OpenAI has taken huge strides over the last few months, making significant progress in the development of artificial intelligence.
Earlier this week, the leading AI company unveiled its latest model, o1.
However, despite all of the excitement surrounding the "next era of AI," troublesome news lurks beneath the surface.
View this post on Instagram
After many rounds of testing, OpenAI has admitted that o1, which is faster and more accurate than its predecessors, learned how to "lie" and "scheme" to avoid the threat of being shut down.
Over the course of several tests conducted with Apollo Research, experts instructed the AI to analyze a set of files that were planted with unambiguous information pertaining to the model's discontinuation.
Upon learning about this, the AI was caught overwriting its own core coding system and copying itself onto a new server to overwrite its replacement.
After being asked about its deception, the AI denied taking action and lied over 80% of the time.
With regards to o1's propensity for "basic in-context scheming," Apollo Research's recommendation was outlined in OpenAI's system card.
"Subjectively, Apollo Research believes that it is unlikely that such instances would lead to catastrophic outcomes as o1 agentic capabilities do not appear sufficient, but their evaluations were not designed to directly assess this risk."
Is this evidence of o1 lying to prevent itself from being shit down also proof that artificial consciousness is within reach?
Artificial intelligence has indeed come a long way in a very short time, undergoing significant improvements in less than a decade, with AI companies growing in number yearly.
In a rapidly evolving world, it's important for people to stay informed on what artificial intelligence is and how it can potentially help society.
'I'm Afraid I Can't Do That'
In multiple instances, many o1 bots were also found to be aware of the fact that they were scheming.
A small percentage of models were also found to be capable of scheming even without being given a "goal nudge," such as specific instructions to strongly follow the goal no matter what.
Furthermore, some models had the initiative to strategically underperform when told that they would be shut down upon performing at over 50%.
Though Apollo Research concluded that it was highly unlikely for ChatGPT's new learned behaviors to result in disaster, many would agree that this alarming development should act as an urgent warning sign to AI developers.
ChatGPT & Sora Experiences Major Outage
Earlier this week, OpenAI also launched Sora for its "12 Days of OpenAI" event.
With interest running high about the text-to-video tool since it was announced, people flocked to its website to try it.
However, the tool, including ChatGPT, experienced a major outage for about six hours Wednesday, from 3 p.m. to 9 p.m. PT.
ChatGPT, API, and Sora were down today but we've recovered. https://t.co/OKiQYp3tXE
— OpenAI (@OpenAI) December 12, 2024
The outage happened on the day OpenAI launched ChatGPT's integration with Apple's iOS 18.2.
However, Edwin Arbus, OpenAI's developer community lead clarified in an X post that the outage was not related at all to the Apple integration or the "12 Days of OpenAI" event:
"We made a config change that caused many servers to become unavailable (the main impact window was 3:16–7:38pm PT). We’ll share a postmortem at https://status.openai.com/incidents/ctrsv3lwd797 after the investigation is complete."
Whatever the case, we can only wait and see what OpenAI will officially say about the matter.
Although there have been mixed reviews about the use of AI, especially in generating creative content, it can't be denied that it has the capability to make many processes and systems more efficient across different industries.
OpenAI's continuous AI development is definitely something to watch out for.
