DeepSeek V4 Drops, World Models Heat Up, and China Blocks Meta’s Manus Deal

DeepSeek V4: Not Just Another Model Release

Look, I’ve been around long enough to get cynical about model announcements. But DeepSeek’s V4 preview, dropped last Friday, actually has some teeth. The headline numbers are decent — it matches Anthropic, OpenAI, and Google on performance while staying open source. That alone is a flex. But what caught my eye is the architecture change: V4 handles much longer prompts efficiently. That’s not just a spec bump; it means you can throw entire documents or long conversation histories at it without the usual context window headaches.

More interestingly, this is DeepSeek’s first model optimized for Huawei’s Ascend chips. That’s a direct test of China’s ability to wean itself off Nvidia. If V4 runs well on domestic hardware, it signals a real shift in the supply chain dynamics. Whether it actually delivers in practice remains to be seen — I’ve seen too many “optimized” models that choke on real workloads. But the intent is clear.

Caiwei Chen at MIT Tech Review nailed it: this could shake up the AI landscape, especially if the open-source community embraces it. I’m watching to see if the community finds V4 more hackable than its predecessors.

Why World Models Are Suddenly Everywhere

Meanwhile, the physical world is still stubbornly analog. We can generate novels and code, but ask an AI to fold laundry or navigate a sidewalk and it falls apart. That’s the gap world models are supposed to bridge. Fei-Fei Li and Yann LeCun are pushing hard on this, arguing that LLMs hit a wall without a grounded understanding of physics and causality.

Grace Huckins’ piece on this is worth a read. She points out that world models aren’t just another academic curiosity — they’re the key to making robotics actually useful outside labs. I’ve seen demos of robots that can pick up objects they’ve never seen before, and it’s impressive, but we’re still years from a robot that can fold a fitted sheet. The bet is that a world model gives the system common sense about how objects behave, which LLMs lack.

MIT Tech Review included world models in their “10 Things That Matter in AI Right Now” list, and I think that’s fair. It’s one of those areas where hype and genuine progress are actually aligned for once.

The Must-Reads: China Blocks Meta, Google Bets Big on Anthropic

The week’s news was a whirlwind. China blocked Meta’s $2 billion acquisition of Manus, an AI startup, citing national security. That’s a big deal — it’s the kind of move that escalates the US-China tech rivalry into outright asset blocking. Beijing called it a “conspiratorial” attempt to hollow out its tech base. I’ve seen this play out before in semiconductors, and it rarely ends well for anyone. The MIT Tech Review piece on this is spot-on: there are no winners in this competition.

In other news, Google is pumping up to $40 billion into Anthropic, valuing it at $350 billion. That’s a lot of zeroes. The funding is supposedly for compute capacity, which tells you how desperate the compute arms race has become. Anthropic and OpenAI are basically in a bidding war for GPU clusters. It’s getting ridiculous.

And President Trump just fired the entire National Science Board. The NSF has been a backbone of US science funding for decades. This move is raising alarms about political interference, and rightly so. The Verge and Nature both covered it well.

Oh, and conspiracy theories about the Washington shooting are already spreading like wildfire. That’s the internet we live in.

Quick Hits

• Google’s $40B Anthropic bet: The deal values the firm at $350B, with funds earmarked for compute. TechCrunch has details.
• China tightens grip on AI firms trying to leave: TechCrunch reports on the crackdown.
• No winners in US-China AI rivalry: MIT Tech Review’s take is sobering.

I’ll be keeping an eye on DeepSeek V4’s real-world performance and whether world models actually make it out of labs. The next few months should be telling.