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

DeepSeek finally released a preview of V4 on Friday, and it’s not just another incremental update. The Chinese outfit’s new flagship can handle significantly longer prompts thanks to a redesigned architecture that chews through large text blocks more efficiently. That alone would be notable, but the real story is elsewhere.

V4 is DeepSeek’s first model optimized for Huawei’s Ascend chips. This matters because it’s a direct test of how far China can wean itself off Nvidia hardware. The model is still open source, and its performance reportedly matches Anthropic’s Claude, OpenAI’s GPT-4o, and Google’s Gemini. If Huawei’s chips can carry that weight, the geopolitical implications for the AI supply chain are huge. Caiwei Chen at MIT Technology Review laid out three reasons this matters, and I’d add a fourth: it shows open-source models can keep pace with closed-source giants without relying on the same silicon.

Meanwhile, the world model conversation is heating up again. Stanford’s Fei-Fei Li and AMI Labs’ Yann LeCun are pushing the idea that we need systems that understand physics and causality, not just text patterns. LLMs are great at writing novels or generating code, but they fall apart when asked to fold laundry or navigate a crowded street. World models aim to bridge that gap by giving AI a grounded understanding of how the physical world works. Grace Huckins covered this for MIT Tech Review, and it’s refreshing to see serious researchers calling out the limitations of pure language models. The robotics industry has been waiting for this shift for years.

Now for the messy stuff. 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. That’s strong language, and it signals that China is tightening control over AI firms trying to leave the country. The decision escalates the US-China AI rivalry, but as MIT Tech Review noted, there will be no winners in this competition. Both sides lose when talent and capital can’t flow freely.

Google, meanwhile, is investing up to $40 billion in Anthropic, valuing the company at $350 billion. That’s a staggering sum, and it reflects the insane compute demands these frontier models require. Anthropic and OpenAI are effectively fighting for the same pool of GPU capacity, and Google is betting big that Anthropic’s safety-first approach will pay off. Whether that bet works depends on how quickly Anthropic can scale without losing its edge.

President Trump just fired the entire National Science Board. The NSF has been a quiet powerhouse in developing foundational technology, and this move heightens fears about political interference in US science. Nature called it alarming, and I agree. Removing an entire oversight board mid-stride is a recipe for chaos, especially when the US is trying to maintain its lead in AI research.

And of course, conspiracy theories about the Washington shooting are proliferating online. Over 300,000 posts in the first 24 hours, according to some trackers. The usual pattern: bad actors exploit confusion before facts settle. Platforms are scrambling to label misinformation, but by the time they act, the narratives have already hardened.

This week felt like a stress test for the AI ecosystem. DeepSeek is proving open source can compete, world models are challenging the supremacy of LLMs, and geopolitics is reshaping who gets to build what. The next few months will tell us whether this is a turning point or just another cycle of hype.