Just been reading through Meta's latest moves and honestly, the pattern is getting hard to ignore. December acquisition of Manus, then Moltbook just last week. Two billion here, undisclosed amount there. But what's actually interesting isn't what Meta bought—it's what they didn't.

Let me back up. Earlier this year Zuckerberg went on this unprecedented recruitment spree, supposedly offering signing bonuses up to $100 million, meeting people at his places in Lake Tahoe and Palo Alto. The targets? Perplexity AI, Runway, Safe Superintelligence, and Thinking Machines Lab. All four said no. That rejection list tells you more about Meta's problem than any successful deal ever could.

Consider the contrast with what OpenAI did. When Peter Steinberger's OpenClaw framework hit GitHub—this was the underlying tech for Moltbook—it exploded. 200k stars in weeks, 2 million weekly visits. OpenAI looked at this and thought: we need this person. They brought him in, made him head of personal Agents, kept OpenClaw open source as his only condition. They got the builder.

Meanwhile Meta went after the data labeling infrastructure. Alexandr Wang from Scale AI, $14.3 billion deal. Wang's company has never trained a complete large model—they organize human labelers to categorize training data. It's important work, sure, but it's not research. It's not building. And when Yann LeCun, the Turing Award winner who built Meta's credibility in AI research, refused to report to Wang, he left. Eleven of the original fourteen Llama researchers have already departed.

Here's what really gets me: Meta's flagship model Llama 4 Behemoth finished training but internal evals didn't hit the mark. Release shelved. A company spending over $100 billion annually can't ship its main product. So what's the response? Buy Manus, buy Moltbook. Acquire the narrative instead of building it.

The Moltbook story is instructive. Matt Schlicht and Ben Parr are smart people—good storytellers, solid connectors in the Agent ecosystem. But they're not Ilya Sutskever leaving OpenAI to start Safe Superintelligence because he had conviction about AI safety. They're not Steinberger building something that resonates with 200k developers in weeks. They're connectors, not builders. And Meta paid for connectors.

Back in 2012, when Zuckerberg offered $1 billion for Instagram with thirteen employees, that made sense. Instagram had already proven mobile photo sharing was irreversible. WhatsApp showed instant messaging could replace SMS globally. Facebook had the distribution to amplify what was already working. The founders sold because leverage was scarce and distribution was everything.

Now? Capital isn't scarce. Distribution isn't scarce. What's scarce is conviction and independence. Sutskever didn't need Meta's users to build AI safety research. Perplexity didn't need Meta's infrastructure to compete in search. Steinberger didn't need Meta's resources to create OpenClaw. These people had their own narrative.

Meta AI hit 1 billion monthly active users in early 2025. Sounds big until you realize no one changed their behavior because of it. ChatGPT rewired search habits in two months. Gemini got embedded into Android. Claude became the enterprise standard. Meta AI is a feature in Instagram and WhatsApp that people use occasionally but don't think about.

Even Manus is revealing. The Agent capabilities? Powered by Anthropic's Claude. Meta spent billions acquiring a working AI Agent, but the actual intelligence comes from a competitor. That's not a research win. That's a dependency.

The brutal part is what Zuckerberg's choices say about where this goes. When your most qualified internal voice—a Turing Award winner—disagrees with your direction and you choose the 28-year-old data labeling founder instead, you're signaling something. You're saying direction matters more than dissent. You're saying we're committed to this path regardless.

But the asset pool willing to join that path has gotten noticeably smaller. In 2012, the puzzle had a rational answer. Today it doesn't. And buying Moltbook doesn't solve that.