We cracked @sama's death star vaguepost. It's been one week since the GPT-5 launch, and we think we finally understand Sam’s death star vaguepost. In the lead-up to the GPT-4 launch, there was a data visualization that showed the model’s parameter count compared to GPT-3’s that went viral. The implication: a bigger pre-train meant a much better model. That turned out to be true for the GPT-3 to GPT-4 leap. But GPT-4.5 taught us that pre-training scale isn’t enough — you also need a mixture of techniques, experts, researchers, etc. The death star post represents an even bigger circle from the GPT-3 -> GPT-4 meme, but OpenAI blew up that metaphorical big circle with a model that isn’t just bigger but better engineered — “the router is the release,” as SemiAnalysis put it. Some expected GPT-5 to be superintelligence and were disappointed by what felt incremental. But there are plenty of reasons to be optimistic here. GPT-5 has clearly been RL’ed in a way that reduces hallucinations, probably by training on unknowable / impossible research tasks and including examples where experts modeled saying, “A good answer to this question just isn’t out there.” That’s a huge upgrade. We’ve also talked on the show about how GPT-5’s router sets OpenAI up to really dominate in agentic commerce / a monetizable free tier — a bull case for superintelligence. The Masa card can really only be played once, and the singularity won’t be subsidized by venture capitalists. It’ll be funded by free cash flow. The trillion dollar cluster will be built, and it will be paid for by a highly profitable tech company that can directly underwrite the investment with an aligned business model. The singularity is still thousands of days away, but generally on track.