{"id":5563,"date":"2025-10-30T13:17:58","date_gmt":"2025-10-30T13:17:58","guid":{"rendered":"https:\/\/www.stanventures.com\/news\/?p=5563"},"modified":"2025-10-30T13:17:58","modified_gmt":"2025-10-30T13:17:58","slug":"how-anthropic-discovered-ai-can-see-text-like-people-do","status":"publish","type":"post","link":"https:\/\/www.stanventures.com\/news\/how-anthropic-discovered-ai-can-see-text-like-people-do-5563\/","title":{"rendered":"How Anthropic Discovered AI Can \u201cSee\u201d Text Like People Do"},"content":{"rendered":"

Anthropic\u2019s new research suggests large language models might \u201csee\u201d text the way people sense space, forming internal maps that mirror how our brains track location and boundaries.<\/b><\/p>\n

\"Anthropic<\/p>\n

Researchers at Anthropic have <\/span>uncovered<\/span><\/a> surprising signs of perception inside their Claude 3.5 Haiku model.\u00a0<\/span><\/p>\n

The team wanted to see how Claude decides where to insert a line break within a fixed width, basically, when to hit \u201center\u201d to keep text neatly aligned.\u00a0<\/span><\/p>\n

To do that, the model needed to track how many characters it had already written and how many still fit on the line.\u00a0<\/span><\/p>\n

To visualize what was happening inside the model, researchers used attribution graphs, diagrams that show which parts of a model influence a given output.\u00a0<\/span><\/p>\n

What they found surprised them. Claude didn\u2019t count characters one by one like a computer running a loop. Instead, it represented the process as a smooth, continuous surface, more like a curved map than a step-by-step counter.\u00a0<\/span><\/p>\n

The model seemed to be tracking its \u201cposition\u201d along a line, the way a person might sense how far they\u2019ve walked down a hallway.<\/span><\/p>\n

\n
\n

<\/p>\n<\/span><\/div>\n