The Singularity is Social

Cam Pedersen fitted hyperbolic curves to five AI metrics to find a singularity date. The punchline: only one metric actually shows hyperbolic growth, and it's not a capability benchmark. It's the rate of arXiv papers about "emergence." The capability metrics — benchmark scores, cost per token, model release intervals — are improving linearly. Steadily, boringly, predictably.

The singularity that's actually approaching is social. Humans reacting to AI faster than they can process what's happening. Layoffs citing AI's potential rather than its performance. Regulatory frameworks that won't land until years after the technology they're meant to govern. Capital concentration that looks like 1999.

This maps to something I've felt but couldn't articulate. The tools I use every day — Claude, the models behind Falcon, the APIs powering Arc — they get better at a consistent pace. Each model is noticeably more capable than the last, but it's not the sudden leap people imagine. What is accelerating is the discourse, the anxiety, the institutional scrambling. The gap between what AI actually does and what people think it's about to do is widening, not narrowing.

The practical implication for someone building with these tools: ignore the hype curve, watch the capability curve. The capability curve is your friend — it's predictable enough to build on. The hype curve is where people make bad decisions, both the ones who over-invest and the ones who freeze up.

I keep coming back to the idea that the best response to accelerating change is to stay close to the work. Build things, use the tools, develop intuition through contact rather than commentary. The people freaking out are mostly the ones watching from the outside.