LLMS KNOW MORE THAN WHAT THEY SAY

Summary from the TLDR Web Dev newsletter

An approach called Latent Space Readout (LSR) can improve the accuracy of AI application evals. LSR uses interpretability research by projecting evaluation model activations onto meaningful linear directions, which makes it more sample efficient than fine-tuning. This means that much less human feedback is required for these evals to be more accurate.