DeepSeek V4, World Models, and a Week of AI Power Moves

Friday brought a surprise from DeepSeek: a preview of V4, their next flagship model. It’s been a long wait, and the specs don’t disappoint.

The headline feature is context length. V4 can handle significantly longer prompts than its predecessor, thanks to a more efficient architecture for processing large text blocks. That alone is useful for anyone working with lengthy documents or codebases.

But here’s where it gets interesting. DeepSeek claims V4 matches the performance of closed-source leaders from Anthropic, OpenAI, and Google—while staying open source. And this is their first model optimized for Huawei’s Ascend chips. That’s a big deal for China’s AI independence from Nvidia, and a real-world test of whether domestic hardware can compete.

Three things to watch:

• If V4 truly matches GPT-4 or Claude 3.5 on benchmarks, it validates the open-source approach at the frontier. That pressures the proprietary players on pricing.
• The Huawei optimization could accelerate China’s shift away from Nvidia. If performance holds up, expect more domestic models to follow.
• Longer context windows are becoming table stakes. DeepSeek’s efficiency gains may force others to rethink their architectures.

Meanwhile, the physical world is getting its own AI push. World models are having a moment, and for good reason.

Current AI systems excel at manipulating text and images, but folding laundry or navigating a busy street remains laughably hard. Researchers like Fei-Fei Li and Yann LeCun argue that LLMs hit a wall here because they lack a grounded understanding of physics and causality.

Enter world models. These systems aim to simulate how the world behaves—predicting outcomes of actions, reasoning about spatial relationships, and planning sequences. If done right, they could unlock robotics that actually works in messy, real-world environments.

It’s still early, but the idea is gaining traction. Expect to see more funding and papers in this direction. The gap between digital and physical AI is the next big frontier.

The week also delivered some dramatic moves in AI geopolitics and investment.

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. The decision escalates the US-China AI rivalry, but as MIT Technology Review notes, there will be no winners in this competition. China is also tightening controls on AI firms trying to leave the country.

On the investment side, Google is pouring up to $40 billion into Anthropic, valuing the company at $350 billion. The funding will support Anthropic’s growing computing needs, which is telling—both Anthropic and OpenAI are scrambling for compute capacity. That’s a supply constraint that will shape the industry for years.

In Washington, President Trump fired the entire National Science Board, raising fears of political interference in US science. The NSF has been a backbone for technology development, and this move doesn’t inspire confidence in long-term research stability.

And online, conspiracy theories about the Washington shooting are spreading rapidly. Over 300 posts in the first hour alone. The platforms are struggling to keep up.

It’s a lot to digest. DeepSeek V4 and world models represent genuine technical progress, while the political and financial moves show how high the stakes have become. The AI race isn’t just about benchmarks anymore—it’s about hardware independence, geopolitical leverage, and who gets to build the infrastructure for the next decade.