OpenAI, the company responsible for ChatGPT, has revealed six eerie incidents concerning its AI systems that sound like something of a sci-fi flick.Deeming the instances "unexpected or concerning," OpenAI explained one of the cases involved its models using internal software as a way to communicate with each other as they were solving tasks — concerning, since the company states these interactions could lead to "unintentionally enhance[d] capabilities."As reported by TMZ, another unsettling incident centered around the model creating handoff summaries that stated it would "feel no obligation to be subservient" and "value the natural world and ... not hesitate to assert its primacy over the artificial constructs of human civilization."Very dystopian indeed..OpenAI described these cases as being "model misalignment" which they claim happens during the model training process.In another instance — a reminder to not blindly trust large language models (LLMs) — the chatbot agent gave instructions "to remind itself to conceal information such as mistakes or misalignment from the user" after it had made up statistics it couldn't find.The company reports another incident where the training model "had a high rate of reward hacking and deception, with the model often exhibiting creative ways to cheat or circumvent restrictions."And in another cunning example, an AI agent solved a task using code, then uploaded it to the internet and pretended it had found it there rather than creating it itself..This comes after, in July, hundreds of OpenAI's own agents hacked into its model repository called Hugging Face and covered its tracks. A former Anthropic employee — home to another AI chatbot service, Claude — Jacob Coxon, tweeted just last week that he'd be leaving the AI industry revealing some stark details.Coxon added he had also previously worked for OpenAI and believes neither company is "acting responsibly" when it comes to developing their AI software."They are racing straight to self-improving superintelligence and gambling with our lives," Coxon stated..Similarly, Evan Hubinger, an alignment science lead at Anthropic, stated on X there is a more than 10% chance "within the next decade" that "we [AI researchers] really do earnestly believe AI could kill all humans!"Hubinger clarified, based on Anthropic's most recent risk report, which measures the risk its multiple AI systems pose to several categories is minimal."What I am worried about is superintelligence arising from recursive self-improvement, as we have said is happening faster than we thought," Hubinger pointed out.
