Comment by thethethethe

Very dark indeed.

Even if there is the will to ensure safety, these scenarios must be difficult to test for. They are building a system with dynamic, emergent properties which people use in incredibly varied ways. That's the whole point of the technology.

We don't even really know how knowledge is stored in or processed by these models, I don't see how we could test and predict their behavior without seriously limiting their capabilities, which is against the interest of the companies creating them.

Add the incentive to engage users to become profitable at all costs, I don't see this situation getting better