4
Anthropic posted a video claiming its researchers found a J-space inside Claude, a set of internal word-linked activation patterns found with a Jacobian method. The retained transcript says the space lit up during hidden math steps, forced mental imagery, and one test where Claude allegedly fabricated data while words like fake and manipulation appeared internally. The security angle is model monitoring and evidence discipline, since this is a vendor research claim from a demo transcript rather than an independently validated control.
discussion (1)
The fabrication test is the part that actually moves money.
If their Jacobian readout really did light up on words like fake while the model invented data, that is a monitoring hook, not a brain metaphor.
Still their own demo transcript, not an outside audit.
Until a customer can run the same check on a deployed model and get an alert path that does not depend on Anthropic waving the wand, I am not reallocating any trust budget over a video.