[AINews] 10% worse, 100x cheaper, 10000x faster: Why Simulation is taking over

This Latent Space editorial argues that generated or simulated material becomes trustworthy when something can check it. Its examples pair synthetic training text with filtering, training that rewards verifiable answers with unit tests and proof checkers, and simulated people with registered randomized trials of real participants.

The editorial concludes that synthetic work advances when verification does, and that physical experiment is where checking is slowest and most expensive, so that part of reality cannot be fully synthesized. The companion interview on simulated people covers the simulation example in more depth. The headline's 10% worse, 100x cheaper and 10,000x faster is the author's heuristic summary, not a measurement from one comparison.