DeepSeek V4, World Models, and the Week AI Got Real

Friday brought us DeepSeek V4, the long-awaited new flagship from the Chinese AI lab. The preview isn’t a full release, but it’s already making noise for three reasons.

First, it handles much longer prompts than V3. That’s thanks to a new architecture that processes large chunks of text more efficiently. If you’ve ever hit a context window wall while trying to feed a model a whole codebase or a book, this is a real improvement.

Second, the performance matches leading closed-source models from Anthropic, OpenAI, and Google. That’s not a shock—DeepSeek has been competitive for a while—but it reinforces that open-source AI isn’t a charity project anymore. It’s a genuine threat to the big labs.

Third, and most interesting: V4 is DeepSeek’s first model optimized for Huawei’s Ascend chips. This is a direct test of China’s dependence on Nvidia. If Ascend can deliver competitive training and inference, it’s a big deal for Beijing’s semiconductor self-sufficiency goals. I’m skeptical about performance parity, but the fact that they tried says a lot.

Separately, the world model conversation is heating up again. The pitch is simple: LLMs are great at text, but they fail at physical tasks like folding laundry or driving a car. World models—systems that learn the physics and dynamics of the real world—are supposed to fix that.

Fei-Fei Li and Yann LeCun are the loudest proponents. They argue that without a world model, AI will always be stuck in the digital realm. I’ve seen this argument before, and it’s compelling in theory. But building a world model that actually works in robotics has been a graveyard of PhD theses. The hardware is still the bottleneck. Still, the attention is shifting, and that matters.

Elsewhere this week, the geopolitical AI chess game got a lot more interesting.

China blocked Meta’s $2 billion acquisition of AI startup Manus. Regulators cited national security, and Beijing called the deal a “conspiratorial” attempt to hollow out its tech base. This isn’t just about one acquisition; it’s a signal that China is tightening control over AI firms that try to leave or sell to foreign buyers. No winners here.

Google is investing up to $40 billion in Anthropic, valuing the firm at $350 billion. That’s a lot of cash, even by AI standards. The money will go toward computing needs, which tells you how desperate the compute arms race is. Both Anthropic and OpenAI are fighting for capacity, and Google just wrote a check that says “we’re all in.”

President Trump fired the entire National Science Board. The NSF has been a quiet but crucial engine for US technology development. This move heightens fears about political interference in science. It’s not a good look.

Conspiracy theories about a Washington shooting are spreading online. Over 300 accounts were suspended, but the damage is done. The platforms still can’t handle real-time disinformation.

It’s a week where the lines between corporate strategy, national security, and research are blurring faster than anyone expected. DeepSeek V4 is a reminder that open-source AI isn’t slowing down. World models remind us that we’re still bad at physical tasks. And the deals and firings remind us that AI is now a geopolitical weapon, not just a tech product.