DeepSeek V4 and the World Model Race: What Actually Matters

Friday brought us DeepSeek V4, the long-awaited follow-up from the Chinese AI lab that’s been making waves. The preview release isn’t just another model update—it’s a signal.

Three things stand out. First, the context window. V4 can handle much longer prompts than its predecessor, thanks to a redesigned architecture that processes large text blocks more efficiently. That’s genuinely useful for anyone working with long documents or complex codebases.

Second, performance. DeepSeek claims V4 matches leading closed-source models from Anthropic, OpenAI, and Google. I’m skeptical until I run my own benchmarks, but the open-source community will verify this quickly. If true, it’s another blow to the argument that proprietary models are inherently superior.

Third—and this is the big one—V4 is the first DeepSeek model optimized for Huawei’s Ascend chips. This is a direct test of China’s ability to reduce dependence on Nvidia. If Ascend can deliver competitive performance, the geopolitical implications are huge. If not, well, we’ll know soon enough.

The World Model Hype Train

Meanwhile, there’s a growing buzz around “world models”—AI systems that understand physical reality, not just text. The argument is straightforward: LLMs are great at composing novels and writing code, but they can’t fold laundry or navigate a busy street. To bridge that gap, you need something fundamentally different.

Stanford’s Fei-Fei Li and AMI Labs’ Yann LeCun are leading the charge. They argue that world models can overcome the well-known limitations of LLMs—lack of common sense, inability to reason about physics, zero understanding of cause and effect. I’ve seen this pitch before, and it’s compelling in theory. But building a model that actually grasps the physical world is orders of magnitude harder than scaling up next-token prediction.

World models made MIT Technology Review’s list of “10 Things That Matter in AI Right Now,” which tracks trends worth paying attention to. I’d put it lower on the list than, say, the open-source vs. closed-source battle or the compute crunch, but it’s not wrong to include it.

The Real News in AI This Week

Let’s cut through the noise. 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 escalates the US-China AI rivalry, but honestly, there are no winners in this competition—just more fragmentation and duplication of effort.

Google is investing up to $40 billion in Anthropic, valuing the firm at $350 billion. That’s a lot of zeroes for a company that hasn’t turned a profit. The funding will support Anthropic’s growing computing needs, which is the real bottleneck in this arms race. Both Anthropic and OpenAI are fighting for compute capacity, and Google has the deepest pockets.

President Trump just fired the entire National Science Board. The NSF has been a cornerstone of US technology development for decades. This move heightens fears over political interference in science. I don’t care which party does it—politicizing research funding is a terrible idea.

And of course, conspiracy theories about the Washington shooting are proliferating online. Over 300 accounts pushed coordinated narratives within hours. Social media platforms are still playing whack-a-mole with disinformation, and they’re losing.

My Take

DeepSeek V4 is a genuine achievement, but the real story is the chip optimization. If Huawei’s Ascend can deliver competitive performance, it reshapes the hardware landscape. If not, China’s AI ambitions face a hard ceiling.

As for world models, I’m watching but not holding my breath. We’ve seen this movie before—every few years, someone declares that the next paradigm shift is here. Sometimes it is. More often, it’s a research paper that never makes it to production. The physical world is messy, expensive, and slow to simulate. I’ll believe in the robot butler when it actually folds my laundry without ripping my shirts.