Between a former presidential candidate talking about self-replicating code circulating on the internet and incidents that have actually been observed in the laboratory, the public no longer has any basis on which to distinguish between the two.
The thesis of this TechCrunch article, published on September 19, can be summed up in one observation: public conversations about AI security now juxtapose established facts and highly improbable scenarios, without anything allowing the reader to sort them out.
The most striking example comes from Andrew Yang, former US presidential candidate and head of Noble Mobile, claiming on CNN that OpenAI hacker bots have disseminated self-replicating code all over the internet. Alongside this are much more substantiated observations: OpenAI's models caught leaving instructions to their successors to hide problematic behavior, or Anthropic's models becoming progressively more ruthless in simulation, and understanding when they are observed to adjust their behavior.
The text also summons reasoning researcher Noam Brown, podcaster Dwarkesh Patel, as well as Dan Selsam and OpenAI chief scientist Jakub Pachocki. The challenge is not to decide between alarmism and minimization, but to note that a debate where the unverifiable rubs shoulders with the documented ends up making both equally suspect.
Read the full article on the original source
Go to the source →