs. In other words, DeepSeek models needed billions in training cost done for them before they could train their models for $5.5m in compute cost.

The core of Khan's argument still stands, U.S. companies have been claiming they need astronomical resources ($7T in Altman's case) and massive GPU clusters (Meta targeting 600,000 H100s) while a smaller competitor demonstrated that better efficiency is possible, even if they built on existing work.

The article's argument isn't really about absolute costs, it's about how monopolistic practices led to inefficient resource utilization and reduced innovation pressure. Whether DeepSeek used existing model outputs or not, they still managed to create a competitive model with significantly fewer resources than what U.S. companies claim they need.

This is exactly the kind of technical discussion we should be having instead of dismissing DeepSeek's achievement outright or accepting the narrative that AI development requires unfathomable resources.

You're trying to tie two unrelated things together.

We know competition is good for consumers. But overregulation is also bad for consumers.

Khan wasn't just advocating for generic regulation, she specifically warned that allowing unchecked consolidation would make the companies themselves weaker and less innovative.

She argued that "maintaining open, fair, and competitive markets, especially at technological inflection points, is a key way to ensure America benefits from the innovation these tools may catalyze." It's mentioned that U.S. tech companies "would buy the company, kill it, and absorb some of its staff" instead of competing through innovation. That's why it isn't about overregulation, it's about the lack of enforcement of existing antitrust laws during "the Bush and Obama eras [which] saw unbridled growth and mergers."

The DeepSeek situation is a symptom of this problem. When companies could simply acquire or crush competitors rather than out-innovate them, they "got fat and happy inside their castles" and "pivoted from technological innovation to performing alchemy with spreadsheets."

So this isn't just a binary choice between competition and regulation. The article's argument is that failing to enforce basic antitrust laws actually hurt competition and innovation. Whether you agree with that interpretation or not, it's more nuanced than just "competition good, regulation bad."

You are the one trying to separate these issues despite the article making a clear case for their connection.

Reply on news.ycombinator.com