Latent Space traces eight things AI stopped hiring humans for, ending with the physical world
An August 22 AINews essay argues the field repeatedly trades 10% accuracy for 100x cost and 10000x speed, and walks eight substitutions: reward signals (InstructGPT, Constitutional AI), training data (Phi, Apple's WRAP at ~3x pretraining efficiency, Nemotron-4 340B), teachers (Alpaca's $600 clone, DeepSeek-R1 distills), curriculum (Self-Instruct, STaR, Meta's Self-Rewarding LMs), researchers (Karpathy's autoresearch stacking 700 experiments into 20 kept improvements, cutting time-to-GPT-2 from 2.02 to 1.80 hours), environments, human subjects (Simile's twins at 85% of human self-reproduction accuracy), and physical experiments (CZ Biohub, in silico ~1000x cheaper than in vivo). The pattern it names is the useful part: each flip was blocked by a model-collapse objection until a verification mechanism made the synthetic version checkable.
Source
↳ Follow the thread