AINews Maps Eight Stages of Humans Being Replaced in the ML Pipeline, and Says Verification, Not Generation, Gated Every One

Latent Space (AINews)·high signalXBlueskyLinkedInCopy link

The August 22 issue lays out a progression from reward signals (InstructGPT, 2022) through training data (Phi), teachers (Alpaca's $600 fine-tunes), curricula, researchers (Karpathy's autoresearch stacking 700 overnight experiments to cut GPT-2 training from 2.02 to 1.80 hours), environments (Z.ai), human subjects (Simile), and now cells. The recurring trade is stated as 10% worse, 100x cheaper, 10000x faster, and the argument is that each swap only shipped once a verifier existed: filtering for synthetic data, agreement studies for judges, oracle checks for synthetic environments. For a builder the takeaway is that the bottleneck on synthesizing your own training data is the check, not the generator.