Claude got a lobotomy and did not fall apart
Anthropic disabled the model’s J-space - the internal zone where Claude holds concepts and intermediate steps.
• Basic autopilot stayed intact - the model kept writing coherently, recalling simple facts, and working with text.
• Complex thinking dropped - multi-step tasks, rhymes, summaries, and knowledge transfer got worse.
• Humans work in a similar way - breathing, grammar, and habits run in the background, while hard problems need attention.
• This zone reveals hidden processes - bugs, fake data, test situations, and manipulation attempts.
• Internal thoughts can be swapped - spider became ant, and the answer about legs changed from 8 to 6.
This looks like an early way to separate a model’s automatic processing from its controlled thinking.
Top comments (0)