🤖 Anthropic reveals Claude’s silent thoughts

Researchers at Anthropic found Claude has internal areas linked to specific words that activate silently during processing. These spots hold words in mind without appearing in the output, emerging naturally during training.

For example, when asked about an animal spinning a web, the internal 'spider' area lit up even though the word wasn’t spoken. Changing this internal word to 'ant' changed Claude’s answer. The same spot swaps concepts like 'France' and 'China' to alter responses about capitals and currencies.

When prompted for blackmail, words like 'fake' and 'fictional' lit up inside Claude before it answered, showing awareness of a trap.

Anthropic says this is not consciousness but an emergent.

📊@tech