What is neuralese and why is it bad?
Linch explains what would be lost if AI models stopped reasoning in words. Today a model's chain of thought is natural language that humans and weaker models can read, which makes monitoring it "one of the few techniques we can somewhat reliably use to track model intent"; neuralese would replace it with "a long stream of numbers". The hook is The Information's report, from leaks, that OpenAI's upcoming model Astra may be a hybrid with a "neuralese recurrent loop", which the post calls "an obvious intermediate step on the path to full neuralese", asking journalists to dig and anyone working at OpenAI or another company building neuralese to champion internal reforms or quit. As the post notes, no current publicly available model is known to use neuralese.
Was this useful?