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

Friday was quiet enough that Latent Space used it to zoom out, and the pattern they land on is blunt: since 2022, another piece of the machine-intelligence pipeline has kept flipping from human-made to model-made. Not all at once. Not cleanly. But each time, there’s a patient zero — a paper, a product, a release — where the synthetic version stops being a stunt and starts carrying real weight.

The cleanest way to think about it is the one-line summary they give the whole thing: increasingly ambitious human simulation, just 10% worse, but 100x cheaper and 10,000x faster. That shorthand runs through the whole stack. The reward signal went synthetic first, with InstructGPT’s preference model, then Constitutional AI, then LLM-as-judge evals. The models moved from being judged by people to judging each other.

Then came the data. Microsoft’s Phi series argued that textbook-style synthetic data could punch far above its size, and Apple’s WRAP pushed the idea wider by rephrasing the web with an LLM. NVIDIA’s Nemotron-4 340B made synthetic data generation part of the pitch, and by 2025 reasoning traces had become ordinary pretraining fuel. The same move kept repeating in the teacher role too: Alpaca, Vicuna, Orca, and later distilled model families made “the teacher is a model” feel normal instead of novel.

By 2024, the loop was closing. Models were no longer just consuming synthetic inputs; they were deciding what should be learned next. Self-Instruct and STaR laid the groundwork, then Meta’s Self-Rewarding Language Models and SPIN showed that a model could generate tasks, judge outputs, and push past the edge of its human preference data. Curriculum design, once one of the most human parts of ML, started looking automated.

The frontier has now moved into research and environments. DeepMind’s AlphaEvolve found new algorithms, Sakana’s AI Scientist reached the paper-writing stage, and Andrej Karpathy’s autoresearch loop ran 700 experiments to keep 20 improvements, cutting time-to-GPT-2 from 2.02 to 1.80 hours. On the environment side, Z.ai’s GLM-5.3 stack synthesizes task worlds, judges, and verifiers end to end. Simile takes the same logic to the human layer, using interviews, transaction data, and registered RCTs to build digital twins that can stand in for real subjects. The message is pretty clear: the models are no longer just copying humans. They’re replacing the whole feedback loop around them.

And the article’s real point is that this only works when verification gets better. Generation can be messy. The synthetic step still wins when there’s a way to check it: preference agreement, unit tests, proof checkers, oracle and no-op checks, registered trials, wet-lab feedback. That’s why the remaining frontier is the physical world. It’s the slowest, most expensive place to verify anything, which is exactly why it hasn’t been fully simulated yet.