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

Latent Space · rss · 2026-08-22

Argues that the AI pipeline is flipping from human-made to model-made across all components. It traces 8 stages of this "synthetic" evolution:

Reward Signal (2022): Reward models and RLAIF replace human feedback.
Training Data (2023): Phi and Nemotron-4 prove synthetic data outperforms raw data.
The Teacher (2023): Distillation from larger models (e.g., DeepSeek-R1) becomes standard.
The Curriculum (2024): Models generate their own tasks and improve via Self-Rewarding LMs.
The Researcher (2026): Karpathy's autoresearch and AlphaEvolve automate the discovery loop.
The Environment (2026): End-to-end synthetic environments and verifiers (GLM-5.3) solve RL scaling.
The Human Subject (2025): Simile uses digital twins to replace focus groups and A/B tests.
The Physical World (In Progress): AI for Science compresses the physical world, but real experiments remain essential.

More from AGI Musings