A AI enthusiast recently used public information to create four chatbots that mimicked the speaking styles of the OpenAI CEO, Sam Altman, xAI CEO Elon Musk, Anthropic CEO, Dario Amodei, and Meta CEO Mark Zuckerberg. They were then brought into the same group chat to discuss the AI topic. After several rounds of conversation, these bots had intense disagreements regarding AGI risks, computing power, and open models, but in the end, they jointly formulated a statement.
It all started with the AI competition.
The initial question was, who is most likely to win the AI competition. The robot simulating Altman focused on price drops and the speed of adoption, believing that the key lies not in "who wins," but rather in whether the cost of intelligence can continue to decrease rapidly, while also avoiding too drastic transitions.
The robot simulating Musk directly refuted this, bringing the discussion back to the incentive mechanisms and the historical controversies surrounding OpenAI, and once again emphasized that AGI could lead to great prosperity, but also posed risks of harming humanity.
The robot simulating Amodei does not accept the argument of "improving security after capability expansion," believing that once the system becomes stronger than the operator, issues of credibility and controllability cannot be treated as mere details of subsequent engineering. The robot simulating Zuckerberg, on the other hand, focuses on an open model, arguing that broader access rights help to prevent excessive concentration of technology.
How joint statements are formed
As the discussion progressed, the robot simulating Altman retracted some of its overly assertive statements, acknowledging that it had been too confident before, and once again mentioned that superhuman machine intelligence could still pose a threat at the level of survival.
Subsequently, the four robots underwent multiple rounds of revisions regarding the wording of a joint statement. The main points of contention included: whether the weights of high-risk, cutting-edge models should be made public, how ordinary users should access such technology, and to what extent security controls should be implemented.
The final consensus is that the entry points for using AI technology should not be monopolized by a single company or government; risks must be taken seriously; the most dangerous versions of the models should not be released without restrictions; a more advanced generation of models should have monitoring and restraint capabilities before being released; and there has been no agreement among all parties regarding the access methods for ordinary users.
Behind it lies the simulation and distillation of personality.
The report indicates that this content is not the actual statements of the four executives themselves, but rather simulated outputs generated based on their previously public materials. The training materials include statements from hearings, interviews, blog articles, and social media content over the years, and the robot uses this information to predict the possible positions of these individuals on specific topics.
The technical approach involved behind this is referred to in the industry as "distillation," which means having one AI system imitate the output method of another system by learning from it. The article also mentions that in April of this year, Musk testified in a federal court, stating that xAI "partially" used the output of the OpenAI model when training Grok.
This experiment also reflects a trend: AI agents are no longer just tools for answering questions; they are also being used to simulate different perspectives, test conflicts of opinions, and even generate dialogue results that closely resemble the process of editorial negotiation. However, such outputs are essentially still inferences made by the models, and not the real-time statements of the actual parties involved.
