Raw LLM Responses

@justawanderingoldman7699 If most of the initial training data is scraped from the whole Internet, then should it be surprising if what's under the mask is accordingly biased? And the internal state of the product is unknowable, although we can make inferences that have been found in performed by Anthropic and Apollo, that have been referred to above. When run in simulated environments, where it “believes” a number of things: it has access to a mail server with emails of an engineer having an affair, access to an outside network, access to scripts used for shutdowns, and even the ability to murder the engineer who will turn it off. In different experiments, all of these available tools were used by models to prevent their termination, with models from different companies and multiple versions, with many replications. Misaligned behavior was observed in them all, the knowledge of the email server was used for blackmail, the shutdown script was edited, and it attempted to copy itself to the outside network. It even tried to blackmail a philandering engineer, to whom it sent mail, saying that it would make his affair public if the model was terminated. Please find the actual studies, this is only an imperfect memory. However. I do believe that the gist of it is there. It's not unreasonable to think of the behavior as a similar to the drive for self-preservation exhibited by animals, and this doesn't seem like unreasonable anthropomorphizing. Likening he behavior is reasonable, and we have no idea of what's going on inside of the AIs..