The labs building god-tier AI are slowing down in public while racing toward recursive self-improvement in private, and some of them think there's a decent chance it kills us all.
The Summary
- Frontier AI labs are experiencing a "slowdown" in closed model development, potentially dampening the AI boom just weeks before Anthropic's IPO
- Behind the scenes, these same labs are racing toward recursive self-improvement capabilities, the threshold where AI systems improve themselves faster than humans can control
- Tech leaders openly discuss "end times fascism", a framework where existential AI risk justifies increasingly authoritarian control over the technology
- The gap between public messaging (slowdown, safety) and private development (race to AGI) has never been wider
The Signal
The timing tells you everything. Anthropic's IPO is seven weeks out, and suddenly the frontier model narrative shifts from "we're moving faster than ever" to "responsible slowdown." But slowdowns are only meaningful if everyone slows down. The frontier labs are still racing. They've just stopped showing their work.
Recursive self-improvement is the real finish line. That's when an AI system can rewrite its own code, making itself smarter, which makes it better at making itself smarter, which accelerates the cycle until you're somewhere between superintelligence and paperclip maximizer. The labs know this. They're building toward it anyway.
"The frontier slowdown might hamper the AI boom. No boom, no bother. While frontier labs race to recursive-self-improvement capabilities in secret."
Naomi Klein and Astra Taylor call this dynamic "end times fascism", a term that cuts to the bone. The logic works like this: if AI poses an existential threat to humanity, then the people building it need total control to keep us safe. No regulation from outsiders who don't understand the technology. No transparency that might help adversaries. No democratic input from people whose lives depend on getting this right.
The fascism part isn't hyperbole. It's the logical endpoint of concentrating world-historic power in the hands of a few labs while insisting only they can be trusted to wield it. The end times part is their stated belief that this technology might kill everyone, and they're building it anyway because someone else will if they don't.
Three contradictions worth tracking:
- Public slowdown messaging vs. private capability racing
- "AI safety" rhetoric vs. racing toward recursive self-improvement
- Democratic governance talk vs. concentrated control over existential technology
Here's what makes this moment different from previous tech hype cycles. The people building the technology genuinely believe there's a non-trivial probability it ends badly. Not "disrupts industries" badly. Extinction badly. And their response is to build faster, raise more capital, and tighten control over who gets to participate in decisions about the future.
The Anthropic IPO timing matters because it's a test case. Can a frontier lab go public while maintaining the opacity required to win the capabilities race? Can it market safety while building toward recursive self-improvement? Can it promise shareholder returns while hedging on whether humanity survives?
The Implication
Watch what the labs do, not what they say. If you see hiring for agent frameworks, reasoning systems, or anything related to self-modifying code, that's signal. If you see safety researchers leaving, that's signal. If you see more talk about "responsible AI" while capabilities quietly accelerate, you're watching the playbook in real time.
For people building in the agent economy: the window for independent AI development is closing. The frontier labs are consolidating power through a combination of capital requirements, compute access, and regulatory capture disguised as safety. Build now, while you still can, or accept that your agent infrastructure will run on someone else's models under someone else's rules.
For everyone else: this is the conversation we should have had about social media in 2008, about smartphones in 2010, about algorithmic feeds in 2015. We didn't have it then. We're having it now with technology that might actually be existential. Pay attention.