The world is being quietly rearranged by people who write very long documents.


The title they went with Humanline: Online Alignment as Perceptual Loss Noisy translates that to

Making AI training 6x faster by mimicking how humans actually perceive


Researchers found that the expensive way to train AI models (online, real-time feedback) works better than the cheap way (offline, batch data) because of how human brains perceive probability — not because of any fundamental requirement. They built cheaper offline training methods that deliberately encode these human perceptual quirks, matching the performance of expensive online methods while running six times faster.
If this holds up in practice, it collapses a false constraint: the entire AI training industry has assumed you need expensive real-time human feedback loops to get good results, but the bottleneck was never the feedback loop itself — it was not accounting for how humans actually perceive uncertainty.

If you insist
Read the original →