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

Two things caught my eye this week, and they’re more connected than they look.

First, DeepSeek quietly released a preview of V4 on Friday. It’s their new flagship model, and the headline feature is context length — it can handle much longer prompts than the previous generation, thanks to a redesigned architecture that chews through large text blocks more efficiently. That alone is useful for anyone working with long documents or multi-turn conversations.

But what actually matters is the chip play. V4 is DeepSeek’s first model optimized for Huawei’s Ascend chips. This is a direct test of how far China can get without Nvidia. The model is still open source, and its performance reportedly matches the closed-source leaders from Anthropic, OpenAI, and Google. If that holds up, it’s a real signal that the US export controls might not be as decisive as some hoped.

Caiwei Chen over at MIT Tech Review laid out three reasons this matters, and I agree with the gist: it’s a capability jump, it’s a geopolitical flex, and it’s a reminder that open-source models aren’t slowing down.

The second big thread this week is the resurgence of “world models.” This is one of those ideas that keeps bubbling up in AI research, but it’s now getting serious backing from heavy hitters like Stanford’s Fei-Fei Li and Yann LeCun. The argument is straightforward: large language models are great at manipulating symbols and text, but they’re terrible at understanding physics, spatial reasoning, or cause and effect in the real world. A world model attempts to learn a compressed representation of how the world works — not just language patterns — so an AI can actually plan, navigate, and manipulate objects.

This matters most for robotics. You can’t have a robot fold laundry or walk down a crowded street if it can’t predict what happens when it moves. LLMs alone won’t cut it. World models aim to give AI that missing intuition. Grace Huckins has a solid explainer on why this is suddenly at the forefront, and I’d recommend it if you’re curious about where the field is heading beyond chatbots.

Meanwhile, the geopolitical side of AI is getting messier by the day.

China blocked Meta’s $2 billion acquisition of AI startup Manus, citing national security. Beijing called the deal a “conspiratorial” attempt to hollow out its tech base. This is a big deal — it shows China is tightening control over AI firms trying to leave or get acquired by foreign entities. The decision escalates the US-China AI rivalry, and honestly, I don’t see winners there. Just two superpowers locking down their ecosystems.

On the other side of the Pacific, Google is throwing up to $40 billion at Anthropic, valuing the company at $350 billion. That’s an eye-watering number for a company that’s still burning cash on compute. The funding is meant to cover Anthropic’s growing infrastructure needs, but it also signals that the compute arms race between Anthropic and OpenAI is getting brutal. Both are fighting for access to the same limited GPU supply.

And in a move that feels straight out of a dystopian novel, President Trump just fired the entire National Science Board. The NSF has been a backbone of US technology research for decades. Firing the whole board raises serious concerns about political interference in science. The Verge and Nature both covered it, and the tone is grim.

Oh, and conspiracy theories about the Washington shooting are already spreading online. Over 300 posts flagged so far. Social media platforms are scrambling, but we all know how this goes. It’s exhausting.

If I had to pick one thing to watch this week, it’s the world model work. DeepSeek V4 is impressive, but it’s still an LLM at heart. The real leap will come when AI can understand the physical world as well as it understands text. That’s still years away, but the pieces are starting to line up.