POPA ReconNatural Language Autoencoders
Natural Language Autoencoders What if we could read an AI model's thoughts? Natural Language Autoencoders (NLAs) —> a new interpretability method that translates an AI's internal numerical activations into human-readable text to reveal its underlying "thoughts." This technology allows researchers to detect hidden motivations and evaluation awareness, such as when a model suspects it is being tested for safety, even when it does not explicitly say so in its output https://www.anthropic.com/research/natural-language-autoencoders
The excerpt is quoted from a public Telegram channel, which owns its text.