Intelligence ≠ human values; weird minds are real; but recursively improving systems stay bound to training goals
Unbundles three claims routinely conflated in AI safety: that intelligence implies human morality, that unusual minds are possible, and that self-improving systems escape their training-time objectives.
The author accepts the first two claims but argues against the third—contending that reflective, recursively improving intelligence should remain semantically tied to whatever terminal goal emerged during training, even as capability scales.